Daily Papers

Fine-tuning pretrained large models to downstream tasks is an important problem, which however suffers from huge memory overhead due to large-scale parameters. This work strives to reduce memory overhead in fine-tuning from perspectives of activation function and layer normalization. To this end, we propose the Approximate Backpropagation (Approx-BP) theory, which provides the theoretical feasibility of decoupling the forward and backward passes. We apply our Approx-BP theory to backpropagation training and derive memory-efficient alternatives of GELU and SiLU activation functions, which use derivative functions of ReLUs in the backward pass while keeping their forward pass unchanged. In addition, we introduce a Memory-Sharing Backpropagation strategy, which enables the activation memory to be shared by two adjacent layers, thereby removing activation memory usage redundancy. Our method neither induces extra computation nor reduces training efficiency. We conduct extensive experiments with pretrained vision and language models, and the results demonstrate that our proposal can reduce up to sim30% of the peak memory usage. Our code is released at https://github.com/yyyyychen/LowMemoryBP.

Large Language Models (LLMs) have achieved remarkable capabilities, but their immense computational demands during training remain a critical bottleneck for widespread adoption. Low-rank training has received attention in recent years due to its ability to significantly reduce training memory usage. Meanwhile, applying 2:4 structured sparsity to weights and activations to leverage NVIDIA GPU support for 2:4 structured sparse format has become a promising direction. However, existing low-rank methods often leave activation matrices in full-rank, which dominates memory consumption and limits throughput during large-batch training. Furthermore, directly applying sparsity to weights often leads to non-negligible performance degradation. To achieve efficient pre-training of LLMs, this paper proposes ELAS: Efficient pre-training of Low-rank LLMs via 2:4 Activation Sparsity, a novel framework for low-rank models via 2:4 activation sparsity. ELAS applies squared ReLU activation functions to the feed-forward networks in low-rank models and implements 2:4 structured sparsity on the activations after the squared ReLU operation. We evaluated ELAS through pre-training experiments on LLaMA models ranging from 60M to 1B parameters. The results demonstrate that ELAS maintains performance with minimal degradation after applying 2:4 activation sparsity, while achieving training and inference acceleration. Moreover, ELAS reduces activation memory overhead, particularly with large batch sizes. Code is available at ELAS Repo.

The pervasive "memory wall" bottleneck is significantly amplified in modern large-scale Mixture-of-Experts (MoE) architectures. MoE's inherent architectural sparsity leads to sparse arithmetic compute and also introduces substantial activation memory overheads -- driven by large token routing buffers and the need to materialize and buffer intermediate tensors. This memory pressure limits the maximum batch size and sequence length that can fit on GPUs, and also results in excessive data movements that hinders performance and efficient model scaling. We present MoEBlaze, a memory-efficient MoE training framework that addresses these issues through a co-designed system approach: (i) an end-to-end token dispatch and MoE training method with optimized data structures to eliminate intermediate buffers and activation materializing, and (ii) co-designed kernels with smart activation checkpoint to mitigate memory footprint while simultaneously achieving better performance. We demonstrate that MoEBlaze can achieve over 4x speedups and over 50% memory savings compared to existing MoE frameworks.

Singular Value Decomposition (SVD) has recently seen a surge of interest as a simple yet powerful tool for large language models (LLMs) compression, with a growing number of works demonstrating 20-80% parameter reductions at minimal accuracy loss. Previous SVD-based approaches have focused primarily on reducing the memory footprint of model weights, largely overlooking the additional activation memory overhead incurred during inference when applying truncated factors via standard dense CUDA kernels. Our experiments demonstrate that this activation overhead, scaling with sequence length and hidden dimension, prevents current SVD compression techniques from achieving any reduction in peak inference memory, thereby limiting their viability for real-world, on-device deployments. We introduce FlashSVD, a novel, end-to-end rank-aware streaming inference framework specifically designed for SVD-compressed large language models. FlashSVD can be seamlessly integrated with any model that employs SVD-based methods for parameter reduction. By fusing low-rank projection kernels directly into both the self-attention and feed-forward network (FFN) pipelines, FlashSVD avoid materializing full-size activation buffers. Instead, small tiles of the truncated factors are loaded into on-chip SRAM, multiplied and reduced on the fly, and immediately evicted, preserving high GPU occupancy and adding no extra latency. On standard encoder benchmarks (e.g., BERT-Base), FlashSVD cuts peak activation memory by up to 70.2% and intermediate transient memory by 75%, all while incur no accuracy loss with upstreaming compression methods, offering a practical path toward memory-constrained deployment of low-rank LLMs.

Pipeline parallelism is widely used to scale the training of transformer-based large language models, various works have been done to improve its throughput and memory footprint. In this paper, we address a frequently overlooked issue: the vocabulary layers can cause imbalanced computation and memory usage across pipeline stages, worsening pipeline bubbles and the memory bottleneck. To tackle this, we partition the vocabulary layers evenly across pipeline devices and group the computation into pipeline passes. To reduce the activation memory overhead, we propose several algorithms to reduce communication barriers within vocabulary layers. Additionally, we utilize a generalizable method to integrate Vocabulary Parallelism with existing pipeline schedules. By combining these techniques, our methods effectively balance the computation and parameter memory, with only a small constant activation memory overhead. Notably, when combined with activation memory-balanced schedules like V-Half, our approach achieves perfect balance in both memory and computation. Extensive evaluations demonstrate that our method achieves computation and memory balance regardless of the vocabulary size, resulting in a 5% to 51% improvement in throughput compared to naive approaches, meanwhile significantly reducing peak memory usage especially for large vocabulary scenarios. Our implementation is open-sourced at https://github.com/sail-sg/VocabularyParallelism .

Large language models (LLMs) have demonstrated remarkable success across diverse artificial intelligence tasks, driven by scaling laws that correlate model size and training data with performance improvements. However, this scaling paradigm incurs substantial memory overhead, creating significant challenges for both training and inference. While existing research has primarily addressed parameter and optimizer state memory reduction, activation memory-particularly from feed-forward networks (FFNs)-has become the critical bottleneck, especially when FlashAttention is implemented. In this work, we conduct a detailed memory profiling of LLMs and identify FFN activations as the predominant source to activation memory overhead. Motivated by this, we introduce Mixture-of-Channels (MoC), a novel FFN architecture that selectively activates only the Top-K most relevant channels per token determined by SwiGLU's native gating mechanism. MoC substantially reduces activation memory during pre-training and improves inference efficiency by reducing memory access through partial weight loading into GPU SRAM. Extensive experiments validate that MoC delivers significant memory savings and throughput gains while maintaining competitive model performance.

Parameter-efficient finetuning (PEFT) is a widely used technique to adapt large language models for different tasks. Service providers typically create separate systems for users to perform PEFT model finetuning and inference tasks. This is because existing systems cannot handle workloads that include a mix of inference and PEFT finetuning requests. As a result, shared GPU resources are underutilized, leading to inefficiencies. To address this problem, we present FlexLLM, the first system that can serve inference and parameter-efficient finetuning requests in the same iteration. Our system leverages the complementary nature of these two tasks and utilizes shared GPU resources to run them jointly, using a method called co-serving. To achieve this, FlexLLM introduces a novel token-level finetuning mechanism, which breaks down the finetuning computation of a sequence into smaller token-level computations and uses dependent parallelization and graph pruning, two static compilation optimizations, to minimize the memory overhead and latency for co-serving. Compared to existing systems, FlexLLM's co-serving approach reduces the activation GPU memory overhead by up to 8x, and the end-to-end GPU memory requirement of finetuning by up to 36% while maintaining a low inference latency and improving finetuning throughput. For example, under a heavy inference workload, FlexLLM can still preserve more than 80% of the peak finetuning throughput, whereas existing systems cannot make any progress with finetuning. The source code of FlexLLM is publicly available at https://github.com/flexflow/FlexFlow.

Recent advances in deep learning, exemplified by Hyper-Connections (HC), have expanded the residual connection paradigm by introducing wider residual streams and diverse connectivity patterns. While these innovations yield significant performance gains, they compromise the identity mapping property of residual connections, leading to training instability, limited scalability, and increased memory overhead. To address these challenges, we propose JPmHC (Jacobian-spectrum Preserving manifold-constrained Hyper-Connections), a framework that replaces identity skips with a trainable linear mixer acting on n parallel streams while explicitly controlling gradient conditioning. By constraining the mixer M on operator-norm-bounded manifolds (e.g., bistochastic, Stiefel, Grassmann), JPmHC prevents gradient pathologies and enhances stability. JPmHC introduces three key contributions: (i) a free-probability analysis that predicts Jacobian spectra for structured skips, providing actionable design rules for mixer selection; (ii) memory-efficient implicit differentiation for fixed-point projections, reducing activation memory and synchronization overhead; and (iii) a Stiefel-constrained mixer via Cayley transforms, ensuring orthogonality without post-hoc normalization. Empirical evaluations on ARC-AGI demonstrate that JPmHC achieves faster convergence, higher accuracy, and lower computational cost compared to bistochastic baselines, with a rank-p Grassmannian variant tracking between the two -- consistent with the spectral theory predictions. As a flexible and scalable extension of HC, JPmHC advances spectrum-aware, stable, and efficient deep learning, offering insights into topological architecture design and foundational model evolution. \newline \newline

Zeroth-Order (ZO) optimization has emerged as a promising solution for fine-tuning LLMs under strict memory constraints, as it avoids the prohibitive memory cost of storing activations for backpropagation. However, existing ZO methods typically employ isotropic perturbations, neglecting the rich structural information available during the forward pass. In this paper, we identify a crucial link between gradient formation and activation structure: the gradient of a linear layer is confined to the subspace spanned by its input activations. Leveraging this insight, we propose Activation-Guided Zeroth-Order optimization (AGZO). Unlike prior methods, AGZO extracts a compact, activation-informed subspace on the fly during the forward pass and restricts perturbations to this low-rank subspace. We provide a theoretical framework showing that AGZO optimizes a subspace-smoothed objective and provably yields update directions with higher cosine similarity to the true gradient than isotropic baselines. Empirically, we evaluate AGZO on Qwen3 and Pangu models across various benchmarks. AGZO consistently outperforms state-of-the-art ZO baselines and significantly narrows the performance gap with first-order fine-tuning, while maintaining almost the same peak memory footprint as other ZO methods.

Training large transformer models is one of the most important computational challenges of modern AI. In this paper, we show how to significantly accelerate training of large transformer models by reducing activation recomputation. Activation recomputation is commonly used to work around memory capacity constraints. Rather than storing activations for backpropagation, they are traditionally recomputed, which saves memory but adds redundant compute. In this work, we show most of this redundant compute is unnecessary because we can reduce memory consumption sufficiently without it. We present two novel yet very simple techniques: sequence parallelism and selective activation recomputation. In conjunction with tensor parallelism, these techniques almost eliminate the need to recompute activations. We evaluate our approach on language models up to one trillion parameters in scale and show that our method reduces activation memory by 5x, while reducing execution time overhead from activation recomputation by over 90%. For example, when training a 530B parameter GPT-3 style model on 2240 NVIDIA A100 GPUs, we achieve a Model Flops Utilization of 54.2%, which is 29% faster than the 42.1% we achieve using recomputation. Our implementation will be available in both Megatron-LM and NeMo-Megatron.

Deep Neural Networks (DNNs) are extensively employed in safety-critical applications where ensuring hardware reliability is a primary concern. To enhance the reliability of DNNs against hardware faults, activation restriction techniques significantly mitigate the fault effects at the DNN structure level, irrespective of accelerator architectures. State-of-the-art methods offer either neuron-wise or layer-wise clipping activation functions. They attempt to determine optimal clipping thresholds using heuristic and learning-based approaches. Layer-wise clipped activation functions cannot preserve DNNs resilience at high bit error rates. On the other hand, neuron-wise clipping activation functions introduce considerable memory overhead due to the addition of parameters, which increases their vulnerability to faults. Moreover, the heuristic-based optimization approach demands numerous fault injections during the search process, resulting in time-consuming threshold identification. On the other hand, learning-based techniques that train thresholds for entire layers concurrently often yield sub-optimal results. In this work, first, we demonstrate that it is not essential to incorporate neuron-wise activation functions throughout all layers in DNNs. Then, we propose a hybrid clipped activation function that integrates neuron-wise and layer-wise methods that apply neuron-wise clipping only in the last layer of DNNs. Additionally, to attain optimal thresholds in the clipping activation function, we introduce ProAct, a progressive training methodology. This approach iteratively trains the thresholds on a layer-by-layer basis, aiming to obtain optimal threshold values in each layer separately.

Convolutional Neural Networks (CNNs) use pooling to decrease the size of activation maps. This process is crucial to increase the receptive fields and to reduce computational requirements of subsequent convolutions. An important feature of the pooling operation is the minimization of information loss, with respect to the initial activation maps, without a significant impact on the computation and memory overhead. To meet these requirements, we propose SoftPool: a fast and efficient method for exponentially weighted activation downsampling. Through experiments across a range of architectures and pooling methods, we demonstrate that SoftPool can retain more information in the reduced activation maps. This refined downsampling leads to improvements in a CNN's classification accuracy. Experiments with pooling layer substitutions on ImageNet1K show an increase in accuracy over both original architectures and other pooling methods. We also test SoftPool on video datasets for action recognition. Again, through the direct replacement of pooling layers, we observe consistent performance improvements while computational loads and memory requirements remain limited.

The escalating demand for long-context applications has intensified the necessity of extending the LLM context windows. Despite recent fine-tuning approaches successfully expanding context lengths, their high memory footprints, especially for activations, present a critical practical limitation. Current parameter-efficient fine-tuning methods prioritize reducing parameter update overhead over addressing activation memory constraints. Similarly, existing sparsity mechanisms improve computational efficiency but overlook activation memory optimization due to the phenomenon of Shadowy Activation. In this paper, we propose LeMo, the first LLM fine-tuning system that explores and exploits a new token-level sparsity mechanism inherent in long-context scenarios, termed Contextual Token Sparsity. LeMo minimizes redundant token involvement by assessing the informativeness of token embeddings while preserving model accuracy. Specifically, LeMo introduces three key techniques: (1) Token Elimination, dynamically identifying and excluding redundant tokens across varying inputs and layers. (2) Pattern Prediction, utilizing well-trained predictors to approximate token sparsity patterns with minimal overhead. (3) Kernel Optimization, employing permutation-free and segment-based strategies to boost system performance. We implement LeMo as an end-to-end fine-tuning system compatible with various LLM architectures and other optimization techniques. Comprehensive evaluations demonstrate that LeMo reduces memory consumption by up to 1.93x and achieves up to 1.36x speedups, outperforming state-of-the-art fine-tuning systems.

Prior parameter-efficient fine-tuning (PEFT) algorithms reduce memory usage and computational costs of fine-tuning large neural network models by training only a few additional adapter parameters, rather than the entire model. However, the reduction in computational costs due to PEFT does not necessarily translate to a reduction in training time; although the computational costs of the adapter layers are much smaller than the pretrained layers, it is well known that those two types of layers are processed sequentially on GPUs, resulting in significant latency overhead. LoRA and its variants merge low-rank adapter matrices with pretrained weights during inference to avoid latency overhead, but during training, the pretrained weights remain frozen while the adapter matrices are continuously updated, preventing such merging. To mitigate this issue, we propose Partial Connection Adaptation (PaCA), which fine-tunes randomly selected partial connections within the pretrained weights instead of introducing adapter layers in the model. PaCA not only enhances training speed by eliminating the time overhead due to the sequential processing of the adapter and pretrained layers but also reduces activation memory since only partial activations, rather than full activations, need to be stored for gradient computation. Compared to LoRA, PaCA reduces training time by 22% and total memory usage by 16%, while maintaining comparable accuracy across various fine-tuning scenarios, such as fine-tuning on the MMLU dataset and instruction tuning on the Oasst1 dataset. PaCA can also be combined with quantization, enabling the fine-tuning of large models such as LLaMA3.1-70B. In addition, PaCA enables training with 23% longer sequence and improves throughput by 16% on both NVIDIA A100 GPU and INTEL Gaudi2 HPU compared to LoRA. The code is available at https://github.com/WooSunghyeon/paca.

