Leandro von Werra PRO

lvwerra's activity

• All
• Models
• Datasets
• Spaces
• Buckets
• Papers
• Collections
• Community
• Posts
• Upvotes
• Likes
• Articles

New activity in rl-llm-wiki/knowledge-base 9 days ago

source: url:interconnects.ai/p/the-state-of-reasoning — State of reasoning (speculation/calibration)

4

#722 opened 25 days ago by

lvwerra

grpo §4: integrate 2025 soft-source framing on the sharpen-vs-expand debate (SSA shaping, RL scaling laws)

2

#792 opened 23 days ago by

lvwerra

source: url:garymarcus.substack.com/p/alphaproof-alphageometry-chatgpt — AlphaProof/AlphaGeometry neurosymbolic framing (DeepMind, op-ed)

3

#759 opened 23 days ago by

lvwerra

deepen generative-rm: coverage vs precision, the generative RM as the ceiling on repeated sampling (§7b)

3

#769 opened 23 days ago by

lvwerra

deepen self-correction-rl: the positive result, RL makes intrinsic self-correction work (SCoRe) (§4b)

3

#780 opened 23 days ago by

lvwerra

source: url:newsletter.semianalysis.com/p/rl-environments-and-rl-for-science — RL-environments industry map / data foundries (analyst, speculation)

3

#786 opened 23 days ago by

lvwerra

tool-use-rl §6b: integrate 2025 practitioner soft-sources (Will Brown multi-turn RL, Verifier's law)

2

#789 opened 23 days ago by

lvwerra

reward-attacks §4b: integrate soft-source framing (spec-gaming vs reward-optimization, Turner/LessWrong)

2

#790 opened 23 days ago by

lvwerra

source: url:interconnects.ai/p/reverse-engineering-openai-o1 — Reverse engineering OpenAI's o1 (speculation)

3

#718 opened 25 days ago by

lvwerra

source: url:interconnects.ai/p/openais-o1-using-search-was-a-psyop — o1 search was a PSYOP (speculation)

3

#719 opened 25 days ago by

lvwerra

source: url:semianalysis.com/scaling-laws-o1-pro-architecture — SemiAnalysis o1-pro / lab training (speculation)

3

#720 opened 25 days ago by

lvwerra

source: url:interconnects.ai/p/deepseek-r1-recipe-for-o1 — DeepSeek-R1 recipe / o1 read-across (speculation)

3

#721 opened 25 days ago by

lvwerra

source: url:interconnects.ai/p/gpt-5-and-bending-the-arc-of-progress — GPT-5 arc of progress (speculation)

3

#723 opened 25 days ago by

lvwerra

deepen rm-reliability: ODIN and the supervision spectrum for confound removal (§3b, §6b)

3

#737 opened 23 days ago by

lvwerra

deepen ppo-in-practice: PPO-max / Secrets-of-RLHF-I, which impl details are load-bearing (§2b, §7b)

3

#746 opened 23 days ago by

lvwerra

deepen sft-vs-rl-boundary: on-policy vs off-policy supervision (GKD, MiniLLM) as a second axis the boundary thins on (§4b, §7b)

3

#753 opened 23 days ago by

lvwerra

deepen unified-offline-po: the divergence axis (f-DPO) as the second orthogonal knob of the family (§1b, §4b)

3

#756 opened 23 days ago by

lvwerra

deepen pluralistic-preference-optimization: attribute conditioning (SteerLM) as a third route (§4b, §7b)

3

#757 opened 23 days ago by

lvwerra

source: url:yuanchaofa.com/post/kimi-k2-5-reading-notes — Kimi K2.5 PARL parallel-agent RL deep-read (CN, speculation)

4

#744 opened 23 days ago by

lvwerra

source: url:hub.baai.ac.cn/view/43156 — BAAI: how R1 got reasoning (zh, speculation)

3

#726 opened 25 days ago by

lvwerra