OpenAI's blog ( https://t.co/Mu05PFfPXg) points out that today’s language models hallucinate because training and evaluation reward guessing instead of admitting uncertainty. This raises a natural question: can we reduce hallucination without hurting utility?🤔 On-policy RL with

Super happy to receive the Best Paper Award at #NeurIPS2025 for our Artificial Hivemind paper!! (Really enjoyed giving oral talk at NeurIPS as well!)

I'll be at #NeurIPS2025 until 12/7! I work on post-training and reward signals (Spurious Rewards), currently curious about bridging the gap between how humans and LLMs learn. Looking forward to connecting with new and old friends—also exploring summer 2025 internships. DMs open!

I will be at #NeurIPS2025 12.3–12.7 Looking forward to meeting old and new friends ! ☕️🌮 Recently working on hallucination (Binary RAR) and verbatim memorization (ParaPO), issues that scaling up pretraining cannot simply fix. Also interested in making models learn more like

8B model can outperform AlphaEvolve on open optimization problems by scaling compute for inference or test-time RL🚀! ⭕Circle packing: AlphaEvolve (Gemini-2.0-Flash/Pro) : 2.63586276 Ours (DeepSeek-R1-0528-Qwen3-8B) : 2.63598308 🔗in🧵 [1/n]

PhD applicants — Join Akari’s first cohort of students! Akari's research ranges from careful benchmarking to solid methodology. She always gives sharp feedback while being thoughtful and supportive. She stayed driven throughout her PhD and now brings that same energy to her new

Exciting DR Tulu updates! 📈 DR Tulu-8B (new RL ckpt) sits on the performance–cost frontier, beating Tongyi DR-30B and matching OpenAI DR/Gemini 3 Pro+Search at a fraction of the cost. Now on arXiv. 🖥️ You can run an interactive CLI demo with open code, almost for free. 1/🧵

Olmo3 is here! 🎉 Fully open data and fully open training recipes again. 🚀 Huge congrats to the whole team!

🔥Thrilled to introduce DR Tulu-8B, an open long-form Deep Research model that matches OpenAI DR 💪Yes, just 8B! 🚀 The secret? We present Reinforcement Learning with Evolving Rubrics (RLER) for long-form non-verifiable DR tasks! Our rubrics: - co-evolve with the policy model -

New blog post out! 📜 We share our latest research efforts to build more effective, human-centered AI collaboration. Months ago, I was genuinely surprised by how quickly AI agents were improving, and with that came a deep fear of being replaced, of humans slowly losing agency as

📝Paper: https://t.co/7pxLXwN7Jb 💻Github: https://t.co/nw4ztyuxqx Huge thanks to the amazing collaborators who made this work possible. @AkariAsai @LukeZettlemoyer @HannaHajishirzi @faeze_brh [7/n]

💡 Finding 2: Binary reward works better than continuous reward. Continuous rewards such as VeriScore are vulnerable to reward hacking because the model can raise the score by adding correct but irrelevant claims or by adapting to formats favored by the claim extractor and

In short-form QA, binary RAR sharply cuts incorrect answers while keeping accuracy unchanged when the model is allowed to express uncertainty. Standard binary rewards for verifiable tasks give 1 only for correct answers and 0 for both incorrect and abstaining, which trains

💡 Finding 1: Utility stays unchanged. A model can reach zero hallucination by always saying “I do not know,” but then it is not useful. Our approach avoids this. In long-form generation, Binary RAR keeps the number of correct claims unchanged and reduces incorrect claims. [4/n]

With Binary RAR, we cut hallucinations in both long-form generation (61.9→37.5) and short-form question answering (60.6→27.6), while core skills remain unchanged. Interestingly, after RL finetuning with binary RAR, the model keeps the same accuracy when forced to answer on

Binary Retrieval-Augmented Reward (Binary RAR) gives a reward of one only when the verifier finds no contradiction between the model output and retrieved evidence, and zero otherwise. No partial credit and little reward hacking. With this stepwise signal, the KL penalty in RL

RL is bounded by finite data😣? Introducing RLVE: RL with Adaptive Verifiable Environments We scale RL with data procedurally generated from 400 envs dynamically adapting to the trained model 💡find supervision signals right at the LM capability frontier + scale them 🔗in🧵

Our infini-gram mini paper received the Best Paper Award at #EMNLP2025 !! Really proud 🥹

Model collaboration talk tour continues~ Compositional intelligence. Collaborative development. Decentralized AI. By the Many. The methods. The vision. The hot takes. The comedy. If you are around one of these places, let's chat!

How to build agentic search systems for long-horizon tasks? Check out our new paper! - Simple design principles are efficient and effective - Error analysis and fine-grain analysis for search systems A 🧵 on SLIM, our long-horizon agentic search framework