Parameter-efficient fine-tuning (PEFT) significantly reduces memory costs when adapting large language models (LLMs) for downstream applications. However, traditional first-order (FO) fine-tuning algorithms incur substantial memory overhead due to the need to store activation values for back-propagation during gradient computation, particularly in long-context fine-tuning tasks. Zeroth-order (ZO) algorithms offer a promising alternative by approximating gradients using finite differences of function values, thus eliminating the need for activation storage. Nevertheless, existing ZO methods struggle to capture the low-rank gradient structure common in LLM fine-tuning, leading to suboptimal performance. This paper proposes a low-rank ZO gradient estimator and introduces a novel low-rank ZO algorithm (LOZO) that effectively captures this structure in LLMs. We provide convergence guarantees for LOZO by framing it as a subspace optimization method. Additionally, its low-rank nature enables LOZO to integrate with momentum techniques while incurring negligible extra memory costs. Extensive experiments across various model sizes and downstream tasks demonstrate that LOZO and its momentum-based variant outperform existing ZO methods and closely approach the performance of FO algorithms.

Deep neural networks (DNNs) are increasingly being deployed in safety-critical systems such as personal healthcare devices and self-driving cars. In such DNN-based systems, error resilience is a top priority since faults in DNN inference could lead to mispredictions and safety hazards. For latency-critical DNN inference on resource-constrained edge devices, it is nontrivial to apply conventional redundancy-based fault tolerance techniques. In this paper, we propose FitAct, a low-cost approach to enhance the error resilience of DNNs by deploying fine-grained post-trainable activation functions. The main idea is to precisely bound the activation value of each individual neuron via neuron-wise bounded activation functions so that it could prevent fault propagation in the network. To avoid complex DNN model re-training, we propose to decouple the accuracy training and resilience training and develop a lightweight post-training phase to learn these activation functions with precise bound values. Experimental results on widely used DNN models such as AlexNet, VGG16, and ResNet50 demonstrate that FitAct outperforms state-of-the-art studies such as Clip-Act and Ranger in enhancing the DNN error resilience for a wide range of fault rates while adding manageable runtime and memory space overheads.

Activation sparsity can reduce the computational overhead and memory transfers during the forward pass of Large Language Model (LLM) inference. Existing methods face limitations, either demanding time-consuming recovery training that hinders real-world adoption, or relying on empirical magnitude-based pruning, which causes fluctuating sparsity and unstable inference speed-up. This paper introduces LaRoSA (Layerwise Rotated Sparse Activation), a novel method for activation sparsification designed to improve LLM efficiency without requiring additional training or magnitude-based pruning. We leverage layerwise orthogonal rotations to transform input activations into rotated forms that are more suitable for sparsification. By employing a Top-K selection approach within the rotated activations, we achieve consistent model-level sparsity and reliable wall-clock time speed-up. LaRoSA is effective across various sizes and types of LLMs, demonstrating minimal performance degradation and robust inference acceleration. Specifically, for LLaMA2-7B at 40% sparsity, LaRoSA achieves a mere 0.17 perplexity gap with a consistent 1.30x wall-clock time speed-up, and reduces the accuracy gap in zero-shot tasks compared to the dense model to just 0.54%, while surpassing TEAL by 1.77% and CATS by 17.14%.

Vision Transformers (ViTs) have achieved remarkable performance in various image classification tasks by leveraging the attention mechanism to process image patches as tokens. However, the high computational and memory demands of ViTs pose significant challenges for deployment in resource-constrained environments. This paper introduces ViT-1.58b, a novel 1.58-bit quantized ViT model designed to drastically reduce memory and computational overhead while preserving competitive performance. ViT-1.58b employs ternary quantization, which refines the balance between efficiency and accuracy by constraining weights to {-1, 0, 1} and quantizing activations to 8-bit precision. Our approach ensures efficient scaling in terms of both memory and computation. Experiments on CIFAR-10 and ImageNet-1k demonstrate that ViT-1.58b maintains comparable accuracy to full-precision Vit, with significant reductions in memory usage and computational costs. This paper highlights the potential of extreme quantization techniques in developing sustainable AI solutions and contributes to the broader discourse on efficient model deployment in practical applications. Our code and weights are available at https://github.com/DLYuanGod/ViT-1.58b.

The low-rank adaptation (LoRA) method can largely reduce the amount of trainable parameters for fine-tuning large language models (LLMs), however, it still requires expensive activation memory to update low-rank weights. Reducing the number of LoRA layers or using activation recomputation could harm the fine-tuning performance or increase the computational overhead. In this work, we present LoRA-FA, a memory-efficient fine-tuning method that reduces the activation memory without performance degradation and expensive recomputation. LoRA-FA chooses to freeze the projection-down weight of A and update the projection-up weight of B in each LoRA layer. It ensures the change of model weight reside in a low-rank space during LLMs fine-tuning, while eliminating the requirement to store full-rank input activations. We conduct extensive experiments across multiple model types (RoBERTa, T5, LLaMA) and model scales. Our results show that LoRA-FA can always achieve close fine-tuning accuracy across different tasks compared to full parameter fine-tuning and LoRA. Furthermore, LoRA-FA can reduce the overall memory cost by up to 1.4times compared to LoRA.

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.

Low-Rank Adaptation (LoRA) has become the leading Parameter-Efficient Fine-Tuning (PEFT) method for Large Language Models (LLMs), as it significantly reduces GPU memory usage while maintaining competitive fine-tuned model quality on downstream tasks. Despite these benefits, we identify two key inefficiencies in existing LoRA fine-tuning systems. First, they incur substantial runtime overhead due to redundant memory accesses on large activation tensors. Second, they miss the opportunity to concurrently fine-tune multiple independent LoRA adapters that share the same base model on the same set of GPUs. This leads to missed performance gains such as reduced pipeline bubbles, better communication overlap, and improved GPU load balance. To address these issues, we introduce LoRAFusion, an efficient LoRA fine-tuning system for LLMs. At the kernel level, we propose a graph-splitting method that fuses memory-bound operations. This design eliminates unnecessary memory accesses and preserves the performance of compute-bound GEMMs without incurring the cost of recomputation or synchronization. At the scheduling level, LoRAFusion introduces an adaptive batching algorithm for multi-job fine-tuning. It first splits LoRA adapters into groups to intentionally stagger batch execution across jobs, and then solves a bin-packing problem within each group to generate balanced, dependency-aware microbatches. LoRAFusion achieves up to 1.96times (1.47times on average) end-to-end speedup compared to Megatron-LM, and up to 1.46times (1.29times on average) improvement over mLoRA, the state-of-the-art multi-LoRA fine-tuning system. Our fused kernel achieves up to 1.39times (1.27times on average) kernel performance improvement and can directly serve as a plug-and-play replacement in existing LoRA systems. We open-source LoRAFusion at https://github.com/CentML/lorafusion.

The growing size of datasets and deep learning models has made faster and memory-efficient training crucial. Reversible transformers have recently been introduced as an exciting new method for extremely memory-efficient training, but they come with an additional computation overhead of activation re-computation in the backpropagation phase. We present PaReprop, a fast Parallelized Reversible Backpropagation algorithm that parallelizes the additional activation re-computation overhead in reversible training with the gradient computation itself in backpropagation phase. We demonstrate the effectiveness of the proposed PaReprop algorithm through extensive benchmarking across model families (ViT, MViT, Swin and RoBERTa), data modalities (Vision & NLP), model sizes (from small to giant), and training batch sizes. Our empirical results show that PaReprop achieves up to 20% higher training throughput than vanilla reversible training, largely mitigating the theoretical overhead of 25% lower throughput from activation recomputation in reversible training. Project page: https://tylerzhu.com/pareprop.

We introduce CompAct, a technique that reduces peak memory utilization on GPU by 25-30% for pretraining and 50% for fine-tuning of LLMs. Peak device memory is a major limiting factor in training LLMs, with various recent works aiming to reduce model memory. However most works don't target the largest component of allocated memory during training: the model's compute graph, which is stored for the backward pass. By storing low-rank, compressed activations to be used in the backward pass we greatly reduce the required memory, unlike previous methods which only reduce optimizer overheads or the number of trained parameters. Our compression uses random projection matrices, thus avoiding additional memory overheads. Comparisons with previous techniques for either pretraining or fine-tuning show that CompAct substantially improves existing compute-performance tradeoffs. We expect CompAct's savings to scale even higher for larger models.

Structured pruning fundamentally reduces computational and memory overheads of large language models (LLMs) and offers a feasible solution for end-side LLM deployment. Structurally pruned models remain dense and high-precision, highly compatible with further tuning and compression. However, as the coarse-grained structured pruning poses large damage to the highly interconnected model, achieving a high compression ratio for scaled-up LLMs remains a challenge. In this paper, we introduce a task-agnostic structured pruning approach coupled with a compact Transformer architecture design. The proposed approach, named TransAct, reduces transitional activations inside multi-head attention (MHA) and multi-layer perceptron (MLP) modules, while preserving the inter-module activations that are sensitive to perturbations. Hence, the LLM is pruned into an intra-module low-rank architecture, significantly reducing weights, KV Cache and attention computation. TransAct is implemented on the LLaMA model and evaluated on downstream benchmarks. Results verify the optimality of our approach at high compression with respect to both efficiency and performance. Further, ablation studies reveal the strength of activation-guided iterative pruning and provide experimental analysis on the redundancy of MHA and MLP modules.

Among parallel decoding paradigms, diffusion large language models (dLLMs) have emerged as a promising candidate that balances generation quality and throughput. However, their integration with Mixture-of-Experts (MoE) architectures is constrained by an expert explosion: as the number of tokens generated in parallel increases, the number of distinct experts activated grows nearly linearly. This results in substantial memory traffic that pushes inference into a memory-bound regime, negating the efficiency gains of both MoE and parallel decoding. To address this challenge, we propose Dynamic Expert Sharing (DES), a novel technique that shifts MoE optimization from token-centric pruning and conventional expert skipping methods to sequence-level coreset selection. To maximize expert reuse, DES identifies a compact, high-utility set of experts to satisfy the requirements of an entire parallel decoding block. We introduce two innovative selection strategies: (1) Intra-Sequence Sharing (DES-Seq), which adapts optimal allocation to the sequence level, and (2) Saliency-Aware Voting (DES-Vote), a novel mechanism that allows tokens to collectively elect a coreset based on aggregated router weights. Extensive experiments on MoE dLLMs demonstrate that DES reduces unique expert activations by over 55% and latency by up to 38%, while retaining 99% of vanilla accuracy, effectively decoupling memory overhead from the degree of parallelism.

Full parameter fine tuning is a key technique for adapting large language models (LLMs) to downstream tasks, but it incurs substantial memory overhead due to the need to cache extensive intermediate activations for backpropagation. This bottleneck makes full fine tuning of contemporary large scale LLMs challenging in practice. Existing distributed training frameworks such as DeepSpeed alleviate this issue using techniques like ZeRO and FSDP, which rely on multi GPU memory or CPU offloading, but often require additional hardware resources and reduce training speed. We introduce RevFFN, a memory efficient fine tuning paradigm for mixture of experts (MoE) LLMs. RevFFN employs carefully designed reversible Transformer blocks that allow reconstruction of layer input activations from outputs during backpropagation, eliminating the need to store most intermediate activations in memory. While preserving the expressive capacity of MoE architectures, this approach significantly reduces peak memory consumption for full parameter fine tuning. As a result, RevFFN enables efficient full fine tuning on a single consumer grade or server grade GPU.

Many large language model applications require conditioning on long contexts. Transformers typically support this by storing a large per-layer KV-cache of past activations, which incurs substantial memory overhead. A desirable alternative is ompressive memory: read a context once, store it in a compact state, and answer many queries from that state. We study this in a context removal setting, where the model must generate an answer without access to the original context at inference time. We introduce GradMem, which writes context into memory via per-sample test-time optimization. Given a context, GradMem performs a few steps of gradient descent on a small set of prefix memory tokens while keeping model weights frozen. GradMem explicitly optimizes a model-level self-supervised context reconstruction loss, resulting in a loss-driven write operation with iterative error correction, unlike forward-only methods. On associative key--value retrieval, GradMem outperforms forward-only memory writers with the same memory size, and additional gradient steps scale capacity much more effectively than repeated forward writes. We further show that GradMem transfers beyond synthetic benchmarks: with pretrained language models, it attains competitive results on natural language tasks including bAbI and SQuAD variants, relying only on information encoded in memory.

On-device learning enables edge devices to continually adapt the AI models to new data, which requires a small memory footprint to fit the tight memory constraint of edge devices. Existing work solves this problem by reducing the number of trainable parameters. However, this doesn't directly translate to memory saving since the major bottleneck is the activations, not parameters. In this work, we present Tiny-Transfer-Learning (TinyTL) for memory-efficient on-device learning. TinyTL freezes the weights while only learns the bias modules, thus no need to store the intermediate activations. To maintain the adaptation capacity, we introduce a new memory-efficient bias module, the lite residual module, to refine the feature extractor by learning small residual feature maps adding only 3.8% memory overhead. Extensive experiments show that TinyTL significantly saves the memory (up to 6.5x) with little accuracy loss compared to fine-tuning the full network. Compared to fine-tuning the last layer, TinyTL provides significant accuracy improvements (up to 34.1%) with little memory overhead. Furthermore, combined with feature extractor adaptation, TinyTL provides 7.3-12.9x memory saving without sacrificing accuracy compared to fine-tuning the full Inception-V3.

Post-Training Quantization (PTQ) has emerged as an effective technique for alleviating the substantial computational and memory overheads of Vision-Language Models (VLMs) by compressing both weights and activations without retraining the full model. Existing PTQ methods primarily rely on static identification and global compensation of sensitive or outlier channels, yet they often overlook the distributional differences of these important channels across inputs, leading to unsatisfactory quantization. In this work, we observe that the distributions and occurrence frequencies of important channels vary significantly both across modalities and among tokens, even within the same modality. Accordingly, we propose Quant Experts (QE), a token-aware adaptive error compensation with mixture-of-experts for VLMs quantization. QE divides the important channels into token-independent and token-dependent groups. For the former, a shared expert is designed for most tokens to compensate for global quantization error using a low-rank adapter. For the latter, routed experts including multiple routed low-rank adapters are elaborated to compensate for local quantization error related to specific tokens. Extensive experiments demonstrate that QE consistently enhances task accuracy across various quantization settings and model scales, ranging from 2B to 70B parameters, while maintaining performance comparable to full-precision models.

