Excited to share our new work on Hybrid Reinforcement (HERO) — combining verifiable and reward-model signals for reasoning RL. Verifiers are precise but brittle. Reward models are rich but noisy. In our new paper HERO, we show how to combine both!

Human preference data is noisy: inconsistent labels, annotator bias, etc. No matter how fancy the post-training algorithm is, bad data can sink your model. 🔥 @Samuel861025 and I are thrilled to release PrefCleanBench — a systematic benchmark for evaluating data cleaning

💃New Multi-Agent RL Method: WaltzRL💃 📝: https://t.co/KE8dM9kX1r - Makes LLM safety a positive-sum game between a conversation & feedback agent - At inference feedback is adaptive, used when needed -> Improves safety & reduces overrefusals without degrading capabilities! 🧵1/5

Your LVLM says: “There’s a cat on the table.” But… there’s no cat in the image. Not even a whisker. This is object hallucination — one of the most persistent reliability failures in multi-modal language models. Our new #NeurIPS2025 paper introduces GLSim, a simple but

Thrilled to announce our #NeurIPS2025 paper on LENS (Latent EmbeddiNg Synthesis)! LENS learns to synthesize preference pairs directly in latent space — 🚀 18× faster data generation ⚙️ 16 000× smaller model Efficient, theoretically grounded, and scalable reward modeling.

Excited to share our #NeurIPS2025 paper: Visual Instruction Bottleneck Tuning (Vittle) Multimodal LLMs do great in-distribution, but often break in the wild. Scaling data or models helps, but it’s costly. 💡 Our work is inspired by the Information Bottleneck (IB) principle,

🌀New Self-Driven RL Method: RESTRAIN 🌀 📝: https://t.co/x4EgHfxZfG - RESTRAIN turns spurious votes → self-Improving signals. No labels needed - Does this through self-penalizing unreliable reasoning paths: ✔️ Uses all rollouts, not just the majority, ✔️ Offsets

🌀New work: Era of Real-World Human Interaction 🌀 📝: https://t.co/qB1roKk6Ou - RL *directly* from User Conversations - Organic replies + long-term history are learning signal - Trained on WildChat, beats RLHF at *user* level -> the future for personal Super Intelligence? 🧵1/6

🌀Diversity Aware RL (DARLING)🌀 📝: https://t.co/MH0tui34Cb - Jointly optimizes for quality & diversity using a learned partition function - Outperforms standard RL in quality AND diversity metrics, e.g. higher pass@1/p@k - Works for both non-verifiable & verifiable tasks 🧵1/5

🪜Introducing: StepWiser🦉 📝: https://t.co/RXOjaMjHI1 - Reframes stepwise reward modeling as a reasoning task: outputs CoT + judgment. - Trained by RL using relative outcomes of rollouts. Results: (1) SOTA performance on ProcessBench! (2) Improves policy at train time. (3)

BloomScrub🧽 is now accepted to EMNLP 2025 as a main conference paper! Check out our post below for a detailed summary⬇️

A smarter way to teach LLMs. Meta and NYU propose CoT-Self-Instruct: a new method where LLMs first reason & plan via Chain-of-Thought (CoT), then generate high-quality synthetic prompts, filtered with automatic metrics. 🧠 Outperforms s1k/OpenMath on MATH500, AMC23, AIME24,

Congratulations! My best advisor!!!

Excited to be in Vancouver for ICML2025! I'll be presenting "Position: Challenges and Future Directions of Data-Centric AI Alignment" in East Exhibition Hall A-B #E-601 on Tuesday, 7/15, from 4:30 pm. Please come if you are interested in AI alignment! #ICML2025 #aialignment

🌍 GeoArena is live! Evaluate how well large vision-language models (LVLMs) understand the world through image geolocalization. Help us compare models via human preference — your feedback matters! 🔗 Try it now: https://t.co/x5oe9B26tu #GeoArena #Geolocation #LVLM #AI

🚨 If you care about reliable, low-cost LLM hallucination detection, our #ICML2025 paper offers a powerful and data-efficient solution. 💡We introduce TSV: Truthfulness Separator Vector — a single vector injected into a frozen LLM that reshapes its hidden space to better

🚨 We’re hiring! The Radio Lab @ NTU Singapore is looking for PhD, master, undergrads, RAs, and interns to build responsible AI & LLMs. Remote/onsite from 2025. Interested? Email us: radiolab.ntu.recruiting@gmail.com 🔗 https://t.co/377UaRL9Ic Please spread the word if you can!

Excited to join the College of Computing and Data Science at Nanyang Technological University, Singapore (@NTUsg) as an Assistant Professor this fall! 🙌 Grateful to my advisor @SharonYixuanLi and all who supported me along the way. Looking forward to the new chapter! 😄 🇸🇬

📢 Looking for new research ideas in AI alignment? Check out our new #ICML2025 position paper: "Challenges and Future Directions of Data-Centric AI Alignment". TL;DR: Aligning powerful AI systems isn't just about better algorithms — it's also about better feedback data, whether

Excited to be in Singapore for ICLR 2025! We’ll be presenting “Your Weak LLM is Secretly a Strong Teacher for Alignment “ in Hall 3 + Hall 2B on April 24th from 3:00 p.m. If you’re interested in LLM post-training or AI safety, come by! Looking forward to meeting everyone!