Yao Dou retweeted
Grok 4.7 is here.
It's a notable improvement over Grok 4.6 at the same price and speed.
Glad to share that our paper Gavel was accepted to EMNLP 2026 Main! 🎉
We study frontier LLMs on legal case summarization with inputs exceeding 500K tokens. Models struggle beyond 256K tokens and are much more likely to omit key information than hallucinate.
We also built Gavel-Agent, an agentic, reference-free evaluator that searches source documents for evidence.
More details: yao-dou.github.io/gavel/
Yao Dou retweeted
We are recruiting reviewers for the Evaluation of Interactive Agents workshop at NeurIPS Atlanta this year! If you are interested, please fill out the reviewer interest form :D
📅 Workshop date: December 12 or 13 (to be confirmed)
⏰ Submission deadline: August 29, AoE
🌐 Workshop website: eval-interactive-agents-work…
📝 Submit via OpenReview: openreview.net/group?id=Neur…
👥 Interested in reviewing? Sign up here: docs.google.com/forms/d/e/1F…
Yao Dou retweeted
We're organizing the Workshop on Evaluation of Interactive Agents at NeurIPS 2026!
Consider submitting your work (non-archival) and joining us in Atlanta.
Topics include (not limited to): evaluation benchmarks, trajectory-level metrics, user/environment simulation, long-horizon reliability, safety, and ethics.
⏰ Deadline: Aug. 29 (AoE)
🔗 Website: eval-interactive-agents-work…
#NeurIPS
We are excited to announce Evaluation of Interactive Agents Workshop (IAEval) at NeurIPS 2026 in Atlanta.
As AI agents become more capable and engage with users on long, open-ended tasks, static benchmarks are no longer enough. We need evaluation methods that capture how agents behave in realistic, multi-turn interactions.
If your research is relevant, we invite you to submit. More details below.
🤖 Made with AI
📅 Workshop date: December 12 or 13 (to be confirmed)
⏰ Submission deadline: August 29, AoE
🌐 Workshop website: eval-interactive-agents-work…
📝 Submit via OpenReview: openreview.net/group?id=Neur…
👥 Interested in reviewing? Sign up here: docs.google.com/forms/d/e/1F…
Yao Dou retweeted
Excited to attend #ICML2026 in Seoul 🇰🇷 together with my Ph.D. students: @EthanMendes3, @hyungjoochae, @CherylolGuo!
We’ll be presenting work on:
- Multilingual reasoning
- Self-distillation from expert reasoning
- Cloning synthetic web environments
Poster info + paper links below. Looking forward to catching up!
Yao Dou retweeted
Knowledge is not evenly distributed across languages. For a given question, what if better information comes from another language?
🔥 We propose Language-Routed Policy Optimization (LRPO), an online reinforcement learning framework that lets LLMs explore and learn from cross-lingual generations during training.
#ICML2026
Yao Dou retweeted
Real-world LLM interactions are conversational, yet LLMs are far more vulnerable to multi-turn than single-turn attacks.
Introducing DialTree: RL + tree search for automated multi-turn attack discovery.
📊 81.5% avg ASR@1 across 12 LLMs
📈 +44.2% over prior SOTA
#ICLR2026 [1/6]
Yao Dou retweeted
Nice images😃. Now back to agents. Excited to share a dedicated agent research lab just came out of stealth! Our mission never changed and our path towards it is clearer than ever: Building the agent learning for any specialized domains. We are hiring, come join us!🚀 neocognition.io/
Introducing @NeoCognition, the agent lab for specialized intelligence.
Everyone needs experts, but human expertise does not scale.
Backed by $40M seed funding, we build self-learning agents that specialize across domains to make expertise abundant.
RT @alan_ritter: New paper by my Ph.D. student @hyungjoochae: How can we build a safe, scalable learning environment for web agents?
https…
Yao Dou retweeted
We released Nemotron Cascade 2 30B A3B.
What makes this release especially meaningful to me is that it reflects a 1.5-year journey at NVIDIA around one core idea: improving AI math reasoning through self-improvement at test time.
Each project tackled a different part of that problem.
With AceMath (24 Q4), we built an external verifier model to identify the right solution during test-time scaling.
With AceReason (24 Q1-2), we scaled the reasoning capabilities of the model through RL so the model could spend more time reflecting while solving problems. Along the way, we found a general, simple and effective RL recipe that we’ve kept using since.
And now with Cascade 2 (25 Q1), we’ve pushed that effort further: the model can generate hypotheses, verify them, and refine them on its own. That self-improvement loop is what enabled IMO gold-level performance at 30B level.