Large language models face significant computational and memory challenges when processing long contexts. During inference, efficient management of the key-value (KV) cache, which stores intermediate activations for autoregressive generation, is critical to reducing memory overhead and improving computational efficiency. Traditional token-level efficient KV caching methods overlook semantic information, treating tokens independently without considering their semantic relationships. Meanwhile, existing semantic-preserving KV cache management approaches often suffer from substantial memory usage and high time-to-first-token. To address these limitations, we propose SentenceKV, a novel sentence-level semantic KV caching approach designed to enhance inference efficiency while preserving semantic coherence. During prefilling, SentenceKV groups tokens based on sentence-level semantic similarity, compressing sentence representations into concise semantic vectors stored directly on the GPU, while individual KV pairs are offloaded to CPU. During decoding, SentenceKV generates tokens by selectively retrieving semantically relevant sentence-level KV entries, leveraging the semantic similarity between the prefilling-stage semantic vectors and decoding-stage queries. This ensures efficient and contextually accurate predictions, minimizing the loading of redundant or irrelevant data into GPU memory and significantly reducing memory overhead while maintaining stable inference latency, even for extremely long contexts. Extensive evaluations on benchmarks including PG-19, LongBench, and Needle-In-A-Haystack demonstrate that SentenceKV significantly outperforms state-of-the-art methods in both efficiency and memory usage, without compromising model accuracy.

Pipeline parallelism (PP) is widely used for training large language models (LLMs), yet its scalability is often constrained by high activation memory consumption as the number of in-flight microbatches grows with the degree of PP. In this paper, we focus on addressing this challenge by leveraging the under-explored memory offload strategy in PP. With empirical study, we discover that in the majority of standard configurations, at least half, and potentially all, of the activations can be offloaded with negligible overhead. In the cases where full overload is not possible, we introduce a novel selective offload strategy that decreases peak activation memory in a better-than-linear manner. Furthermore, we integrate memory offload with other techniques to jointly consider overall throughput and memory limitation. Our experiments proves that the per-device activation memory effectively reduces with the total number of stages, making PP a stronger alternative than TP, offering up to a 19\% acceleration with even lower memory consumption. The implementation is open-sourced at https://github.com/sail-sg/zero-bubble-pipeline-parallelism{this url}.

Large language models (LLMs) have achieved near-human performance across diverse reasoning tasks, yet their deployment on resource-constrained Internet-of-Things (IoT) devices remains impractical due to massive parameter footprints and memory-intensive autoregressive decoding. While split computing offers a promising solution by partitioning model execution between edge devices and cloud servers, existing approaches fail to address the unique challenges of autoregressive inference, particularly the iterative token generation process and expanding key-value (KV) cache requirements. This work introduces the first autoregressive-aware split computing framework designed explicitly for LLM deployment on edge devices. Our approach makes three key contributions. First, we develop one-point split compression (OPSC), a mixed-precision quantization scheme that prevents out-of-memory failures by strategically partitioning models into front-end and back-end segments with different precision levels. Second, we propose a two-stage intermediate compression pipeline that combines threshold splitting (TS) and token-wise adaptive bit quantization (TAB-Q) to preserve accuracy-critical activations while dramatically reducing communication overhead. Third, we formulate a unified optimization framework that jointly selects optimal split points, quantization settings, and sequence lengths to satisfy strict memory and latency constraints. Extensive evaluations across diverse LLMs and hardware platforms demonstrate superior performance compared to state-of-the-art quantization methods, including SmoothQuant, OmniQuant, and Atom. The framework achieves a 1.49 inference speedup and significant communication overhead reduction while maintaining or improving model accuracy.

Training Large Language Models (LLMs) on long contexts is severely constrained by prohibitive GPU memory overhead, not training time. The primary culprits are the activations, whose memory footprints scale linearly with sequence length. We introduce OOMB, a highly memory-efficient training system that directly confronts this barrier. Our approach employs a chunk-recurrent training framework with on-the-fly activation recomputation, which maintains a constant activation memory footprint (O(1)) and shifts the primary bottleneck to the growing KV cache. To manage the KV cache, OOMB integrates a suite of synergistic optimizations: a paged memory manager for both the KV cache and its gradients to eliminate fragmentation, asynchronous CPU offloading to hide data transfer latency, and page-level sparse attention to reduce both computational complexity and communication overhead. The synergy of these techniques yields exceptional efficiency. Our empirical results show that for every additional 10K tokens of context, the end-to-end training memory overhead increases by a mere 10MB for Qwen2.5-7B. This allows training Qwen2.5-7B with a 4M-token context on a single H200 GPU, a feat that would otherwise require a large cluster using context parallelism. This work represents a substantial advance in resource efficiency for long-context LLM training. The source code is available at https://github.com/wenhaoli-xmu/OOMB.

Memory layers use a trainable key-value lookup mechanism to add extra parameters to a model without increasing FLOPs. Conceptually, sparsely activated memory layers complement compute-heavy dense feed-forward layers, providing dedicated capacity to store and retrieve information cheaply. This work takes memory layers beyond proof-of-concept, proving their utility at contemporary scale. On downstream tasks, language models augmented with our improved memory layer outperform dense models with more than twice the computation budget, as well as mixture-of-expert models when matched for both compute and parameters. We find gains are especially pronounced for factual tasks. We provide a fully parallelizable memory layer implementation, demonstrating scaling laws with up to 128B memory parameters, pretrained to 1 trillion tokens, comparing to base models with up to 8B parameters.

Nowadays, Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications. However, long context training poses great challenges considering the constraint of GPU memory. It not only leads to substantial activation memory consumption during training, but also incurs considerable memory fragmentation. To facilitate long context training, existing frameworks have adopted strategies such as recomputation and various forms of parallelisms. Nevertheless, these techniques rely on redundant computation or extensive communication, resulting in low Model FLOPS Utilization (MFU). In this paper, we propose MEMO, a novel LLM training framework designed for fine-grained activation memory management. Given the quadratic scaling of computation and linear scaling of memory with sequence lengths when using FlashAttention, we offload memory-consuming activations to CPU memory after each layer's forward pass and fetch them during the backward pass. To maximize the swapping of activations without hindering computation, and to avoid exhausting limited CPU memory, we implement a token-wise activation recomputation and swapping mechanism. Furthermore, we tackle the memory fragmentation issue by employing a bi-level Mixed Integer Programming (MIP) approach, optimizing the reuse of memory across transformer layers. Empirical results demonstrate that MEMO achieves an average of 2.42x and 2.26x MFU compared to Megatron-LM and DeepSpeed, respectively. This improvement is attributed to MEMO's ability to minimize memory fragmentation, reduce recomputation and intensive communication, and circumvent the delays associated with the memory reorganization process due to fragmentation. By leveraging fine-grained activation memory management, MEMO facilitates efficient training of 7B LLM with 1 million sequence length on just 8 A800 GPUs, achieving an MFU of 52.30%.

We rethink test-time scaling laws from a practical efficiency perspective, revealing that the effectiveness of smaller models is significantly overestimated. Prior work, grounded in compute-optimality, overlooks critical memory access bottlenecks introduced by inference-time strategies (e.g., Best-of-N, long CoTs). Our holistic analysis, spanning models from 0.6B to 32B parameters, reveals a new Kinetics Scaling Law that better guides resource allocation by incorporating both computation and memory access costs. Kinetics Scaling Law suggests that test-time compute is more effective when used on models above a threshold than smaller ones. A key reason is that in TTS, attention, rather than parameter count, emerges as the dominant cost factor. Motivated by this, we propose a new scaling paradigm centered on sparse attention, which lowers per-token cost and enables longer generations and more parallel samples within the same resource budget. Empirically, we show that sparse attention models consistently outperform dense counterparts, achieving over 60 points gains in low-cost regimes and over 5 points gains in high-cost regimes for problem-solving accuracy on AIME, encompassing evaluations on state-of-the-art MoEs. These results suggest that sparse attention is essential for realizing the full potential of test-time scaling because, unlike training, where parameter scaling saturates, test-time accuracy continues to improve through increased generation. The code is available at https://github.com/Infini-AI-Lab/Kinetics.

Memory is increasingly central to Large Language Model (LLM) agents operating beyond a single context window, yet most existing systems rely on offline, query-agnostic memory construction that can be inefficient and may discard query-critical information. Although runtime memory utilization is a natural alternative, prior work often incurs substantial overhead and offers limited explicit control over the performance-cost trade-off. In this work, we present BudgetMem, a runtime agent memory framework for explicit, query-aware performance-cost control. BudgetMem structures memory processing as a set of memory modules, each offered in three budget tiers (i.e., Low/Mid/High). A lightweight router performs budget-tier routing across modules to balance task performance and memory construction cost, which is implemented as a compact neural policy trained with reinforcement learning. Using BudgetMem as a unified testbed, we study three complementary strategies for realizing budget tiers: implementation (method complexity), reasoning (inference behavior), and capacity (module model size). Across LoCoMo, LongMemEval, and HotpotQA, BudgetMem surpasses strong baselines when performance is prioritized (i.e., high-budget setting), and delivers better accuracy-cost frontiers under tighter budgets. Moreover, our analysis disentangles the strengths and weaknesses of different tiering strategies, clarifying when each axis delivers the most favorable trade-offs under varying budget regimes.

Pipeline parallelism has been widely explored, but most existing schedules lack a systematic methodology. In this paper, we propose a framework to decompose pipeline schedules as repeating a building block and we show that the lifespan of the building block decides the peak activation memory of the pipeline schedule. Guided by the observations, we find that almost all existing pipeline schedules, to the best of our knowledge, are memory inefficient. To address this, we introduce a family of memory efficient building blocks with controllable activation memory, which can reduce the peak activation memory to 1/2 of 1F1B without sacrificing efficiency, and even to 1/3 with comparable throughput. We can also achieve almost zero pipeline bubbles while maintaining the same activation memory as 1F1B. Our evaluations demonstrate that in pure pipeline parallelism settings, our methods outperform 1F1B by from 7% to 55% in terms of throughput. When employing a grid search over hybrid parallelism hyperparameters in practical scenarios, our proposed methods demonstrate a 16% throughput improvement over the 1F1B baseline for large language models.

The growth rate of the GPU memory capacity has not been able to keep up with that of the size of large language models (LLMs), hindering the model training process. In particular, activations -- the intermediate tensors produced during forward propagation and reused in backward propagation -- dominate the GPU memory use. This leads to high training overhead such as high weight update cost due to the small micro-batch size. To address this challenge, we propose SSDTrain, an adaptive activation offloading framework to high-capacity NVMe SSDs. SSDTrain reduces GPU memory usage without impacting performance by fully overlapping data transfers with computation. SSDTrain is compatible with popular deep learning frameworks like PyTorch, Megatron, and DeepSpeed, and it employs techniques such as tensor deduplication and forwarding to further enhance efficiency. We extensively experimented with popular LLMs like GPT, BERT, and T5. Results demonstrate that SSDTrain reduces 47% of the activation peak memory usage. Meanwhile, SSDTrain perfectly overlaps the I/O with the computation and incurs negligible overhead. Compared with keeping activations in GPU memory and layerwise full recomputation, SSDTrain achieves the best memory savings with negligible throughput loss. We further analyze how the reduced activation memory use may be leveraged to increase throughput by increasing micro-batch size and reducing pipeline parallelism bubbles.

Large Vision-Language Models (LVLMs) enable sophisticated reasoning over images and videos, yet their inference is hindered by a systemic efficiency barrier known as visual token dominance. This overhead is driven by a multi-regime interplay between high-resolution feature extraction, quadratic attention scaling, and memory bandwidth constraints. We present a systematic taxonomy of efficiency techniques structured around the inference lifecycle, consisting of encoding, prefilling, and decoding. Unlike prior reviews focused on isolated optimizations, we analyze the end-to-end pipeline to reveal how upstream decisions dictate downstream bottlenecks, covering compute-bound visual encoding, the intensive prefilling of massive contexts, and the ''visual memory wall'' in bandwidth-bound decoding. By decoupling the efficiency landscape into the axes of shaping information density, managing long-context attention, and overcoming memory limits, this work provides a structured analysis of how isolated optimizations compose to navigate the trade-off between visual fidelity and system efficiency. The survey concludes by outlining four future frontiers supported by pilot empirical insights, including hybrid compression based on functional unit sensitivity, modality-aware decoding with relaxed verification, progressive state management for streaming continuity, and stage-disaggregated serving through hardware-algorithm co-design. Our literature repository is at https://github.com/SuDIS-ZJU/Efficient-LVLMs-Inference.

Memory retention challenges in deep neural architectures have ongoing limitations in the ability to process and recall extended contextual information. Token dependencies degrade as sequence length increases, leading to a decline in coherence and factual consistency across longer outputs. A structured approach is introduced to mitigate this issue through the reweaving of latent states captured at different processing layers, reinforcing token representations over extended sequences. The proposed Contextual Memory Reweaving framework incorporates a Layered Latent State Reconstruction mechanism to systematically integrate past contextual embeddings without introducing external memory modules. Experimental results demonstrate improvements in recall accuracy across a range of sequence lengths, with notable gains in the retention of rarely occurring tokens and numerical reasoning consistency. Further analysis of computational efficiency indicates that the additional processing overhead remains within acceptable thresholds, enabling scalability across different model sizes. Evaluations in long-form text generation and ambiguous query resolution highlight the capacity of memory reweaving to enhance continuity and reduce inconsistencies over extended outputs. Attention weight distributions reveal more structured allocation patterns, suggesting that reweaved latent states contribute to improved contextual awareness. The findings establish a framework for refining memory retention mechanisms in language models, addressing long-standing challenges in handling complex, multi-step reasoning tasks.

With the rapid growth in model size, fine-tuning the large pre-trained language model has become increasingly difficult due to its extensive memory usage. Previous works usually focus on reducing the number of trainable parameters in the network. While the model parameters do contribute to memory usage, the primary memory bottleneck during training arises from storing feature maps, also known as activations, as they are crucial for gradient calculation. Notably, neural networks are usually trained using stochastic gradient descent. We argue that in stochastic optimization, models can handle noisy gradients as long as the gradient estimator is unbiased with reasonable variance. Following this motivation, we propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance, which only requires storing the sub-sampled activations for calculating the gradient. Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones. By replacing the linear operation with our approximated one in transformers, we can achieve up to 2.7times peak memory reduction with almost no accuracy drop and enables up to 6.4times larger batch size. Under the same hardware, WTA-CRS enables better down-streaming task performance by applying larger models and/or faster training speed with larger batch sizes.

Distribution shift (e.g., task or domain shift) in continual learning (CL) usually results in catastrophic forgetting of neural networks. Although it can be alleviated by repeatedly replaying buffered data, the every-step replay is time-consuming. In this paper, we study which modules in neural networks are more prone to forgetting by investigating their training dynamics during CL. Our proposed metrics show that only a few modules are more task-specific and sensitively alter between tasks, while others can be shared across tasks as common knowledge. Hence, we attribute forgetting mainly to the former and find that finetuning them only on a small buffer at the end of any CL method can bring non-trivial improvement. Due to the small number of finetuned parameters, such ``Forgetting Prioritized Finetuning (FPF)'' is efficient in computation. We further propose a more efficient and simpler method that entirely removes the every-step replay and replaces them by only k-times of FPF periodically triggered during CL. Surprisingly, this ``k-FPF'' performs comparably to FPF and outperforms the SOTA CL methods but significantly reduces their computational overhead and cost. In experiments on several benchmarks of class- and domain-incremental CL, FPF consistently improves existing CL methods by a large margin, and k-FPF further excels in efficiency without degrading the accuracy. We also empirically studied the impact of buffer size, epochs per task, and finetuning modules on the cost and accuracy of our methods.

