🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Optimizing LLM Test-Time Compute Involves Solving a Meta-RL Problem. Check out our new blog below:

Last day to submit papers to our workshop SSI-FM ICLR 2025! Both long and tiny papers are welcome.

is digit multiplication the true agi test?

Thrilled to share that I recently joined Google DeepMind I’ll be working on LLM RL reasoning. Exciting times ahead and I'm eager to collaborate with and learn from the brilliant minds here.

Is BIG-Bench Hard too easy for your LLM? We just unleashed BIG-Bench EXTRA Hard (BBEH)! 😈 Every task, harder! Every model, humbled! (Poem Credit: Gemini 2.0 Flash) Massive headroom for progress across various areas in general reasoning 🤯

Happy to see a flavor 🧂 of Compositional GSM in this work Comp GSM: arxiv.org/abs/2410.01748 BBEH: arxiv.org/abs/2502.19187

🚀 New Paper Alert! Can we generate informative synthetic data that truly helps a downstream learner? Introducing Deliberate Practice for Synthetic Data (DP)—a dynamic framework that focuses on where the model struggles most to generate useful synthetic training examples. 🔥

Our team in London is hiring! If you want to come work with a wonderful group of researchers on investigating the frontiers of autonomous open-ended agents that help humans be better at doing things we love, come have a look. Link in post below 👇

Test-time scaling with GenRMs being “more efficient” than self-consistency (SC) is misleading 😶‍🌫️ In our new work, in a compute-matched setup, we show that SC outperforms GenRM across most budgets! 💰 Check out details/code and the paper here : arxiv.org/abs/2504.01005

this pretty much sums up the paper

At higher budgets GenRM beats SC, but a question remains: How to allocate the budget to generate more solutions or verifications? We derive scaling laws for optimal # of solutions and verifications. Showing that # of solution should grow 1.5-2x faster than # of verifications

We've released the models and data for this paper on huggingface 🤗: huggingface.co/sc-genrm-scali… 👉 Fine-tuned generative verifiers 👉 GPT-4o generated synthetic data for fine-tuning your own verifiers 👉 Verifications from lots of models (incl. QwQ-32B) on lots of datasets!

is there a study on what sort of concepts should we let LLMs learn by themselves rather than injecting our preferences or inductive biases? For instance, Is CoT reasoning the kind of thing that needs (language) preference/bias?

After a lovely week away from LLMs in Vietnam, I’ll be at ICLR 2026 🇸🇬 Will be at posters for GenRM and Training LLM Reasoners via Compute-Optimal Sampling . Also Google DeepMind booth on 25th (and more). Come chat about LLM Reasoning and RL, or for Vietnam recommendations

Great work by Kusha Sareen and morgane Paper : huggingface.co/papers/2505.04…

Upon some requests, we now have a BBEH Mini with 460 examples (20 per task) for faster and cheaper experimentation. The set can be downloaded from: github.com/google-deepmin… The results are reported in Table 3 of arxiv.org/pdf/2502.19187

Idea: Merging generative verification and solution generarion during RL training of LLM reasoners. Why? This allows you to scale inference compute both sequentially (long CoT) and parallel (Best of N, weighted maj voting). Next? Generation and verification to be trained end

🚨Announcing RAM 2 workshop @ COLM25 - call for papers🚨 - 10 years on, we present the sequel to the classic RAM🐏 (Reasoning, Attention, Memory) workshop that took place in 2015 at the cusp of major change in the area. Now in 2025 we reflect on what's happened and discuss the

Says more about MATH500 and AMC than RL