From MATH500, to AIME, and now IMO Proof.
This team is THE BEST.
Technical report:
huggingface.co/papers/2603.1…
🚀 Introducing Nemotron-Cascade 2 🚀
Just 3 months after Nemotron-Cascade 1, we’re releasing Nemotron-Cascade 2: an open 30B MoE with 3B active parameters, delivering best-in-class reasoning and strong agentic capabilities.
🥇 Gold Medal-level performance on IMO 2025, IOI 2025, and ICPC World Finals 2025:
• Capabilities once thought achievable only by frontier proprietary models (e.g. Gemini Deep Think) or frontier-scale open models (i.e. DeepSeek-V3.2-Speciale-671B-A37B).
• Remarkably high intelligence density with 20× fewer parameters.
🏆 Best-in-class across math, code reasoning, alignment, and instruction following:
• Outperforms the latest Qwen3.5-35B-A3B (2026-02-24) and even larger Qwen3.5-122B-A10B (2026-03-11).
🧠 Powered by Cascade RL + multi-domain on-policy distillation:
• Significantly expand Cascade RL across a much broader range of reasoning and agentic domains than Nemotron-Cascade 1, while distilling from the strongest intermediate teacher models throughout training to recover regressions and sustain gains.
🤗 Model + SFT + RL data:
👉 huggingface.co/collections/n…
📄 Technical report:
👉 research.nvidia.com/labs/nem…
Yao Dou retweeted
🥈 Silver Medal at IOI 2025 & Outperforms DeepSeek-R1-0528 on LiveCodeBench.
Instead of mixing different tasks together, we scale *Cascade RL* to develop general LLMs in curriculum (RLFH -> Instruct -> Math -> Code -> SWE).
So many learnings, check out our report!👇
🚀 Introducing Nemotron-Cascade! 🚀
We’re thrilled to release Nemotron-Cascade, a family of general-purpose reasoning models trained with cascaded, domain-wise reinforcement learning (Cascade RL), delivering best-in-class performance across a wide range of benchmarks.
💻 Coding powerhouse
After RL, our 14B model:
• Surpasses DeepSeek-R1-0528 (671B) on LiveCodeBench v5/v6/Pro.
• Achieves silver-medal performance at IOI 2025 🥈.
• Reaches a 43.1% pass@1 on SWE-Bench Verified, and 53.8% with test-time scaling.
🧠 What is Cascade RL?
Instead of mixing heterogeneous prompts across domains, Cascade RL trains sequentially, domain by domain, which reduces engineering complexity, mitigates heterogeneous verification latencies, and enables domain-specific curricula and tailored hyperparameter tuning.
✨ Key insight
Using RLHF for alignment as a pre-step dramatically boosts complex reasoning—far beyond preference optimization. Subsequent domain-wise RLVR stages rarely hurt the benchmark performance attained in earlier domains and may even improve it, as illustrated in the following figure.
🤗 Models & training data 🔥
👉 huggingface.co/collections/n…
📄 Technical report with detailed training and data recipes
👉 arxiv.org/pdf/2512.13607
Yao Dou retweeted
Introducing early experience: using future states resulting from agent’s own action as scalable supervision to train itself - without reward🧠!
1️⃣Reward-free: can train directly in real-world environments.
2️⃣Better RL warm-start: when continued with RL, leads to higher final performance than imitation-only warm-ups.
3️⃣Data-efficient & scalable: outperforms imitation with even 1/8 data.👇
🌀Agent Learning via Early Experience🌀
📝: arxiv.org/abs/2510.08558
- SFT for agents is sparse; RL on long-horizons is hard
We provide new mid-training signals that work:
1) Implicit next state world modeling task
2) Self-reflection on alternate states
- Strong improvements over 8 environments and multiple model families
- Works well for subsequent RL!
🧵1/5
Can LLM-simulated users replace expensive human evaluation for multi-turn conversations? Short answer: yes, if you model the user right.
With our SimulatorArena, we find that detailed user profiles (knowledge + message style) improve alignment with real human evaluation by 26% at <3% the cost.
#EMNLP2025 [1/6] 🧵
[6/6]
🌐Website: simulatorarena.ai/
📄Paper: arxiv.org/abs/2510.05444
💻Github: github.com/microsoft/Simulat…
Huge thanks to my amazing coauthors—Michel Galley, Baolin Peng, Chris Kedzie, Weixin Cai, @alan_ritter, Chris Quirk, @cocoweixu, @JianfengGao0217—for an incredible collaboration at @MSFTResearch and @ICatGT.