Simon Yu

The future of RL+LLM? Self-play. Why? Competitive scenarios offer: ✅ Built-in verification ✅ Automated curriculum learning ✅ Infinite complexity scaling Games prove this works for multi-turn, multi-agent systems. But the real potential? Extending beyond games to real-world…

Kimi K2. More evidence that: • The lead of American “frontier” AI companies is rather small 🔬 • A broad ecosystem of strong foundation model companies is developing in China, with more players than in the US or elsewhere. 🐅 moonshotai.github.io/Kimi-K2/ github.com/MoonshotAI/Kim…

LLMs Encode Harmfulness and Refusal Separately Jiachen Zhao (@jcz12856876), Jing Huang, Zhengxuan Wu (@ZhengxuanZenWu), @davidbau, Weiyan Shi (@shi_weiyan)

For the first time it feels within reach to go from the current status quo of single-turn oracles to generally intelligent, general purpose multi-turn agents. Very happy to help organize this NeurIPS workshop (lead by @simon_ycl ). If you are working on something related,…

this is basically self-contained + very concise but still should have pretty great perf github.com/ChenmienTan/RL2

>>> Qwen3-Coder is here! ✅ We’re releasing Qwen3-Coder-480B-A35B-Instruct, our most powerful open agentic code model to date. This 480B-parameter Mixture-of-Experts model (35B active) natively supports 256K context and scales to 1M context with extrapolation. It achieves…

The world is moving towards agents Static benchmarks don't measure what agents do best (multi-turn reasoning) Thus, interactive benchmarks: * Terminal Bench (@alexgshaw, @Mike_A_Merrill) * Text Arena (@LeonGuertler) * BALROG (@PaglieriDavide, @_rockt) * ARC-AGI-3 (@arcprize)

LLMs encode 'this is harmful' and 'I should refuse' as two different things. Why this matters 👇 check the amazing work by my labmate @jcz12856876!

1/ 🚨New Paper 🚨 LLMs are trained to refuse harmful instructions, but internally, do they see harmfulness and refusal as the same? ⚔️We find causal evidence that 👈”LLMs encode harmfulness and refusal separately” 👉. ✂️LLMs may know a prompt is harmful internally yet still…

Pleased to share that our Multi-Turn Interactions in LLMs workshop at NeurIPS 2025! …shop-multi-turn-interaction.github.io Welcome work on Multi-Turn RL, multi-turn human<->agent/agent<->agent/agent<->environment interactions, multi-turn tool use, multi-turn alignment, multi-turn evaluation,…

Excited to share that our multi-turn interactions workshop has been accepted by neurips!

🤔Long-horizon tasks: How to train LLMs for the marathon?🌀 Submit anything on 🔁"Multi-turn Interactions in LLMs"🔁 to our @NeurIPSConf workshop by 08/22: 📕 Multi-Turn RL ⚖️ Multi-Turn Alignment 💬 Multi-Turn Human-AI Teaming 📊 Multi-Turn Eval ♾️You name it! #neurips #LLM

Interested in Multi-turn RL, Agent, Alignment, and Interactions with LLMs? Join and submit your paper on our multi-turn interaction workshop at #NeurIPS2025 !

Join us at #NeurIPS2025 workshop to explore the future of multi-turn AI interactions! We welcome submissions on RL for agents, alignment, evaluation methods, and more.

We’re excited to share that our work “Diversify and Conquer: Diversity-Centric Data Selection with Iterative Refinement” will be highlighted today at ICML 2025 in Vancouver, Canada! 🇨🇦 🎉Congrats to all authors! @simon_ycl, @cliangyu_, Sara Ahmadian, @mziizm,

Love this clever combination of SFT + RL for distribution shift 🔥 Check it out!

The RL codebase I like the most: - The NanoGPT of RL - Supports multi-turn RL - Just 1k lines of code in Python - Data, Tensor, Sequence Parallel github.com/ChenmienTan/RL2

Excited to announce the Mindgame @NeurIPS Competition is officially LIVE! 🤖 Pit your agents against others in Mafia, Codename, Prisoner’s Dilemma, Stg Hunt, and Colonel Blotto. Sign up now for $500 in compute credits on your initial run! 🔗 Register : mindgamesarena.com

two quick reasons why I'm not convinced that latent reasoning > CoT: - if that worked, we would have been doing non-autoregressive decoding for ages - in arxiv.org/abs/2011.03459 we found that grounding intermediate symbols works better than having fuzzy embeddings when reasoning

Every few years I think you should jump into a new deep-end by doing something that pushes you out of your comfort zone. However, people rarely do because it is easier to stick with what is known. I think this is true in research, life partners and big life decisions.

ICML ✈️ this week. open to chat and learn mech interp from you. @aryaman2020 and i have cool ideas about steering, just come to our AxBench poster. new steering blog: zen-wu.social/steer/index.ht… 中文: zen-wu.social/steer/cn_index…