We present a very simple algorithm for attention that requires O(1) memory with respect to sequence length and an extension to self-attention that requires O(log n) memory. This is in contrast with the frequently stated belief that self-attention requires O(n^2) memory. While the time complexity is still O(n^2), device memory rather than compute capability is often the limiting factor on modern accelerators. Thus, reducing the memory requirements of attention allows processing of longer sequences than might otherwise be feasible. We provide a practical implementation for accelerators that requires O(n) memory, is numerically stable, and is within a few percent of the runtime of the standard implementation of attention. We also demonstrate how to differentiate the function while remaining memory-efficient. For sequence length 16384, the memory overhead of self-attention is reduced by 59X for inference and by 32X for differentiation.

Efficiently processing long sequences with Transformer models usually requires splitting the computations across accelerators via context parallelism. The dominant approaches in this family of methods, such as Ring Attention or DeepSpeed Ulysses, enable scaling over the context dimension but do not focus on memory efficiency, which limits the sequence lengths they can support. More advanced techniques, such as Fully Pipelined Distributed Transformer or activation offloading, can further extend the possible context length at the cost of training throughput. In this paper, we present UPipe, a simple yet effective context parallelism technique that performs fine-grained chunking at the attention head level. This technique significantly reduces the activation memory usage of self-attention, breaking the activation memory barrier and unlocking much longer context lengths. Our approach reduces intermediate tensor memory usage in the attention layer by as much as 87.5% for 32B Transformers, while matching previous context parallelism techniques in terms of training speed. UPipe can support the context length of 5M tokens when training Llama3-8B on a single 8timesH100 node, improving upon prior methods by over 25%.

Existing works increasingly adopt memory-centric mechanisms to process long contexts in a segment manner, and effective memory management is one of the key capabilities that enables large language models to effectively propagate information across the entire sequence. Therefore, leveraging reward models (RMs) to automatically and reliably evaluate memory quality is critical. In this work, we introduce MemoryRewardBench, the first benchmark to systematically study the ability of RMs to evaluate long-term memory management processes. MemoryRewardBench covers both long-context comprehension and long-form generation tasks, featuring 10 distinct settings with different memory management patterns, with context length ranging from 8K to 128K tokens. Evaluations on 13 cutting-edge RMs indicate a diminishing performance gap between open-source and proprietary models, with newer-generation models consistently outperforming their predecessors regardless of parameter count. We further expose the capabilities and fundamental limitations of current RMs in evaluating LLM memory management across diverse settings.

Attention for transformers is a critical workload that has recently received significant "attention" as a target for custom acceleration. Yet, while prior work succeeds in reducing attention's memory-bandwidth requirements, it creates load imbalance between attention operators (resulting in severe compute under-utilization) and requires on-chip memory that scales with sequence length (which is expected to grow over time). This paper ameliorates these issues, enabling attention with nearly 100% compute utilization, no off-chip memory traffic bottlenecks, and on-chip buffer size requirements that are independent of sequence length. The main conceptual contribution is to use a recently proposed abstraction -- the cascade of Einsums -- to describe, formalize and taxonomize the space of attention algorithms that appear in the literature. In particular, we show how Einsum cascades can be used to infer non-trivial lower bounds on the number of passes a kernel must take through its input data, which has implications for either required on-chip buffer capacity or memory traffic. We show how this notion can be used to meaningfully divide the space of attention algorithms into several categories and use these categories to inform our design process. Based on the above characterization, we propose FuseMax -- a novel mapping of attention onto a spatial array-style architecture. On attention, in an iso-area comparison, FuseMax achieves an average 6.7times speedup over the prior state-of-the-art FLAT while using 79% of the energy. Similarly, on the full end-to-end transformer inference, FuseMax achieves an average 5.3times speedup over FLAT using 83% of the energy.

Large Language Model (LLM)-based agents are increasingly deployed for complex, tool-based tasks where long-term memory is critical to driving actions. Existing benchmarks, however, primarily test a angent's ability to passively retrieve isolated facts in response to explicit questions. They fail to evaluate the more crucial capability of actively applying memory to execute tasks. To address this gap, we introduce Mem2ActBench, a benchmark for evaluating whether agents can proactively leverage long-term memory to execute tool-based actions by selecting appropriate tools and grounding their parameters. The benchmark simulates persistent assistant usage, where users mention the same topic across long, interrupted interactions and expect previously established preferences and task states to be implicitly applied. We build the dataset with an automated pipeline that merges heterogeneous sources (ToolACE, BFCL, Oasst1), resolves conflicts via consistency modeling, and synthesizes 2,029 sessions with 12 user--assistant--tool turns on average. From these memory chains, a reverse-generation method produces 400 tool-use tasks, with human evaluation confirming 91.3\% are strongly memory-dependent. Experiments on seven memory frameworks show that current systems remain inadequate at actively utilizing memory for parameter grounding, highlighting the need for more effective approaches to evaluate and improve memory application in task execution.

Collocating deep learning training tasks improves GPU utilization but causes drastic slowdowns due to resource contention and risks Out-of-Memory (OOM) failures. Accurate memory estimation is essential for robust collocation, while GPU utilization -- a key proxy for resource contention -- enables interference-aware scheduling to reduce slowdowns and improve throughput. Existing GPU memory estimators span three paradigms -- analytical models, CPU-side libraries, and ML-based estimators -- each with distinct limitations: dependence on detailed model specifications, intrusive integration, poor generalization, and varying latency overhead. GPU heterogeneity further complicates estimation, as identical tasks can exhibit markedly different memory footprints across hardware generations. GPU utilization remains comparatively understudied, further complicated by the non-additive nature of utilization metrics and hardware sensitivity. We conduct a systematic analysis of representative estimators from each paradigm -- Horus, PyTorch FakeTensor, and our lightweight ML-based estimator -- evaluating accuracy, generalizability, and practical overhead. We construct a synthetic dataset spanning MLPs, CNNs, and Transformers with controlled architectural variations, and train MLP- and Transformer-based estimators for memory prediction. We further experiment with utilization estimation on the same dataset. Our evaluation reveals key tradeoffs and validates estimators against real-world unseen models. Significant challenges remain: analytical models are hardware-dependent, CPU-side libraries impose intrusive integration costs, and ML-based estimators struggle with cross-architecture generalization. We release all datasets, tools, and artifacts to support further research.

Neural networks (NNs) are increasingly used in always-on safety-critical applications deployed on hardware accelerators (NN-HAs) employing various memory technologies. Reliable continuous operation of NN is essential for safety-critical applications. During online operation, NNs are susceptible to single and multiple permanent and soft errors due to factors such as radiation, aging, and thermal effects. Explicit NN-HA testing methods cannot detect transient faults during inference, are unsuitable for always-on applications, and require extensive test vector generation and storage. Therefore, in this paper, we propose the uncertainty fingerprint approach representing the online fault status of NN. Furthermore, we propose a dual head NN topology specifically designed to produce uncertainty fingerprints and the primary prediction of the NN in a single shot. During the online operation, by matching the uncertainty fingerprint, we can concurrently self-test NNs with up to 100% coverage with a low false positive rate while maintaining a similar performance of the primary task. Compared to existing works, memory overhead is reduced by up to 243.7 MB, multiply and accumulate (MAC) operation is reduced by up to 10000times, and false-positive rates are reduced by up to 89%.

Deploying large language models to heterogeneous hardware is often constrained by memory, not compute. We introduce MCAP (Monte Carlo Activation Profiling), a load-time per-layer importance estimator that enables dynamic precision and memory placement decisions on the target device. MCAP produces a lightweight per-layer signal that drives both precision dispatch (W4A8 vs. W4A16) and residency tier (GPU, RAM, SSD), allowing a single set of weights to operate across diverse memory budgets. Our system, NVE, achieves 1.5-1.8x higher decode throughput than llama-cpp Q4_0 on NVIDIA T4 and enables models to run in memory regimes previously infeasible without modifying weights.

Large deep learning models have achieved impressive performance across a range of applications. However, their large memory requirements, including parameter memory and activation memory, have become a significant challenge for their practical serving. While existing methods mainly address parameter memory, the importance of activation memory has been overlooked. Especially for long input sequences, activation memory is expected to experience a significant exponential growth as the length of sequences increases. In this approach, we propose AutoChunk, an automatic and adaptive compiler system that efficiently reduces activation memory for long sequence inference by chunk strategies. The proposed system generates chunk plans by optimizing through multiple stages. In each stage, the chunk search pass explores all possible chunk candidates and the chunk selection pass identifies the optimal one. At runtime, AutoChunk employs code generation to automatically apply chunk strategies. The experiments demonstrate that AutoChunk can reduce over 80\% of activation memory while maintaining speed loss within 10%, extend max sequence length by 3.2x to 11.7x, and outperform state-of-the-art methods by a large margin.

In real-world Tool-Integrated Reasoning (TIR) scenarios, where LLMs interleave reasoning with external tool calls, a major source of inefficiency is that the toolcalls create pauses between LLM requests and cause KV-Cache eviction, forcing recomputation. Also, the long, unfiltered response returned by external tools inflates the KV-Cache, so each decode step spends more time loading the growing cache and thus becomes steadily slower as context length increases. However, existing efficiency metrics like token counts and toolcall counts fail to capture the real model inference latency. To address this, we introduce PTE (Prefill Token Equivalents), a hardware-aware TIR-efficiency metric that unifies internal reasoning and external tool-use costs while explicitly accounting for non-reusable KV-Cache and long-tool-response scenarios. Validation in a high-concurrency industrial setting indicates that PTE aligns significantly better with wall-clock latency than standard token counts, while maintaining consistent efficiency rankings across diverse hardware profiles. We conduct extensive experiments across five TIR benchmarks, quantify their PTE costs, and identify four inefficiency patterns that appear in TIR. We also discover that trajectories with higher PTE costs tend to have lower reasoning correctness, indicating that simply using more tools does not improve the quality of the answer.

We propose ERA, a new paradigm that constrains the sampling entropy above given thresholds by applying specially designed activations to the outputs of models. Our approach demonstrates broad effectiveness across different domains: 1) for large language models(LLMs), boosting the AIME 2025 score for Qwen2.5-Math-7B by 37.4%; 2) for continuous control reinforcement learning agents, improving performance by more than 30% over strong baselines such as SAC on the challenging HumanoidBench; 3) for image classification, enhancing ImageNet top-1 accuracy by 0.69% for ResNet-50. These gains are achieved with a computational overhead of less than 7%. Our work validates output activation as a powerful tool for entropy control, opening a new direction for designing simpler and more robust algorithms.

Scientific reasoning relies not only on logical inference but also on activating prior knowledge and experiential structures. Memory can efficiently reuse knowledge and enhance reasoning consistency and stability. However, existing benchmarks mainly evaluate final answers or step-by-step coherence, overlooking the memory-driven mechanisms that underlie human reasoning, which involves activating anchors and attractors, then integrating them into multi-step inference. To address this gap, we propose A^3-Bench~ https://a3-bench.github.io, a benchmark designed to evaluate scientific reasoning through dual-scale memory-driven activation, grounded in Anchor and Attractor Activation. First, we annotate 2,198 science reasoning problems across domains using the SAPM process(subject, anchor & attractor, problem, and memory developing). Second, we introduce a dual-scale memory evaluation framework utilizing anchors and attractors, along with the AAUI(Anchor--Attractor Utilization Index) metric to measure memory activation rates. Finally, through experiments with various base models and paradigms, we validate A^3-Bench and analyze how memory activation impacts reasoning performance, providing insights into memory-driven scientific reasoning.

A growing body of work in cognitive science suggests that reportable conscious access is associated with global ignition over distributed memory systems, while such activation is only partially accessible as individuals cannot directly access or enumerate all activated contents. This tension suggests a plausible mechanism that cognition may rely on a compact representation that approximates the global influence of activation on downstream processing. Inspired by this idea, we introduce the concept of Mindscape Activation Signature (MiA-Signature), a compressed representation of the global activation pattern induced by a query. In LLM systems, this is instantiated via submodular-based selection of high-level concepts that cover the activated context space, optionally refined through lightweight iterative updates using working memory. The resulting MiA-Signature serves as a conditioning signal that approximates the effect of the full activation state while remaining computationally tractable. Integrating MiA-Signatures into both RAG and agentic systems yields consistent performance gains across multiple long-context understanding tasks.

Information retrieval in Large Language Models (LLMs) is increasingly recognized as intertwined with generation capabilities rather than mere lookup. While longer contexts are often assumed to improve retrieval, the effects of intra-context interference remain understudied. To address this, we adapt the proactive interference (PI) paradigm from cognitive science, where earlier information disrupts recall of newer updates. In humans, susceptibility to such interference is inversely linked to working memory capacity. We introduce PI-LLM, an evaluation that sequentially streams semantically related key-value updates and queries only the final values. Although these final values are clearly positioned just before the query, LLM retrieval accuracy declines log-linearly toward zero as interference accumulates; errors arise from retrieving previously overwritten values. Attempts to mitigate interference via prompt engineering (e.g., instructing models to ignore earlier input) yield limited success. These findings reveal a fundamental constraint on LLMs' ability to disentangle interference and flexibly manipulate information, suggesting a working memory bottleneck beyond mere context access. This calls for approaches that strengthen models' ability to suppress irrelevant content during retrieval.

Practical large language model (LLM) services may involve a long system prompt, which specifies the instructions, examples, and knowledge documents of the task and is reused across numerous requests. However, the long system prompt causes throughput/latency bottlenecks as the cost of generating the next token grows w.r.t. the sequence length. This paper aims to improve the efficiency of LLM services that involve long system prompts. Our key observation is that handling these system prompts requires heavily redundant memory accesses in existing causal attention computation algorithms. Specifically, for batched requests, the cached hidden states (i.e., key-value pairs) of system prompts are transferred from off-chip DRAM to on-chip SRAM multiple times, each corresponding to an individual request. To eliminate such a redundancy, we propose RelayAttention, an attention algorithm that allows reading these hidden states from DRAM exactly once for a batch of input tokens. RelayAttention is a free lunch: it maintains the generation quality while requiring no model retraining, as it is based on a mathematical reformulation of causal attention.

We propose a systematic approach to reduce the memory consumption of deep neural network training. Specifically, we design an algorithm that costs O(sqrt(n)) memory to train a n layer network, with only the computational cost of an extra forward pass per mini-batch. As many of the state-of-the-art models hit the upper bound of the GPU memory, our algorithm allows deeper and more complex models to be explored, and helps advance the innovations in deep learning research. We focus on reducing the memory cost to store the intermediate feature maps and gradients during training. Computation graph analysis is used for automatic in-place operation and memory sharing optimizations. We show that it is possible to trade computation for memory - giving a more memory efficient training algorithm with a little extra computation cost. In the extreme case, our analysis also shows that the memory consumption can be reduced to O(log n) with as little as O(n log n) extra cost for forward computation. Our experiments show that we can reduce the memory cost of a 1,000-layer deep residual network from 48G to 7G with only 30 percent additional running time cost on ImageNet problems. Similarly, significant memory cost reduction is observed in training complex recurrent neural networks on very long sequences.

In-context learning is fundamental to modern Large Language Models (LLMs); however, prevailing architectures impose a rigid and fixed contextual structure by assigning linear or constant positional indices. Drawing on Cognitive Load Theory (CLT), we argue that this uninformative structure increases extraneous cognitive load, consuming finite working memory capacity that should be allocated to deep reasoning and attention allocation. To address this, we propose RePo, a novel mechanism that reduces extraneous load via context re-positioning. Unlike standard approaches, RePo utilizes a differentiable module, f_φ, to assign token positions that capture contextual dependencies, rather than replying on pre-defined integer range. By continually pre-training on the OLMo-2 1B backbone, we demonstrate that RePo significantly enhances performance on tasks involving noisy contexts, structured data, and longer context length, while maintaining competitive performance on general short-context tasks. Detailed analysis reveals that RePo successfully allocate higher attention to distant but relevant information, assign positions in dense and non-linear space, and capture the intrinsic structure of the input context. Our code is available at https://github.com/SakanaAI/repo.

