Nanyun (Violet) Peng @ ACL26
837 posts
@VioletNPeng
Associated Professor@UCLA (leave), Scientist and manager@Google. AI and Infra, AI for science, AI creativity, event. (she/her/hers). Opinions are my own.
Big congrats 🎉🎊🍾 to UCLA PLUS lab @VioletNPeng and collaborators won 3! !outstanding paper awards at #EMNLP2024 👏👏👏
Does GRPO handle multiple-task advantage normalization effectively? 🤔 🚀We introduce Gaussian GRPO (G²RPO), a novel method that mathematically forcing the advantage distribution of any given task to strictly converge to a standard normal distribution. G²RPO provides: ✅1) intrinsic robustness to outliers, ✅2) symmetric updates for positive and negative rewards, ✅3) uniform variance across diverse tasks. We adopt G²RPO and task-level response length and entropy shaping which leads to OpenVLThinkerV2! 🏆A new model with SOTA open-source visual reasoning and perception performance among the similar size models. (1/n)👇 #VLM #LLM #Multimodal #RL #GRPO
🏆Outstanding Paper is awarded to the following research papers: The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions
Another great paper from Google. Shows general LLMs can solve formal math by planning proofs and checking each step. Raised general LLM performance from under 10% to 70%. A general LLM failed badly when asked to write full formal proofs in 1 try, but became much stronger when it planned, split the work into smaller claims, reused past claims, and learned from Lean’s feedback. The paper shows the weakness was not just the model’s math ability, but the way it was being used - the absence of structured interaction with a verifier. The key idea is that the model does not try to write one giant perfect proof at once, because that usually fails on long and tricky problems. Instead, LEAP stores the proof as a graph of goals and subgoals, so useful lemmas can be reused instead of rediscovered every time. The authors tested LEAP on Putnam 2025 and a new Lean benchmark built from 60 IMO-style problems, where ordinary one-shot proof writing did very poorly. LEAP solved all 12 Putnam 2025 problems and raised general LLM performance on the Lean IMO benchmark from under 10% to 70%. ---- Link – arxiv. org/abs/2606.03303 Title: "LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks"
Another great paper from Google. Shows general LLMs can solve formal math by planning proofs and checking each step. Raised general LLM performance from under 10% to 70%. A general LLM failed badly when asked to write full formal proofs in 1 try, but became much stronger when it planned, split the work into smaller claims, reused past claims, and learned from Lean’s feedback. The paper shows the weakness was not just the model’s math ability, but the way it was being used - the absence of structured interaction with a verifier. The key idea is that the model does not try to write one giant perfect proof at once, because that usually fails on long and tricky problems. Instead, LEAP stores the proof as a graph of goals and subgoals, so useful lemmas can be reused instead of rediscovered every time. The authors tested LEAP on Putnam 2025 and a new Lean benchmark built from 60 IMO-style problems, where ordinary one-shot proof writing did very poorly. LEAP solved all 12 Putnam 2025 problems and raised general LLM performance on the Lean IMO benchmark from under 10% to 70%. ---- Link – arxiv. org/abs/2606.03303 Title: "LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks"
Another great paper from Google. Shows general LLMs can solve formal math by planning proofs and checking each step. Raised general LLM performance from under 10% to 70%. A general LLM failed badly when asked to write full formal proofs in 1 try, but became much stronger when it planned, split the work into smaller claims, reused past claims, and learned from Lean’s feedback. The paper shows the weakness was not just the model’s math ability, but the way it was being used - the absence of structured interaction with a verifier. The key idea is that the model does not try to write one giant perfect proof at once, because that usually fails on long and tricky problems. Instead, LEAP stores the proof as a graph of goals and subgoals, so useful lemmas can be reused instead of rediscovered every time. The authors tested LEAP on Putnam 2025 and a new Lean benchmark built from 60 IMO-style problems, where ordinary one-shot proof writing did very poorly. LEAP solved all 12 Putnam 2025 problems and raised general LLM performance on the Lean IMO benchmark from under 10% to 70%. ---- Link – arxiv. org/abs/2606.03303 Title: "LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks"
🚨Model editing in practice often collapses with catastrophic forgetting! Meet SPHERE🌐: an energy-regularized method that keeps weights uniformly distributed on hyperspheres, making sequential editing stable. Paper: arxiv.org/abs/2510.01172 Code: github.com/PlusLabNLP/SPH…
🚨 You can bypass ALL safety guardrails of GPT-OSS-120B 🚨❗🤯 How? By detecting behavior-associated experts and switching them on/off. 📄 Steering MoE LLMs via Expert (De)Activation 🔗 arxiv.org/abs/2509.09660 🧵👇
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.
Accelerating Science track: A century of scientific progress in one decade - that's the target. 👑Accelerating the Queen of Sciences, @UCLA Teaching AI to wonder, conjecture, and discover the way a mathematician does. Every hard science benefits if this works. @sahaiamit @kaiwei_chang Raghu Meka @VioletNPeng @terrence_tao @WeiWang1973 🌦️Actionable AI Weather Forecasts for Developing Economies, @UChicago Open-source AI weather forecasting infrastructure for developing economies, with millions of farmers already reached with better monsoon forecasts. @WillettBecca @ianfoster @turbulentjet Michael Kremer
Multilingual Routing in Mixture-of-Experts LLMs We present (1) an in-depth analysis of how MoE LLMs route multilingual texts, with very clear patterns + (2) a router intervention (steering) method that leads to consistent multilingual improvements! 🧵1/4 arxiv.org/pdf/2510.04694
(2/4) Nanyun (Violet) Peng @VioletNPeng @UCLA is developing MAP-SE, a Multi-Agent Persuasion Simulation Engine that studies how influence emerges across adaptive agents with long-term memory—moving beyond simple one-on-one interactions. #AIResearch #AgenticAI
Today we introduce humans&, a human-centric frontier AI lab. We believe AI can be reimagined, centering around people and their relationships with each other. At its best, AI should serve as a deeper connective tissue that strengthens organizations and communities
