Can a model learn to break its own reasoning plateau? In our new paper, we show that LLMs can be taught with meta-RL to generate their own "stepping stones" that kickstart learning on hard math problems (0/128 success rate) where direct RL fails. Paper 📝: arxiv.org/abs/2601.18778 Blog post 🌐: ssundaram21.github.io/soar/ (1/n)

Awesome work as always by @SophieLWang w/ @_AmilDravid, with very neat and intuitive insights about how mid-training data affects downstream reasoning!

Synthetic data stays winning 🚀 now applied to xrays! Love the demos and figures too

More efficient test time LLM reasoning with learned concepts: Check out our new paper "Beyond Repeated Sampling"! arxiv.org/abs/2609.26704

Can we tell whether data domains cooperate or compete during pretraining? Adding code to the mix makes models better at math, while some other combinations hurt each other. We call this data synergy. Turns out you can incorporate data synergy into scaling laws and estimate it 🧵

Super excited to be able to release this project we've been working on. The largest open robotics dataset to date and policies capable of dexterous manipulation!

Super cool work from @juliachae_ that fills an important gap in image similarity metrics!

Our team at @AIatMeta is excited to announce ATLAS: one of the largest automated formalization efforts to date. ATLAS contains Lean 4 formalizations of both statements and proofs from 25+ mathematics textbooks, spanning dozens of domains, for a total of 500k lines of code. We are also releasing a flexible formalization harness and a companion paper. External contributions are welcome! Joint work spearheaded by our amazing PhD student Ahmad Rammal (@Ahmad3Rammal), together with Niket Patel (@niketnpatel ), Fabian Gloeckle (@FabianGloeckle), Amaury Hayat (@Amaury_Hayat), Remi Munos (@MunosRemi), Julia Kempe (@KempeLab), Vivien Cabannes, and myself from @AIatMeta, @NYUDataScience , and Ecole des Ponts. This is an ongoing effort; more details in the thread below. (1/9)

What is the right data mix, and how do we find it as the data keeps changing? This is a core, unsolved problem in continual learning. To tackle it, we built a data mixing algo that works everywhere — pretraining, midtraining, instruction tuning Introducing: On-Policy Mix 🧵1/6

"The Truth Lies Somewhere in the Middle (of the Generated Tokens)" In autoregressive language models, mean pooling hidden states across generation yields better representations than any token alone. project page: sophielwang.com/tokens w/ @phillip_isola and @thisismyhat

As a geometric ML researcher, I noticed pseudoscalars don’t get enough attention! Read on to see what pseudoscalars can do for you.

LLMs can learn to self-generate curricula for hard problems that they can't yet solve! Using meta-RL, with rewards grounded in learning progress, models produce their own stepping stones that kickstart learning on hard problems where direct RL plateaus. Poster at the ICLR RSI workshop today!

New blogpost on tokenizing non-sequential data! Language has sequential structure, which gave rise to the next-token prediction paradigm of LLMs. But we increasingly use LLMs for data without inherent order (e.g. images, molecules, sets). What does “next token” mean here? (1/7)

Introducing GASP😮: Guided Asymmetric Self-Play for Coding LLMs We address the goal-agnostic behavior of current asymmetric self-play methods. Key idea: guide the teacher with hard real-data goalposts; first an easier lemma, then a harder lift from the lemma as stepping stones 🧵

Simply adding Gaussian noise to LLMs (one step—no iterations, no learning rate, no gradients) and ensembling them can achieve performance comparable to or even better than standard GRPO/PPO on math reasoning, coding, writing, and chemistry tasks. We call this algorithm RandOpt. To verify that this is not limited to specific models, we tested it on Qwen, Llama, OLMo3, and VLMs. What's behind this? We find that in the Gaussian search neighborhood around pretrained LLMs, diverse task experts are densely distributed — a regime we term Neural Thickets. Paper: arxiv.org/pdf/2603.12228 Code: github.com/sunrainyg/RandOpt Website: thickets.mit.edu

Can language models learn useful priors without ever seeing language? We pre-pre-train transformers on neural cellular automata — fully synthetic, zero language. This improves language modeling by up to 6%, speeds up convergence by 40%, and strengthens downstream reasoning. Surprisingly, it even beats pre-pre-training on natural text! Blog: hanseungwook.github.io/blog/… (1/n)