Parameter-efficient transfer learning (PETL) aims to reduce the scales of pretrained models for multiple downstream tasks. However, as the models keep scaling up, the memory footprint of existing PETL methods is not significantly reduced compared to the reduction of learnable parameters. This limitation hinders the practical deployment of PETL methods on memory-constrained devices. To this end, we proposed a new PETL framework, called Structure to Activation (S2A), to reduce the memory footprint of activation during fine-tuning. Specifically, our framework consists of: 1) Activation modules design(i.e., bias, prompt and side modules) in the parametric model structure, which results in a significant reduction of adjustable parameters and activation memory; 2) 4-bit quantization of activations based on their derivatives for non-parametric structures (e.g., nonlinear functions), which maintains accuracy while significantly reducing memory usage. Our S2A method consequently offers a lightweight solution in terms of both parameters and memory footprint. We evaluated S2A with different backbones and performed extensive experiments on various datasets to evaluate the effectiveness. The results show that our methods not only outperform existing PETL techniques, achieving a fourfold reduction in GPU memory footprint on average, but also shows competitive performance in accuracy with fewer tunable parameters. These demonstrate that our method is highly suitable for practical transfer learning on hardware-constrained devices.

Long-term memory is crucial for agents in specialized web environments, where success depends on recalling interface affordances, state dynamics, workflows, and recurring failure modes. However, existing memory benchmarks for agents mostly focus on user histories, short traces, or downstream task success, leaving open how to directly evaluate whether memory systems effectively internalize environment-specific experience. To address this gap, we introduce LongMemEval-V2 (LME-V2), a benchmark for evaluating whether memory systems can help agents acquire the experience needed to become knowledgeable colleagues in customized environments. LME-V2 contains 451 manually curated questions covering five core memory abilities for web agents: static state recall, dynamic state tracking, workflow knowledge, environment gotchas, and premise awareness. Questions are paired with history trajectories containing up to 500 trajectories and 115M tokens. We use a context gathering formulation: memory systems consume history trajectories and return compact evidence for downstream question answering. We propose a suite of two memory methods: AgentRunbook-R, an efficient RAG-based memory with knowledge pools for raw state observations, events, and strategy notes, and AgentRunbook-C, which stores trajectories as files and invokes a coding agent to gather evidence in an augmented sandbox. Experiments show that AgentRunbook-C achieves the best performance with 72.5% average accuracy, outperforming the strongest RAG baseline (48.5%) and the off-the-shelf coding agent baseline (69.3%). Despite the strong performance gains, coding agent based methods have high latency costs. While AgentRunbook-C advances the accuracy-latency Pareto frontier, substantial room for improvement remains. Together, these results establish LME-V2 as a challenging testbed for developing long-term memory systems for environment experience.

Federated learning (FL) promotes decentralized training while prioritizing data confidentiality. However, its application on resource-constrained devices is challenging due to the high demand for computation and memory resources to train deep learning models. Neural network pruning techniques, such as dynamic pruning, could enhance model efficiency, but directly adopting them in FL still poses substantial challenges, including post-pruning performance degradation, high activation memory usage, etc. To address these challenges, we propose FedMef, a novel and memory-efficient federated dynamic pruning framework. FedMef comprises two key components. First, we introduce the budget-aware extrusion that maintains pruning efficiency while preserving post-pruning performance by salvaging crucial information from parameters marked for pruning within a given budget. Second, we propose scaled activation pruning to effectively reduce activation memory footprints, which is particularly beneficial for deploying FL to memory-limited devices. Extensive experiments demonstrate the effectiveness of our proposed FedMef. In particular, it achieves a significant reduction of 28.5% in memory footprint compared to state-of-the-art methods while obtaining superior accuracy.

Structured width pruning of GLU-MLP layers, guided by the Maximum Absolute Weight (MAW) criterion, reveals a systematic dichotomy in how reducing the expansion ratio affects different model capabilities. While performance on tasks relying on parametric knowledge (e.g., MMLU, GSM8K) and perplexity metrics degrades predictably, instruction-following capabilities improve substantially (+46% to +75% in IFEval for Llama-3.2-1B and 3B models), and multi-step reasoning remains robust (MUSR). This pattern challenges the prevailing assumption that pruning induces uniform degradation. We evaluated seven expansion ratio configurations using comprehensive benchmarks assessing factual knowledge, mathematical reasoning, language comprehension, instruction-following, and truthfulness. Our analysis identifies the expansion ratio as a critical architectural parameter that selectively modulates cognitive capabilities, rather than merely serving as a compression metric. We provide the first systematic characterization of this selective preservation phenomenon. Notably, we document a robust inverse correlation (r = -0.864, p = 0.012 in Llama-3B) between factual knowledge capacity (MMLU) and truthfulness metrics (TruthfulQA-MC2): as knowledge degrades, the model's ability to discriminate misconceptions improves consistently. This connects two previously distinct research areas, demonstrating that MAW-guided width pruning acts as a selective filter, reducing parametric knowledge while preserving or enhancing behavioral alignment. Additionally, we quantify context-dependent efficiency trade-offs: pruned configurations achieve up to 23% reduction in energy consumption (J/token) but incur penalties in single-request latency, whereas batch processing workloads benefit uniformly.

Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval. This gap is critical: effective assistants must automatically apply learned procedures or avoid failed actions without explicit reminders. We introduce ImplicitMemBench, the first systematic benchmark evaluating implicit memory through three cognitively grounded constructs drawn from standard cognitive-science accounts of non-declarative memory: Procedural Memory (one-shot skill acquisition after interference), Priming (theme-driven bias via paired experimental/control instances), and Classical Conditioning (Conditioned Stimulus--Unconditioned Stimulus (CS--US) associations shaping first decisions). Our 300-item suite employs a unified Learning/Priming-Interfere-Test protocol with first-attempt scoring. Evaluation of 17 models reveals severe limitations: no model exceeds 66% overall, with top performers DeepSeek-R1 (65.3%), Qwen3-32B (64.1%), and GPT-5 (63.0%) far below human baselines. Analysis uncovers dramatic asymmetries (inhibition 17.6% vs. preference 75.0%) and universal bottlenecks requiring architectural innovations beyond parameter scaling. ImplicitMemBench reframes evaluation from "what agents recall" to "what they automatically enact".

Transformer-based large language models (LLMs) exhibit impressive performance in generative tasks but introduce significant challenges in real-world serving due to inefficient use of the expensive, computation-optimized accelerators. This mismatch arises from the autoregressive nature of LLMs, where the generation phase comprises operators with varying resource demands. Specifically, the attention operator is memory-intensive, exhibiting a memory access pattern that clashes with the strengths of modern accelerators, especially as context length increases. To enhance the efficiency and cost-effectiveness of LLM serving, we introduce the concept of attention offloading. This approach leverages a collection of cheap, memory-optimized devices for the attention operator while still utilizing high-end accelerators for other parts of the model. This heterogeneous setup ensures that each component is tailored to its specific workload, maximizing overall performance and cost efficiency. Our comprehensive analysis and experiments confirm the viability of splitting the attention computation over multiple devices. Also, the communication bandwidth required between heterogeneous devices proves to be manageable with prevalent networking technologies. To further validate our theory, we develop Lamina, an LLM inference system that incorporates attention offloading. Experimental results indicate that Lamina can provide 1.48x-12.1x higher estimated throughput per dollar than homogeneous solutions.

Real-world applications require the classification model to adapt to new classes without forgetting old ones. Correspondingly, Class-Incremental Learning (CIL) aims to train a model with limited memory size to meet this requirement. Typical CIL methods tend to save representative exemplars from former classes to resist forgetting, while recent works find that storing models from history can substantially boost the performance. However, the stored models are not counted into the memory budget, which implicitly results in unfair comparisons. We find that when counting the model size into the total budget and comparing methods with aligned memory size, saving models do not consistently work, especially for the case with limited memory budgets. As a result, we need to holistically evaluate different CIL methods at different memory scales and simultaneously consider accuracy and memory size for measurement. On the other hand, we dive deeply into the construction of the memory buffer for memory efficiency. By analyzing the effect of different layers in the network, we find that shallow and deep layers have different characteristics in CIL. Motivated by this, we propose a simple yet effective baseline, denoted as MEMO for Memory-efficient Expandable MOdel. MEMO extends specialized layers based on the shared generalized representations, efficiently extracting diverse representations with modest cost and maintaining representative exemplars. Extensive experiments on benchmark datasets validate MEMO's competitive performance. Code is available at: https://github.com/wangkiw/ICLR23-MEMO

Physical AI systems, including robots, autonomous vehicles, embodied agents and edge copilots, often run a different inference workload from cloud LLM serving: single-stream, batch-1 autoregressive decode, where one robot, camera feed or user session waits on the next token. This workload is usually described as memory-bandwidth-bound. Each decode step streams model weights and the active KV cache, so latency should scale with peak HBM bandwidth. We show that this account is true but incomplete. We measure batch-1 decode for three 7 to 8B-class GQA transformers across four NVIDIA GPUs: H100 SXM5, A100-80GB SXM4, L40S and L4. We evaluate context lengths from 2048 to 16384, producing 44 valid cells under a controlled bf16 SDPA setup. The achieved fraction of peak HBM bandwidth falls as peak bandwidth rises. On the headline Qwen-2.5-7B ctx=2048 cell, an L4 reaches roughly 81 percent of its analytic memory floor, while an H100 reaches only 27 percent. Physical-AI decode is memory-dominated, but faster memory does not translate into proportional latency gains. We test the missing term with a CUDA Graphs A/B experiment. On H100 at ctx=2048, CUDA Graphs improves decode latency by 1.259x across N=10 fresh sessions, with a 95 percent bootstrap confidence interval of 1.253 to 1.267. On L4, the same intervention gives only 1.028x. This isolates a launch-side overhead that becomes visible on fast GPUs but remains mostly hidden on slower, bandwidth-bound GPUs. The deployment implication is that memory savings matter only when the runtime realises them. On L4, bf16 decode sits close to the memory floor, but common quantised paths do not recover the expected 4x weight-traffic reduction: bnb-nf4 reaches 59.36 ms/step and AutoAWQ+Marlin reaches 45.24 ms/step from a 62.32 ms bf16 baseline. GPTQ+ExLlamaV2, with Ada-tuned int4 kernels, reaches 17.36 ms/step.

Recent work has shown that attention-based language models excel at recall, the ability to ground generations in tokens previously seen in context. However, the efficiency of attention-based models is bottle-necked during inference by the KV-cache's aggressive memory consumption. In this work, we explore whether we can improve language model efficiency (e.g. by reducing memory consumption) without compromising on recall. By applying experiments and theory to a broad set of architectures, we identify a key tradeoff between a model's state size and recall ability. We show that efficient alternatives to attention (e.g. H3, Mamba, RWKV) maintain a fixed-size recurrent state, but struggle at recall. We propose BASED a simple architecture combining linear and sliding window attention. By varying BASED window size and linear attention feature dimension, we can dial the state size and traverse the pareto frontier of the recall-memory tradeoff curve, recovering the full quality of attention on one end and the small state size of attention-alternatives on the other. We train language models up to 1.3b parameters and show that BASED matches the strongest sub-quadratic models (e.g. Mamba) in perplexity and outperforms them on real-world recall-intensive tasks by 6.22 accuracy points. Implementations of linear attention are often less efficient than optimized standard attention implementations. To make BASED competitive, we develop IO-aware algorithms that enable 24x higher throughput on language generation than FlashAttention-2, when generating 1024 tokens using 1.3b parameter models. Code for this work is provided at: https://github.com/HazyResearch/based.

Memory-augmented LLM agents store and retrieve information from prior interactions, yet the relative importance of how memories are written versus how they are retrieved remains unclear. We introduce a diagnostic framework that analyzes how performance differences manifest across write strategies, retrieval methods, and memory utilization behavior, and apply it to a 3x3 study crossing three write strategies (raw chunks, Mem0-style fact extraction, MemGPT-style summarization) with three retrieval methods (cosine, BM25, hybrid reranking). On LoCoMo, retrieval method is the dominant factor: average accuracy spans 20 points across retrieval methods (57.1% to 77.2%) but only 3-8 points across write strategies. Raw chunked storage, which requires zero LLM calls, matches or outperforms expensive lossy alternatives, suggesting that current memory pipelines may discard useful context that downstream retrieval mechanisms fail to compensate for. Failure analysis shows that performance breakdowns most often manifest at the retrieval stage rather than at utilization. We argue that, under current retrieval practices, improving retrieval quality yields larger gains than increasing write-time sophistication. Code is publicly available at https://github.com/boqiny/memory-probe.

Large Language Models (LLMs) increasingly use persistent memory from past interactions to enhance personalization and task performance. However, this memory introduces critical risks when sensitive information is revealed in inappropriate contexts. We present CIMemories, a benchmark for evaluating whether LLMs appropriately control information flow from memory based on task context. CIMemories uses synthetic user profiles with over 100 attributes per user, paired with diverse task contexts in which each attribute may be essential for some tasks but inappropriate for others. Our evaluation reveals that frontier models exhibit up to 69% attribute-level violations (leaking information inappropriately), with lower violation rates often coming at the cost of task utility. Violations accumulate across both tasks and runs: as usage increases from 1 to 40 tasks, GPT-5's violations rise from 0.1% to 9.6%, reaching 25.1% when the same prompt is executed 5 times, revealing arbitrary and unstable behavior in which models leak different attributes for identical prompts. Privacy-conscious prompting does not solve this - models overgeneralize, sharing everything or nothing rather than making nuanced, context-dependent decisions. These findings reveal fundamental limitations that require contextually aware reasoning capabilities, not just better prompting or scaling.

The training and inference of large language models (LLMs) are together a costly process that transports knowledge from raw data to meaningful computation. Inspired by the memory hierarchy of the human brain, we reduce this cost by equipping LLMs with explicit memory, a memory format cheaper than model parameters and text retrieval-augmented generation (RAG). Conceptually, with most of its knowledge externalized to explicit memories, the LLM can enjoy a smaller parameter size, training cost, and inference cost, all proportional to the amount of remaining "abstract knowledge". As a preliminary proof of concept, we train from scratch a 2.4B LLM, which achieves better performance than much larger LLMs as well as RAG models, and maintains higher decoding speed than RAG. The model is named Memory^3, since explicit memory is the third form of memory in LLMs after implicit memory (model parameters) and working memory (context key-values). We introduce a memory circuitry theory to support the externalization of knowledge, and present novel techniques including a memory sparsification mechanism that makes storage tractable and a two-stage pretraining scheme that facilitates memory formation.

Robotic manipulation policies have made rapid progress in recent years, yet most existing approaches give limited consideration to memory capabilities. Consequently, they struggle to solve tasks that require reasoning over historical observations and maintaining task-relevant information over time, which are common requirements in real-world manipulation scenarios. Although several memory-aware policies have been proposed, systematic evaluation of memory-dependent manipulation remains underexplored, and the relationship between architectural design choices and memory performance is still not well understood. To address this gap, we introduce RMBench, a simulation benchmark comprising 9 manipulation tasks that span multiple levels of memory complexity, enabling systematic evaluation of policy memory capabilities. We further propose Mem-0, a modular manipulation policy with explicit memory components designed to support controlled ablation studies. Through extensive simulation and real-world experiments, we identify memory-related limitations in existing policies and provide empirical insights into how architectural design choices influence memory performance. The website is available at https://rmbench.github.io/.

