Progress of AI is bottlenecked by the quality of evaluation, motivating the need for powerful and generalist LLM judges that can think and reason. Here's our latest paper, J1, on how to train such Thinking-LLM-Judges with RL. 🧵👇.

Check out our new paper where we compared offline and (Semi-)Online DPO with GRPO for post-training LLMs. This led to some interesting findings! 👇.

RT @dair_ai: 3. J1. Introduces a novel training approach for LLMs to act as evaluators (LLM-as-a-Judge) by explicitly incentivizing thought….

RT @johnschulman2: For people who don't like Claude's behavior here (and I think it's totally valid to disagree with it), I encourage you t….

RT @rohanpaul_ai: Evaluation of LLMs is difficult due to judge models using limited reasoning and suffering from biases. This paper propos….

RT @TheTuringPost: The freshest research of the week:. Our top 9:.▪️ Beyond 'Aha!'.▪️ J1: Incentivizing Thinking in LLM-as-a-Judge via Rein….

We're organizing the RAM 2 workshop at COLM 2025 (10 years after the first edition in NeurIPS 2015). Check out our Call of Papers on topics in Reasoning, Attention, and Memory.

RT @chenxi_jw: Presenting new work: Thinking LLM-as-a-Judge via RL!. It’s been great fun working with @swarnaNLP, @jaseweston, @uralik1 and….

RT @NathanThinks: excellent work by @jaseweston & team—extending our "Generative Reward Models" work with RL (GRPO) to optimize LLM reasoni….

Check out our paper for more analysis and ablations, including:. - score distribution of Pointwise-J1 models. - different reward schemes. - different seed thinking prompts. - reward+thought lengths. Fun Fact: Even before the project started, we knew what we wanted to call it 😀

Next, Pointwise-J1 at 8B + 70B scales:. 1⃣ Mitigates position bias. 2⃣ Improves position-consistent accuracy. 3⃣ Reduces ties in pairwise judgments. Finally, test-time scaling of J1 leads to further improvements, for both Pairwise and Pointwise models, at both scales

We tested J1 on 5 benchmarks w/ verifiable + non-verifiable + multilingual instructions at 8B+70B scales. First, Pairwise-J1 outperforms:. 1⃣ open + closed LLM judges . 2⃣ SOTA scalar + generative RMs. 3⃣ R1-distilled-Llama + o1-mini. 4⃣ a much larger R1 on non-verifiable tasks

🧑‍🍳 J1 Recipe:. -- Generate synthetic preference pairs as training data for both verifiable+non-verifiable tasks. -- Train Pairwise-J1 using GRPO with verdict correctness+consistency rewards. -- Train Pointwise-J1 using GRPO with distant pairwise supervision + score-based rewards

Excited to share that EvalPlanner is accepted to #ICML2025! . To make meaningful progress in AI, we need strong evaluators, and specifically those that can reason. Stay tuned for more updates, as we continue to make progress in this space! 😀.

RT @SomnathBrc: 𝐇𝐨𝐰 𝐜𝐚𝐧 𝐰𝐞 𝐩𝐞𝐫𝐟𝐞𝐜𝐭𝐥𝐲 𝐞𝐫𝐚𝐬𝐞 𝐜𝐨𝐧𝐜𝐞𝐩𝐭𝐬 𝐟𝐫𝐨𝐦 𝐋𝐋𝐌𝐬?. Our method, Perfect Erasure Functions (PEF), erases concepts from LLM repre….

RT @tesatory: Ten years ago in 2015 we published a paper called End-to-End Memory Networks (. Looking back, this pa….

RT @jaseweston: 🚨Multi-Token Attention🚨.📝: Attention is critical for LLMs, but its weights are computed by single….

RT @ArchikiPrasad: 🚨 Excited to share: "Learning to Generate Unit Tests for Automated Debugging" 🚨.which introduces ✨UTGen and UTDebug✨ for….

Complex evaluation is in parts planning and in parts reasoning. Hence, we trained an LLM to think before producing a judgment. My first work since joining this awesome team 😄.

RT @mohitban47: 🎉 Congrats to the awesome students, postdocs, & collaborators for this exciting batch of #ICLR2025 and #NAACL2025 accepted….