jhourney retreat solves this x.com/hermittoday/st…

🤯 We cracked RLVR with... Random Rewards?! Training Qwen2.5-Math-7B with our Spurious Rewards improved MATH-500 by: - Random rewards: +21% - Incorrect rewards: +25% - (FYI) Ground-truth rewards: + 28.8% How could this even work⁉️ Here's why: 🧵 Blogpost: tinyurl.com/spurious-rewar…

(1/5) We’ve never enjoyed watching people chop Llamas into tiny pieces. So, we’re excited to be releasing our Low-Latency-Llama Megakernel! We run the whole forward pass in single kernel. Megakernels are faster & more humane. Here’s how to treat your Llamas ethically: (Joint

New paper! We used GRPO to train Qwen 2.5 on 32 randomly-generated Coq programs that don't compile, and it learned to prove the Riemann Hypothesis.

Still not understanding how this can work

Yeah I mean imagine starting from a random policy "RL is just eliciting latent behaviors present in the uniform distribution over tokens"

GDM releases Formal Conjectures Lean 4 repo of formalized open math problems—no proofs. - built to benchmark auto-theorem proving, clarify conjectures, and grow mathlib ↓ github.com/google-deepmin…

This is important, most RL results on qwen need to be reevaluated Seems insane to me that ppl don't run proper baselines

Reminds me of this bad boy: arxiv.org/abs/2503.14286 In it they do find that pushing the effective proportion of neg samples up does improve performance

🚨Beyond 80/20 in LLM reasoning🚨Dropping 80% low-entropy tokens in RL greatly boosts performance 🔗arxiv.org/abs/2506.01939 🏆Zero-RL SoTA: 63.5/68.1 (AIME24), 56.7 (AIME25) 🚀Insights: 1. RL retains base model entropy patterns 2. High-entropy tokens drive all RL improvement ⬇️

Someone needs to use this metric as a non-stationary reward for synthetic data generation

1/ I'm excited to share recent results from my first collaboration with the amazing Aran Nayebi and Leo Kozachkov! We show how autonomous behavior and whole-brain dynamics emerge in embodied agents with intrinsic motivation driven by world models.

I find it amusing these posts were published a day apart. The apparent contradiction pivots around how easily the task context can be factorized. Devin finds coding tasks tend to be less factorizable than research tasks, and therefore advocates against the multi-agent approach

There are two competing narratives about AI: (1) there's too much hype (2) society is being too dismissive and complacent about AI progress. I think both have a kernel of truth. In fact, they feed off of each other. The key to the paradox is to recognize that going from AI

Synthetic ecosystems that autonomously and continuously evolve in silico 👾 An Alifer dream we pursue with Flow-Lenia ! If you are interested in complex systems with (1) emergent creatures and (2) intrinsic evolutionary dynamics, go check our new paper ! A 🧵

Enjoyed this post on open ended agents We need: - cheap open-ended environments - good methods for autocurriculum - effective memory and online learning (from skill libraries to weight updates) Sort of wonder why there aren't so many Voyager followups

In our latest paper, we discovered a surprising result: training LLMs with self-play reinforcement learning on zero-sum games (like poker) significantly improves performance on math and reasoning benchmarks, zero-shot. Whaaat? How does this work? We analyze the results and find

Making sure your tensors are scaled after the residual stream (along others) seems to help with stability and quantized training