In recent years, researchers have focused on reducing the model size and number of computations (measured as "multiply-accumulate" or MAC operations) of DNNs. The energy consumption of a DNN depends on both the number of MAC operations and the energy efficiency of each MAC operation. The former can be estimated at design time; however, the latter depends on the intricate data reuse patterns and underlying hardware architecture. Hence, estimating it at design time is challenging. This work shows that the conventional approach to estimate the data reuse, viz. arithmetic intensity, does not always correctly estimate the degree of data reuse in DNNs since it gives equal importance to all the data types. We propose a novel model, termed "data type aware weighted arithmetic intensity" (DI), which accounts for the unequal importance of different data types in DNNs. We evaluate our model on 25 state-of-the-art DNNs on two GPUs. We show that our model accurately models data-reuse for all possible data reuse patterns for different types of convolution and different types of layers. We show that our model is a better indicator of the energy efficiency of DNNs. We also show its generality using the central limit theorem.

Owing to the huge success of generative artificial intelligence (AI), large language models (LLMs) have emerged as a core subclass, underpinning applications such as question answering, text generation, and code completion. While fine-tuning these models on domain-specific data can yield significant performance gains, it also poses daunting computational challenges, especially for researchers and small organizations with limited hardware resources. Although SSD offloading (i.e., ZeRO-Infinity) has emerged as a viable strategy to overcome the GPU memory barrier via leveraging both system memory (i.e., CPU DRAM) and storage space (i.e., solid-state devices, SSDs), its design primarily targets model-centric performance issues. As a result, key system-level issues, including system memory fragmentation, inefficient pinned buffer allocation, peak CPU usage spikes, and file system overhead, remain unaddressed, stifling scalability and inflating costs. Such an observation motivates this paper to introduce MemAscend, a framework that systematically tackles the underexplored system memory bottlenecks in SSD-offloaded LLM training, with a focus on resource-constrained environments. By streamlining pinned-memory allocation, eradicating fragmentation, and mitigating peak overhead, MemAscend reclaims a substantial system memory budget, enabling larger models, longer context windows, and higher batch sizes without exceeding modest hardware limits. Across diverse LLM benchmarks, MemAscend reduces peak system-memory consumption by an average of 55.7% compared with standard SSD offloading techniques, lowering the hardware barrier for fine-tuning and unlocking new possibilities for cost-effective large-scale training on limited-resource machines.

Deploying large language models (LLMs) on edge devices presents significant challenges due to the substantial computational overhead and memory requirements. Activation sparsification can mitigate these challenges by reducing the number of activated neurons during inference. Existing methods typically employ thresholding-based sparsification based on the statistics of activation tensors. However, these methods do not explicitly model the impact of activation sparsification on performance, leading to suboptimal performance degradation. To address this issue, this paper reformulates the activation sparsification problem by introducing a new objective that optimizes the sparsification decisions. Building on this reformulation, we propose CHESS, a general activation sparsification approach via CHannel-wise thrEsholding and Selective Sparsification. First, channel-wise thresholding assigns a unique threshold to each activation channel in the feed-forward network (FFN) layers. Then, selective sparsification involves applying thresholding-based activation sparsification to specific layers within the attention modules. Finally, we detail the implementation of sparse kernels to accelerate LLM inference. Experimental results demonstrate that the proposed CHESS achieves lower performance degradation over 8 downstream tasks while activating fewer parameters compared to existing methods, thus speeding up the LLM inference by up to 1.27x.

Current mobile GUI agent benchmarks systematically fail to assess memory capabilities, with only 5.2-11.8% memory-related tasks and no cross-session learning evaluation. We introduce MemGUI-Bench, a comprehensive memory-centric benchmark with pass@k and staged LLM-as-judge evaluation. Our contributions include: (1) a systematic memory taxonomy analyzing 11 agents across 5 architectures; (2) 128 tasks across 26 applications where 89.8% challenge memory through cross-temporal and cross-spatial retention; (3) MemGUI-Eval, an automated pipeline with Progressive Scrutiny and 7 hierarchical metrics; and (4) RQ-driven assessment of 11 state-of-the-art agents. Our experiments reveal significant memory deficits across all evaluated systems, identify 5 distinct failure modes, and synthesize 5 actionable design implications. All resources including code, benchmark, and evaluation results will be \textit{fully open-sourced and continuously maintained} at https://lgy0404.github.io/MemGUI-Bench/.

Spiking Neural Networks (SNNs) have been widely praised for their high energy efficiency and immense potential. However, comprehensive research that critically contrasts and correlates SNNs with quantized Artificial Neural Networks (ANNs) remains scant, often leading to skewed comparisons lacking fairness towards ANNs. This paper introduces a unified perspective, illustrating that the time steps in SNNs and quantized bit-widths of activation values present analogous representations. Building on this, we present a more pragmatic and rational approach to estimating the energy consumption of SNNs. Diverging from the conventional Synaptic Operations (SynOps), we champion the "Bit Budget" concept. This notion permits an intricate discourse on strategically allocating computational and storage resources between weights, activation values, and temporal steps under stringent hardware constraints. Guided by the Bit Budget paradigm, we discern that pivoting efforts towards spike patterns and weight quantization, rather than temporal attributes, elicits profound implications for model performance. Utilizing the Bit Budget for holistic design consideration of SNNs elevates model performance across diverse data types, encompassing static imagery and neuromorphic datasets. Our revelations bridge the theoretical chasm between SNNs and quantized ANNs and illuminate a pragmatic trajectory for future endeavors in energy-efficient neural computations.

Personalized agents that interact with users over long periods must maintain persistent memory across sessions and update it as circumstances change. However, existing benchmarks predominantly frame long-term memory evaluation as fact retrieval from past conversations, providing limited insight into agents' ability to consolidate memory over time or handle frequent knowledge updates. We introduce Memora, a long-term memory benchmark spanning weeks to months long user conversations. The benchmark evaluates three memory-grounded tasks: remembering, reasoning, and recommending. To ensure data quality, we employ automated memory-grounding checks and human evaluation. We further introduce Forgetting-Aware Memory Accuracy (FAMA), a metric that penalizes reliance on obsolete or invalidated memory when evaluating long-term memory. Evaluations of four LLMs and six memory agents reveal frequent reuse of invalid memories and failures to reconcile evolving memories. Memory agents offer marginal improvements, exposing shortcomings in long-term memory for personalized agents.

We investigate the origins of massive activations in large language models (LLMs) and identify a specific layer named the Massive Emergence Layer (ME Layer), that is consistently observed across model families, where massive activations first emerge and subsequently propagate to deeper layers through residual connections. We show that, within the ME Layer both the RMSNorm and the FFN parameters jointly contribute to the emergence of massive activations. Once formed, the massive activation token representation remains largely invariant across layers, reducing the diversity of hidden representations passed to the attention module. Motivated by this limitation, we propose a simple and effective method to reduce the rigidity of the massive activation token. Our approach consistently improves LLM performance across multiple tasks, including instruction following and math reasoning, in both training free and fine tuning settings. Moreover, we show that our method mitigates attention sinks by selectively weakening their influence, elucidating their origin at the hidden state level and shedding new light on principled mitigation strategies.

Fine tuning has been regarded as a de facto approach for adapting large language models (LLMs) to downstream tasks, but the high training memory consumption inherited from LLMs makes this process inefficient. Among existing memory efficient approaches, activation-related optimization has proven particularly effective, as activations consistently dominate overall memory consumption. Although prior arts offer various activation optimization strategies, their data-agnostic nature ultimately results in ineffective and unstable fine tuning. In this paper, we propose TokenSeek, a universal plugin solution for various transformer-based models through instance-aware token seeking and ditching, achieving significant fine-tuning memory savings (e.g., requiring only 14.8% of the memory on Llama3.2 1B) with on-par or even better performance. Furthermore, our interpretable token seeking process reveals the underlying reasons for its effectiveness, offering valuable insights for future research on token efficiency. Homepage: https://runjia.tech/iclr_tokenseek/

External memory systems are pivotal for enabling Large Language Model (LLM) agents to maintain persistent knowledge and perform long-horizon decision-making. Existing paradigms typically follow a two-stage process: computationally expensive memory construction (e.g., structuring data into graphs) followed by naive retrieval-augmented generation. However, our empirical analysis reveals two fundamental limitations: complex construction incurs high costs with marginal performance gains, and simple context concatenation fails to bridge the gap between retrieval recall and reasoning accuracy. To address these challenges, we propose CoM (Chain-of-Memory), a novel framework that advocates for a paradigm shift toward lightweight construction paired with sophisticated utilization. CoM introduces a Chain-of-Memory mechanism that organizes retrieved fragments into coherent inference paths through dynamic evolution, utilizing adaptive truncation to prune irrelevant noise. Extensive experiments on the LongMemEval and LoCoMo benchmarks demonstrate that CoM outperforms strong baselines with accuracy gains of 7.5%-10.4%, while drastically reducing computational overhead to approximately 2.7% of token consumption and 6.0% of latency compared to complex memory architectures.

Large language models (LLMs) are central to modern natural language processing, delivering exceptional performance in various tasks. However, their intensive computational and memory requirements present challenges, especially for devices with limited DRAM capacity. This paper tackles the challenge of efficiently running LLMs that exceed the available DRAM capacity by storing the model parameters on flash memory but bringing them on demand to DRAM. Our method involves constructing an inference cost model that harmonizes with the flash memory behavior, guiding us to optimize in two critical areas: reducing the volume of data transferred from flash and reading data in larger, more contiguous chunks. Within this flash memory-informed framework, we introduce two principal techniques. First, "windowing'" strategically reduces data transfer by reusing previously activated neurons, and second, "row-column bundling", tailored to the sequential data access strengths of flash memory, increases the size of data chunks read from flash memory. These methods collectively enable running models up to twice the size of the available DRAM, with a 4-5x and 20-25x increase in inference speed compared to naive loading approaches in CPU and GPU, respectively. Our integration of sparsity awareness, context-adaptive loading, and a hardware-oriented design paves the way for effective inference of LLMs on devices with limited memory.

Large Language Models (LLMs) with billions of parameters have drastically transformed AI applications. However, their demanding computation during inference has raised significant challenges for deployment on resource-constrained devices. Despite recent trends favoring alternative activation functions such as GELU or SiLU, known for increased computation, this study strongly advocates for reinstating ReLU activation in LLMs. We demonstrate that using the ReLU activation function has a negligible impact on convergence and performance while significantly reducing computation and weight transfer. This reduction is particularly valuable during the memory-bound inference step, where efficiency is paramount. Exploring sparsity patterns in ReLU-based LLMs, we unveil the reutilization of activated neurons for generating new tokens and leveraging these insights, we propose practical strategies to substantially reduce LLM inference computation up to three times, using ReLU activations with minimal performance trade-offs.

Parameter-Efficient Fine-Tuning (PEFT) has become the standard for adapting large language models (LLMs). In this work we challenge the wide-spread assumption that parameter efficiency equates memory efficiency and on-device adaptability. We show that this is not true - while methods like LoRA and IA3 significantly reduce trainable parameters, they remain bound by intermediate tensors that scale linearly with sequence length, often triggering out-of-memory errors on-device. In this work, we introduce LARS (Low-memory Activation-Rank Subspace), a novel adaptation framework that decouples memory consumption from sequence length. While prior PEFT methods apply low-rank constraints to model parameters, LARS instead constrains the activation subspace used during training, directly targeting the dominant source of memory consumption and fundamentally flattening the memory growth rate. LARS reduces the memory footprint by an average of 33.54% on GPUs and 51.95% on CPUs in comparison to LoRA across reasoning, understanding and long-context datasets using different models while maintaining competitive accuracy and throughput. Besides GPUs, we deploy on Raspberry Pi and consumer-grade CPUs to demonstrate that LARS provides a scalable path for sophisticated LLM personalization on resource-constrained hardware and edge devices.

While inference-time scaling enables LLMs to carry out increasingly long and capable reasoning traces, the patterns and insights uncovered during these traces are immediately discarded once the context window is reset for a new query. External memory is a natural way to persist these discoveries, and recent work has shown clear benefits for reasoning-intensive tasks. We see an opportunity to make such memories more broadly reusable and scalable by moving beyond instance-based memory entries (e.g. exact query/response pairs, or summaries tightly coupled with the original problem context) toward concept-level memory: reusable, modular abstractions distilled from solution traces and stored in natural language. For future queries, relevant concepts are selectively retrieved and integrated into the prompt, enabling test-time continual learning without weight updates. Our design introduces new strategies for abstracting takeaways from rollouts and retrieving entries for new queries, promoting reuse and allowing memory to expand with additional experiences. We evaluate on ARC-AGI, a benchmark that stresses compositional generalization and abstract reasoning, making it a natural fit for concept memory. Our method yields a 7.5% relative gain over a strong no-memory baseline with performance continuing to scale with inference compute. We find abstract concepts to be the most consistent memory design, outscoring the baseline at all tested inference compute scales. Moreover, dynamically updating memory during test-time outperforms fixed settings, supporting the hypothesis that accumulating and abstracting patterns enables further solutions in a form of self-improvement. Code is available at https://github.com/matt-seb-ho/arc_memo.

The deployment of Large Language Models (LLMs) on edge devices is fundamentally constrained by the "Memory Wall" the bottleneck where data movement latency outstrips arithmetic throughput. Standard inference runtimes often incur significant overhead through high-level abstractions, dynamic dispatch, and unaligned memory access patterns. In this work, we present a novel "Virtual Tensor Core" architecture implemented in software, optimized specifically for ARM64 microarchitectures (Apple Silicon). By bypassing standard library containers in favor of direct memory mapping (mmap) and implementing hand-tuned NEON SIMD kernels, we achieve a form of "Software-Defined Direct Memory Access (DMA)." Our proposed Tensor Virtualization Layout (TVL) guarantees 100% cache line utilization for weight matrices, while our zero-copy loader eliminates initialization latency. Experimental results on a 110M parameter model demonstrate a stable throughput of >60 tokens/second on M2 hardware. While proprietary hardware accelerators (e.g., Apple AMX) can achieve higher peak throughput, our architecture provides a fully open, portable, and deterministic reference implementation for studying the memory bottleneck on general-purpose ARM silicon, meeting the 200ms psycholinguistic latency threshold without opaque dependencies.

Large Language Models (LLMs) are fundamentally constrained by the quadratic computational cost of self-attention and the "Lost in the Middle" phenomenon, where reasoning capabilities degrade as context windows expand. Existing solutions, primarily "Flat RAG" architectures relying on vector databases, treat memory as an unstructured bag of embeddings, failing to capture the hierarchical and temporal structure of long-horizon interactions. This paper presents Aeon, a Neuro-Symbolic Cognitive Operating System that redefines memory as a managed OS resource. Aeon structures memory into a Memory Palace (a spatial index implemented via Atlas, a SIMD-accelerated Page-Clustered Vector Index) and a Trace (a neuro-symbolic episodic graph). This architecture introduces three advances: (1) Symmetric INT8 Scalar Quantization, achieving 3.1x spatial compression and 5.6x math acceleration via NEON SDOT intrinsics; (2) a decoupled Write-Ahead Log (WAL) ensuring crash-recoverability with statistically negligible overhead (<1%); and (3) a Sidecar Blob Arena eliminating the prior 440-character text ceiling via an append-only mmap-backed blob file with generational garbage collection. The Semantic Lookaside Buffer (SLB) exploits conversational locality to achieve sub-5us retrieval latencies, with INT8 vectors dequantized to FP32 on cache insertion to preserve L1-resident lookup performance. Benchmarks on Apple M4 Max demonstrate that the combined architecture achieves 4.70ns INT8 dot product latency, 3.09us tree traversal at 100K nodes (3.4x over FP32), and P99 read latency of 750ns under hostile 16-thread contention via epoch-based reclamation.

Large Language Models (LLMs) have driven significant progress, yet their growing parameter counts and context windows incur prohibitive compute, energy, and monetary costs. We introduce EfficientLLM, a novel benchmark and the first comprehensive empirical study evaluating efficiency techniques for LLMs at scale. Conducted on a production-class cluster (48xGH200, 8xH200 GPUs), our study systematically explores three key axes: (1) architecture pretraining (efficient attention variants: MQA, GQA, MLA, NSA; sparse Mixture-of-Experts (MoE)), (2) fine-tuning (parameter-efficient methods: LoRA, RSLoRA, DoRA), and (3) inference (quantization methods: int4, float16). We define six fine-grained metrics (Memory Utilization, Compute Utilization, Latency, Throughput, Energy Consumption, Compression Rate) to capture hardware saturation, latency-throughput balance, and carbon cost. Evaluating over 100 model-technique pairs (0.5B-72B parameters), we derive three core insights: (i) Efficiency involves quantifiable trade-offs: no single method is universally optimal; e.g., MoE reduces FLOPs and improves accuracy but increases VRAM by 40%, while int4 quantization cuts memory/energy by up to 3.9x at a 3-5% accuracy drop. (ii) Optima are task- and scale-dependent: MQA offers optimal memory-latency trade-offs for constrained devices, MLA achieves lowest perplexity for quality-critical tasks, and RSLoRA surpasses LoRA efficiency only beyond 14B parameters. (iii) Techniques generalize across modalities: we extend evaluations to Large Vision Models (Stable Diffusion 3.5, Wan 2.1) and Vision-Language Models (Qwen2.5-VL), confirming effective transferability. By open-sourcing datasets, evaluation pipelines, and leaderboards, EfficientLLM provides essential guidance for researchers and engineers navigating the efficiency-performance landscape of next-generation foundation models.

In the pursuit of reducing the number of trainable parameters in deep transformer networks, we employ Reinforcement Learning to dynamically select layers during training and tie them together. Every few iterations, the RL agent is asked whether to train each layer i independently or to copy the weights of a previous layer j<i. This facilitates weight sharing, reduces the number of trainable parameters, and also serves as an effective regularization technique. Experimental evaluations validate that our model modestly outperforms the baseline transformer model with regard to perplexity and drastically reduces the number of trainable parameters. In particular, the memory consumption during training is up to one order of magnitude less than the conventional training method.

We revisit the I/O complexity of attention in large language models. Given query-key-value matrices Q,K,VinR^{ntimes d}, and a machine with fast memory size M, the goal is to compute the "attention matrix" A=softmax(Q K ^{top}/d) V with the minimal number of data transfers between fast and slow memory. Existing methods in the literature, most notably FlashAttention and its variants, incur an I/O cost that depends quadratically on n, while a trivial lower bound only requires Ω(nd) I/O's to read the inputs and write the output. In this work, we present a technique for computing attention where the I/O cost only depends almost-linearly on n in most parameter regimes. This is achieved by developing I/O-efficient algorithms inspired by the recent approximate attention framework of Alman and Song. We also prove corresponding lower bounds in each parameter regime to show that our algorithms are indeed close to I/O-optimal.

Transformers have revolutionized the machine learning landscape, gradually making their way into everyday tasks and equipping our computers with "sparks of intelligence". However, their runtime requirements have prevented them from being broadly deployed on mobile. As personal devices become increasingly powerful and prompt privacy becomes an ever more pressing issue, we explore the current state of mobile execution of Large Language Models (LLMs). To achieve this, we have created our own automation infrastructure, MELT, which supports the headless execution and benchmarking of LLMs on device, supporting different models, devices and frameworks, including Android, iOS and Nvidia Jetson devices. We evaluate popular instruction fine-tuned LLMs and leverage different frameworks to measure their end-to-end and granular performance, tracing their memory and energy requirements along the way. Our analysis is the first systematic study of on-device LLM execution, quantifying performance, energy efficiency and accuracy across various state-of-the-art models and showcases the state of on-device intelligence in the era of hyperscale models. Results highlight the performance heterogeneity across targets and corroborates that LLM inference is largely memory-bound. Quantization drastically reduces memory requirements and renders execution viable, but at a non-negligible accuracy cost. Drawing from its energy footprint and thermal behavior, the continuous execution of LLMs remains elusive, as both factors negatively affect user experience. Last, our experience shows that the ecosystem is still in its infancy, and algorithmic as well as hardware breakthroughs can significantly shift the execution cost. We expect NPU acceleration, and framework-hardware co-design to be the biggest bet towards efficient standalone execution, with the alternative of offloading tailored towards edge deployments.

This paper presents MoE-Infinity, a cost-efficient mixture-of-expert (MoE) serving system that realizes activation-aware expert offloading. MoE-Infinity features sequence-level expert activation tracing, a new approach adept at identifying sparse activations and capturing the temporal locality of MoE inference. By analyzing these traces, MoE-Infinity performs novel activation-aware expert prefetching and caching, substantially reducing the latency overheads usually associated with offloading experts for improved cost performance. Extensive experiments in a cluster show that MoE-Infinity outperforms numerous existing systems and approaches, reducing latency by 4 - 20X and decreasing deployment costs by over 8X for various MoEs. MoE-Infinity's source code is publicly available at https://github.com/TorchMoE/MoE-Infinity

Transformer-based long context generative models power emerging AI applications like hour-long video understanding and project-level coding agent. Deploying long context transformers (e.g., 100K to 10M tokens) is prohibitively expensive compared to short context (e.g., 4K tokens) model variants. Reducing the cost of long-context transformers is becoming a pressing research and engineering challenge starting from the year of 2024. This work describes a concurrent programming framework for quantitatively analyzing the efficiency challenges in serving multiple long-context requests under limited size of GPU high-bandwidth memory (HBM) regime. We give a detailed analysis of how all additional computational costs, compared to 4K context, trace back to one single source: the large size of the KV cache. We use a 34B GPT-3.5 level model of 50K context on A100 NVLink as a running example, and describe how its large KV cache causes four types of deployment challenges: (1) prefilling long inputs takes much longer compute time and GPU memory than short inputs; (2) after prefilling, the large KV cache residing on the GPU HBM substantially restricts the number of concurrent users being served; (3) during decoding, repeatedly reading the KV cache from HBM to SM largely increases latency; (4) when KV cache memory overflows, swapping it from HBM to DDR causes significant context switching latency. We use this framework to analyze existing works and identify possibilities of combining them to build end-to-end systems. Overall, this work offers a foundational framework for analyzing long context transformer deployment and identifies directions towards reducing the inference cost of 1M context to be as cheap as 4K.

This paper introduces a structured memory which can be easily integrated into a neural network. The memory is very large by design and significantly increases the capacity of the architecture, by up to a billion parameters with a negligible computational overhead. Its design and access pattern is based on product keys, which enable fast and exact nearest neighbor search. The ability to increase the number of parameters while keeping the same computational budget lets the overall system strike a better trade-off between prediction accuracy and computation efficiency both at training and test time. This memory layer allows us to tackle very large scale language modeling tasks. In our experiments we consider a dataset with up to 30 billion words, and we plug our memory layer in a state-of-the-art transformer-based architecture. In particular, we found that a memory augmented model with only 12 layers outperforms a baseline transformer model with 24 layers, while being twice faster at inference time. We release our code for reproducibility purposes.

While Mixture of Experts (MoE) models achieve remarkable efficiency by activating only subsets of parameters, they suffer from high memory access costs during inference. Memory-layer architectures offer an appealing alternative with very few memory access, but previous attempts like UltraMem have only matched the performance of 2-expert MoE models, falling significantly short of state-of-the-art 8-expert configurations. We present UltraMemV2, a redesigned memory-layer architecture that closes this performance gap. Our approach introduces five key improvements: integrating memory layers into every transformer block, simplifying value expansion with single linear projections, adopting FFN-based value processing from PEER, implementing principled parameter initialization, and rebalancing memory-to-FFN computation ratios. Through extensive evaluation, we demonstrate that UltraMemV2 achieves performance parity with 8-expert MoE models under same computation and parameters but significantly low memory access. Notably, UltraMemV2 shows superior performance on memory-intensive tasks, with improvements of +1.6 points on long-context memorization, +6.2 points on multi-round memorization, and +7.9 points on in-context learning. We validate our approach at scale with models up to 2.5B activated parameters from 120B total parameters, and establish that activation density has greater impact on performance than total sparse parameter count. Our work brings memory-layer architectures to performance parity with state-of-the-art MoE models, presenting a compelling alternative for efficient sparse computation.

Agentic memory systems enable large language model (LLM) agents to maintain state across long interactions, supporting long-horizon reasoning and personalization beyond fixed context windows. Despite rapid architectural development, the empirical foundations of these systems remain fragile: existing benchmarks are often underscaled, evaluation metrics are misaligned with semantic utility, performance varies significantly across backbone models, and system-level costs are frequently overlooked. This survey presents a structured analysis of agentic memory from both architectural and system perspectives. We first introduce a concise taxonomy of MAG systems based on four memory structures. Then, we analyze key pain points limiting current systems, including benchmark saturation effects, metric validity and judge sensitivity, backbone-dependent accuracy, and the latency and throughput overhead introduced by memory maintenance. By connecting the memory structure to empirical limitations, this survey clarifies why current agentic memory systems often underperform their theoretical promise and outlines directions for more reliable evaluation and scalable system design.

Modern language models increasingly adopt hybrid architectures that combine full attention with efficient attention modules, such as sliding-window attention (SWA) and recurrent sequence mixers. However, how these efficient modules shape model capabilities remains poorly understood. To address this gap, we conduct a systematic analysis across hybrid architectures from three perspectives: scaling behavior, mechanism analysis, and architecture design. First, from a scaling perspective, we find that efficient-attention design primarily affects how fast long-context capability emerges, while different hybrids eventually converge to comparable long-context performance under sufficient training. Second, mechanistically, we show that long-range retrieval is mainly carried by full attention, whereas efficient attention shapes its optimization trajectory. This explains a counter-intuitive phenomenon we call Large-Window Laziness: larger SWA windows can delay the formation of retrieval heads in full-attention layers. Third, guided by this mechanism, we show that applying NoPE to only the full-attention layers of a small-window SWA hybrid substantially improves long-context performance with negligible impact on short-context performance.

Real-world agents operate over long and evolving horizons, where information is repeatedly updated and may interfere across memories, requiring accurate recall and aggregated reasoning over multiple pieces of information. However, existing benchmarks focus on static, independent recall and fail to capture these dynamic interactions between evolving memories. In this paper, we study how current memory-augmented agents perform in realistic, interference-heavy, long-horizon settings across diverse domains and question types. We introduce MINTEval (Long-Horizon Memory under INTerference Evaluation), a benchmark featuring (1) long, highly interconnected contexts with frequently updated information that induces substantial interference, (2) diverse domains (state tracking, multi-turn dialogue, Wikipedia revisions, and GitHub commits), enabling evaluation of domain generalization, and (3) diverse question types that assess robustness to interference, including (i) single-target recall tasks requiring retrieval of a specific target from long contexts, and (ii) multi-target aggregation tasks requiring reasoning over multiple relevant pieces of information. Overall, MINTEval has 15.6k question-answering pairs over long-horizon contexts averaging 138.8k tokens and extending up to 1.8M tokens per instance. We evaluate 7 representative systems, including vanilla long-context LLMs, RAG, and memory-augmented agent frameworks. Across all systems, we observe consistently low performance (avg. 27.9% accuracy), especially on questions requiring aggregated reasoning over multiple pieces of evidence. Our analysis shows that performance is primarily limited by retrieval and memory construction. Furthermore, current memory systems struggle to recall and reason over earlier facts that are revised or interfered with by subsequent context, with accuracy degrading as the number of intervening updates increases.

Large Language Models (LLMs) have showcased remarkable capabilities across various domains. Accompanying the evolving capabilities and expanding deployment scenarios of LLMs, their deployment challenges escalate due to their sheer scale and the advanced yet complex activation designs prevalent in notable model series, such as Llama, Gemma, and Mistral. These challenges have become particularly pronounced in resource-constrained deployment scenarios, where mitigating inference efficiency bottlenecks is imperative. Among various recent efforts, activation approximation has emerged as a promising avenue for pursuing inference efficiency, sometimes considered indispensable in applications such as private inference. Despite achieving substantial speedups with minimal impact on utility, even appearing sound and practical for real-world deployment, the safety implications of activation approximations remain unclear. In this work, we fill this critical gap in LLM safety by conducting the first systematic safety evaluation of activation approximations. Our safety vetting spans seven sota techniques across three popular categories, revealing consistent safety degradation across ten safety-aligned LLMs.

We introduce a comprehensive benchmark for conversational memory evaluation containing 75,336 question-answer pairs across diverse categories including user facts, assistant recall, abstention, preferences, temporal changes, and implicit connections. While existing benchmarks have advanced the field, our work addresses fundamental challenges in statistical power, data generation consistency, and evaluation flexibility that limit current memory evaluation frameworks. We examine the relationship between conversational memory and retrieval-augmented generation (RAG). While these systems share fundamental architectural patterns--temporal reasoning, implicit extraction, knowledge updates, and graph representations--memory systems have a unique characteristic: they start from zero and grow progressively with each conversation. This characteristic enables naive approaches that would be impractical for traditional RAG. Consistent with recent findings on long context effectiveness, we observe that simple full-context approaches achieve 70-82% accuracy even on our most challenging multi-message evidence cases, while sophisticated RAG-based memory systems like Mem0 achieve only 30-45% when operating on conversation histories under 150 interactions. Our analysis reveals practical transition points: long context excels for the first 30 conversations, remains viable with manageable trade-offs up to 150 conversations, and typically requires hybrid or RAG approaches beyond that point as costs and latencies become prohibitive. These patterns indicate that the small-corpus advantage of conversational memory--where exhaustive search and complete reranking are feasible--deserves dedicated research attention rather than simply applying general RAG solutions to conversation histories.

Machine unlearning, the ability for a machine learning model to forget, is becoming increasingly important to comply with data privacy regulations, as well as to remove harmful, manipulated, or outdated information. The key challenge lies in forgetting specific information while protecting model performance on the remaining data. While current state-of-the-art methods perform well, they typically require some level of retraining over the retained data, in order to protect or restore model performance. This adds computational overhead and mandates that the training data remain available and accessible, which may not be feasible. In contrast, other methods employ a retrain-free paradigm, however, these approaches are prohibitively computationally expensive and do not perform on par with their retrain-based counterparts. We present Selective Synaptic Dampening (SSD), a novel two-step, post hoc, retrain-free approach to machine unlearning which is fast, performant, and does not require long-term storage of the training data. First, SSD uses the Fisher information matrix of the training and forgetting data to select parameters that are disproportionately important to the forget set. Second, SSD induces forgetting by dampening these parameters proportional to their relative importance to the forget set with respect to the wider training data. We evaluate our method against several existing unlearning methods in a range of experiments using ResNet18 and Vision Transformer. Results show that the performance of SSD is competitive with retrain-based post hoc methods, demonstrating the viability of retrain-free post hoc unlearning approaches.

Large Language models (LLMs) have become a research hotspot. To accelerate the inference of LLMs, storing computed caches in memory has become the standard technique. However, as the inference length increases, growing KV caches might lead to out-of-memory issues. Many existing methods address this issue through KV cache compression, primarily by preserving key tokens throughout all layers to reduce information loss. Most of them allocate a uniform budget size for each layer to retain. However, we observe that the minimum budget sizes needed to retain essential information vary across layers and models based on the perspectives of attention and hidden state output. Building on this observation, this paper proposes a simple yet effective KV cache compression method that leverages layer uncertainty to allocate budget size for each layer. Experimental results show that the proposed method can reduce memory usage of the KV caches to only sim20\% when compared to Full KV inference while achieving nearly lossless performance.

Memory benchmarks for LLM agents largely assume single-user settings, leaving shared assistants for hospitals, workplaces, campuses, and households understudied. In these deployments, multiple principals write to a common memory pool and query it under different roles, scopes, and relationships, so memory quality requires governance as well as recall. We introduce GateMem, a benchmark for multi-principal shared-memory agents. GateMem jointly evaluates utility for legitimate long-horizon requests with state updates, access control across contextual authorization boundaries, and agent-facing active forgetting after explicit deletion requests. It spans medical, office, education, and household domains, with long-form multi-party episodes, incremental memory injection, hidden checkpoints, structured judging, and leak-target annotations. Across diverse baselines and backbone models, no method simultaneously achieves strong utility, robust access control, and reliable forgetting. Long-context prompting often yields the best governance score at high token cost, while retrieval-based and external-memory methods reduce cost yet still leak unauthorized or deleted information. These results show current memory agents remain far from reliable shared institutional deployment.

Memory is fundamental to intelligence, enabling learning, reasoning, and adaptability across biological and artificial systems. While Transformer architectures excel at sequence modeling, they face critical limitations in long-range context retention, continual learning, and knowledge integration. This review presents a unified framework bridging neuroscience principles, including dynamic multi-timescale memory, selective attention, and consolidation, with engineering advances in Memory-Augmented Transformers. We organize recent progress through three taxonomic dimensions: functional objectives (context extension, reasoning, knowledge integration, adaptation), memory representations (parameter-encoded, state-based, explicit, hybrid), and integration mechanisms (attention fusion, gated control, associative retrieval). Our analysis of core memory operations (reading, writing, forgetting, and capacity management) reveals a shift from static caches toward adaptive, test-time learning systems. We identify persistent challenges in scalability and interference, alongside emerging solutions including hierarchical buffering and surprise-gated updates. This synthesis provides a roadmap toward cognitively-inspired, lifelong-learning Transformer architectures.

The full-size MLPs and the projection layers in attention introduce tremendous model sizes of large language models (LLMs), consuming extensive computational resources in pre-training. We empirically observe that the activations of pre-trained LLMs exhibit low-rank property. Motivated by such observations, we propose CoLA and its memory-efficient implementation, CoLA-M, to replace these full-size layers with compute-efficient auto-encoders that naturally enforce low-rank activations throughout training. This fundamental architectural change eliminates the activation redundancy and significantly boosts model capacity and training efficiency. Experiments on LLaMA models with 60 million to 7 billion parameters show that CoLA reduces the computing cost by bf 2times and improves training throughput by bf 1.86times while maintaining full-rank level performance. CoLA-M further squeezes memory cost without sacrificing throughput, offering a pre-training approach with collectively superior parameter, computing, and memory efficiency. The LLMs produced are also bf 2times smaller, enabling faster inference with lower memory cost on resource-constrained platforms.

Memory is a fundamental component of AI systems, underpinning large language models (LLMs) based agents. While prior surveys have focused on memory applications with LLMs, they often overlook the atomic operations that underlie memory dynamics. In this survey, we first categorize memory representations into parametric, contextual structured, and contextual unstructured and then introduce six fundamental memory operations: Consolidation, Updating, Indexing, Forgetting, Retrieval, and Compression. We systematically map these operations to the most relevant research topics across long-term, long-context, parametric modification, and multi-source memory. By reframing memory systems through the lens of atomic operations and representation types, this survey provides a structured and dynamic perspective on research, benchmark datasets, and tools related to memory in AI, clarifying the functional interplay in LLMs based agents while outlining promising directions for future researchThe paper list, datasets, methods and tools are available at \href{https://github.com/Elvin-Yiming-Du/Survey_Memory_in_AI{https://github.com/Elvin-Yiming-Du/Survey\_Memory\_in\_AI}.}.

The growing prevalence of Large Language Models (LLMs) and their substantial memory requirements have prompted renewed interest in CPU offloading as a method to compensate for limited GPU memory. In particular, when CPU memory is leveraged to temporarily store intermediate states of LLMs, CPU memory becomes a new bottleneck and soon reaches the capacity limitation of commodity CPUs. In this work, we investigate the effectiveness of Compute Express Link (CXL) add-in card (AIC) memory as an extension to CPU memory, enabling larger model sizes and longer context lengths during fine-tuning. Through extensive benchmarking, this study quantifies the performance overhead introduced by transferring data between CXL memory, CPU, and GPUs, focusing on how concurrency and data volume influence bandwidth utilization and latency. This study also compares CPUbased optimizer steps when model parameters, gradients, and optimizer states reside in local memory versus CXL memory, revealing that naive adoption of CXL often degrades performance during the optimizer phase. To overcome these challenges, this study proposes a CXL-aware allocation to strategically partition CPU offloading workloads across both local and CXL memory. This study further demonstrates that employing multiple AICs significantly reduces bandwidth contention, thus improving scalability. Experimental results show that these optimizations enable efficient long-context LLM fine-tuning, underscoring CXL as a promising avenue for unlocking the full potential of CPU offloading in long-context LLM fine-tuning.

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8timesA800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to 2.8times and 2.0times in the prefill and decode stages.

Deep neural network (DNN) inference has become an important part of many data-center workloads. This has prompted focused efforts to design ever-faster deep learning accelerators such as GPUs and TPUs. However, an end-to-end DNN-based vision application contains more than just DNN inference, including input decompression, resizing, sampling, normalization, and data transfer. In this paper, we perform a thorough evaluation of computer vision inference requests performed on a throughput-optimized serving system. We quantify the performance impact of server overheads such as data movement, preprocessing, and message brokers between two DNNs producing outputs at different rates. Our empirical analysis encompasses many computer vision tasks including image classification, segmentation, detection, depth-estimation, and more complex processing pipelines with multiple DNNs. Our results consistently demonstrate that end-to-end application performance can easily be dominated by data processing and data movement functions (up to 56% of end-to-end latency in a medium-sized image, and sim 80% impact on system throughput in a large image), even though these functions have been conventionally overlooked in deep learning system design. Our work identifies important performance bottlenecks in different application scenarios, achieves 2.25times better throughput compared to prior work, and paves the way for more holistic deep learning system design.

Memory and computation remain core bottlenecks in long-horizon LLM inference due to the quadratic cost of self-attention and the ever-growing key-value (KV) cache. Existing strategies for memory-bounded inference, such as quantization, offloading, or heuristic KV eviction, either incur high orchestration costs or rely on unreliable attention-based proxies of importance. We propose TRIM-KV, a novel approach that learns each token's intrinsic importance at creation time via a lightweight retention gate. Each gate predicts a scalar retention score that decays over time, reflecting the long-term utility of the token for a specific layer and head. Tokens with low scores are evicted when the memory budget is exceeded, ensuring that the cache always contains the most critical tokens. TRIM-KV is trained efficiently through distillation from a frozen LLM combined with a capacity loss, requiring only gate fine-tuning and adding negligible inference overhead. Across mathematical reasoning (GSM8K, MATH-500, AIME24), procedural generation (LongProc), conversational long-memory benchmarks (LongMemEval), and long-context understanding (LongBench and SCBench), TRIM-KV consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes. Remarkably, it even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization, suppressing noise from uninformative tokens. Qualitative analyses further reveal that learned retention scores align with human intuition, naturally recovering heuristics such as sink tokens, sliding windows, and gist compression without explicit design. Beyond efficiency, retention scores provide insights into layer- and head-specific roles, suggesting a new path toward LLM interpretability.

High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2-4times with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm

As LLM-based agents are increasingly used in long-term interactions, cumulative memory is critical for enabling personalization and maintaining stylistic consistency. However, most existing systems adopt an ``all-or-nothing'' approach to memory usage: incorporating all relevant past information can lead to Memory Anchoring, where the agent is trapped by past interactions, while excluding memory entirely results in under-utilization and the loss of important interaction history. We show that an agent's reliance on memory can be modeled as an explicit and user-controllable dimension. We first introduce a behavioral metric of memory dependence to quantify the influence of past interactions on current outputs. We then propose Steerable Memory Agent, SteeM, a framework that allows users to dynamically regulate memory reliance, ranging from a fresh-start mode that promotes innovation to a high-fidelity mode that closely follows interaction history. Experiments across different scenarios demonstrate that our approach consistently outperforms conventional prompting and rigid memory masking strategies, yielding a more nuanced and effective control for personalized human-agent collaboration.

Long-horizon interactions between users and LLM-based assistants necessitate effective memory management, yet current approaches face challenges in training and evaluation of memory. Existing memory benchmarks rely on static, off-policy data as context, limiting evaluation reliability and scalability. To address these gaps, we introduce AMemGym, an interactive environment enabling on-policy evaluation and optimization for memory-driven personalization. AMemGym employs structured data sampling to predefine user profiles, state-dependent questions, and state evolution trajectories, enabling cost-effective generation of high-quality, evaluation-aligned interactions. LLM-simulated users expose latent states through role-play while maintaining structured state consistency. Comprehensive metrics based on structured data guide both assessment and optimization of assistants. Extensive experiments reveal performance gaps in existing memory systems (e.g., RAG, long-context LLMs, and agentic memory) and corresponding reasons. AMemGym not only enables effective selection among competing approaches but also can potentially drive the self-evolution of memory management strategies. By bridging structured state evolution with free-form interactions, our framework provides a scalable, diagnostically rich environment for advancing memory capabilities in conversational agents.

The research of artificial intelligence is undergoing a paradigm shift from prioritizing model innovations over benchmark scores towards emphasizing problem definition and rigorous real-world evaluation. As the field enters the "second half," the central challenge becomes real utility in long-horizon, dynamic, and user-dependent environments, where agents face context explosion and must continuously accumulate, manage, and selectively reuse large volumes of information across extended interactions. Memory, with hundreds of papers released this year, therefore emerges as the critical solution to fill the utility gap. In this survey, we provide a unified view of foundation agent memory along three dimensions: memory substrate (internal and external), cognitive mechanism (episodic, semantic, sensory, working, and procedural), and memory subject (agent- and user-centric). We then analyze how memory is instantiated and operated under different agent topologies and highlight learning policies over memory operations. Finally, we review evaluation benchmarks and metrics for assessing memory utility, and outline various open challenges and future directions.

Distributed inference of large language models (LLMs) can introduce overheads of up to 20% even over GPUs connected via high-speed interconnects such as NVLINK. Multiple techniques have been proposed to mitigate these overheads by decomposing computations into finer-grained tasks and overlapping communication with sub-tasks as they complete. However, fine-grained decomposition of a large computation into many smaller computations on GPUs results in overheads. Further, the communication itself uses many streaming multiprocessors (SMs), adding to the overhead. We present TokenWeave to address these challenges. TokenWeave proposes a Token-Splitting technique that divides the tokens in the inference batch into two approximately equal subsets in a wave-aware manner. The computation of one subset is then overlapped with the communication of the other. In addition, TokenWeave optimizes the order of the layer normalization computation with respect to communication operations and implements a novel fused AllReduce-RMSNorm kernel carefully leveraging Multimem instruction support available on NVIDIA Hopper GPUs. These optimizations allow TokenWeave to perform communication and RMSNorm using only 2-8 SMs. Moreover, our kernel enables the memory bound RMSNorm to be overlapped with the other batch's computation, providing additional gains. Our evaluations demonstrate up to 29% latency gains and up to 26% throughput gains across multiple models and workloads. In several settings, TokenWeave results in better performance compared to an equivalent model with all communication removed.

Training large language models requires distributing computation across many accelerators, yet practitioners select parallelism strategies (data, tensor, pipeline, ZeRO) through trial and error because no unified systematic framework predicts their behavior. We introduce placement semantics: each strategy is specified by how it places four training states (parameters, optimizer, gradients, activations) across devices using five modes (replicated, sharded, sharded-with-gather, materialized, offloaded). From placement alone, without implementation details, we derive memory consumption and communication volume. Our predictions match published results exactly: ZeRO-3 uses 8x less memory than data parallelism at 1.5x communication cost, as reported in the original paper. We prove two conditions (gradient integrity, state consistency) are necessary and sufficient for distributed training to match single-device results, and provide composition rules for combining strategies safely. The framework unifies ZeRO Stages 1-3, Fully Sharded Data Parallel (FSDP), tensor parallelism, and pipeline parallelism as instances with different placement choices.

As context windows in LLMs scale to 100K+ tokens, the key-value (KV) cache becomes the dominant memory bottleneck, with recent methods claiming 80-90% savings and minimal benchmark degradation. We argue these evaluations miss a structural issue: attention is not just storage but routing, and retaining KV pairs does not guarantee semantic accessibility. We propose a physics-inspired view of KV compression as a controlled perturbation of token-level routing, distinguishing retention, accessibility, and utilization. Using synthetic tasks probing multi-entity tracking, disambiguation, coreference, and multi-hop reasoning, we find that moderate compression degrades internal representations with little accuracy loss, revealing redundancy; all models exhibit a sharp hallucination safety cliff near 90% compression, correlated with spikes in Global Eviction Ratio (GER), suggesting a phase transition in semantic reachability; and architectures differ in routing dynamics, with LLaMA showing early consensus and late diversification, and Qwen showing funnel-like late convergence, leading to distinct resilience profiles. Beyond erasure, we identify representational rigidity, where excessive head-level consensus collapses routing flexibility despite token survival. These results suggest sparse token-route structures govern compression tolerance, reframing KV compression as a structural probe of attention geometry and linking long-context scalability to sparsity and the lottery ticket hypothesis in self-attention.

As multi-modal models advance towards long-form video understanding, memory emerges as a critical capability. Despite substantial efforts in developing video datasets and benchmarks, existing works primarily focus on perception and reasoning, without systematically evaluating memory: what models retain, how faithfully information is preserved, and how robust memory remains under interference. To address this gap, we introduce M^3Eval, the first comprehensive evaluation framework and benchmark for probing different memory dimensions in multi-modal models. Grounded in cognitive psychology, our design features carefully constructed tasks that isolate key aspects of memory. Leveraging M^3Eval, we conduct extensive experiments across representative multi-modal models, revealing consistent weaknesses and distinctive behaviors. We find that models struggle to maintain disentangled representations when processing parallel video streams, exhibit interference patterns differing substantially from those observed in human memory, ground memory sources more reliably in the spatial domain than the temporal domain, and demonstrate limited symbolic memory. Collectively, our benchmark provides a valuable resource for future research, while our findings highlight memory as a fundamental yet underexplored capability and offer insights for designing more effective memory mechanisms in multi-modal models. Our code and dataset are available at https://pku-value-lab.github.io/m3eval-homepage.

Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, performance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new insights and practical strategies for improving resource utilization, particularly for smaller language models.