"JEV-as-a-Judge: Accept When Confident, Escalate When Unsure" This paper shows you can just use JEV for every evaluation instead of expensive LLM. JEV basically acts as a cheap first-pass judge, returning both a verdict and how confident it is. When confidence is high, keep the answer. When it’s low, escalate to a stronger LLM. This simple routing keeps ~99% of GPT-6’s accuracy while reducing evaluation cost by a lot.

Recent research has highlighted the promise of scaling memory embeddings in LLM training. While Engram and STEM index memory by token identity or local n-grams, can we design more flexible memory routing that captures how each token’s meaning changes with context? This paper introduces Mixture-of-Memory Embeddings (MoME), which uses a learned router to sparsely select among multiple memory slots for each token. The architecture outperforms strong memory baselines across three model families. Interestingly, both qualitative and quantitative analyses show that the learned routers’ activations correlate with the context-dependent senses of polysemous words. The authors also train a sub-billion-parameter model that achieves competitive CORE-22 performance against similarly sized base models, including Qwen3-0.6B and Llama 3.2-1B. Code and pretrained models are publicly available!

OpenResearch just hit 5k stars on GitHub! Thanks to those who have kept us on the trending list 🚀 If you're just joining us, OpenResearch turns your coding agents into research agents. Check it out now:

“In-Context Robot Learning with VLM Agents” Most robots need lots of task-specific training before they can do something new. This paper shows you can instead just show a general VLM what to do and let it adapt directly from context. GPT-Policy essentially turns robot learning into prompting. So they gave the frozen VLM an example of the task, even just a human video with no robot action labels, and it can translate what it sees into robot actions on the fly. This brings few-shot learning into the physical world, where teaching a robot a new behavior could look more like showing it an example than retraining a policy.

“Atria Dawn: The Dawn of Agentic Superintelligence” This paper shows AI agents are shifting from executing research tasks to actually participating in the research process. The interesting part is that as agents become better at doing the work, human effort shifts toward deciding which directions are worth pursuing and how to interpret what the experiments reveal. So the real bottleneck toward recursive self-improvement may no longer be execution, but whether AI can develop research judgment itself.

“Dream-RSI: Recursive Self-Improvement through Evolving Worlds” AI agents can search for better solutions, but they’re usually stuck using the same search strategy over and over. Dream-RSI lets the agent learn how to search better by turning its past exploration into a simulator, where it can cheaply replay different strategies before spending compute in the real world. So the agent improves not just its solutions, but also the process it uses to discover them.

New paper! In subliminal learning, LLMs transmit traits (e.g. loving cats) though seemingly unrelated data (e.g. numbers). We proactively detect these effects as readable prompts. To do so, we use the surprising ability of models to verbalize learned soft prompts.🧵

OpenResearch just hit 5k stars on GitHub! Thanks to those who have kept us on the trending list 🚀 If you're just joining us, OpenResearch turns your coding agents into research agents. Check it out now:

Six degrees of Yann LeCun We mapped the coauthorship graph of every AI researcher on arXiv 🚀 Enter any two names to find the shortest path of papers connecting them What’s your @ylecun number?

Check out our paper on @askalphaxiv with a nice summary and a 5-min podcast! Tiny Aya L2-Thinker beats models 2-7x its size on in-language reasoning rate and task accuracy on data-scarce languages.

Can we train a one-step discrete generator without a teacher model? Introducing Discrete Beckmann Transport Models 🛸 — a new family of discrete generative models that provably carries any point in the latent space to a fixed point on the vertices of the simplex in a single step. We achieve the diversity-coherence Pareto frontier for unconditional language modeling and SOTA few-step reasoning performance, with 84.6% accuracy in 4 NFEs on Sudoku-Hard and 16.8% in 32 NFEs on TinyGSM/GSM8K. Joint work with @ShiyiWangML from my visit at Harvard! Also, huge thanks to @msalbergo and @BrianLee794309 for the helpful discussions and guidance :D More in 🧵

Six degrees of Yann LeCun We mapped the coauthorship graph of every AI researcher on arXiv 🚀 Enter any two names to find the shortest path of papers connecting them What’s your @ylecun number?

“Discrete Beckmann Transport Models for One-Step Language Modeling and Reasoning” Few-step diffusion and flow LMs usually require distilling a pretrained teacher, which adds training cost and caps the student at teacher quality. So this paper learns a direct path from noise to discrete tokens, generating the whole sequence in one step without a teacher model or tracking diffusion time. Then when you add extra steps, the model doesn’t just take smaller steps toward an answer. It revisits uncertain tokens and fixes them using the tokens it’s already confident about.

Research papers are filled with citations, but clicking on any of them makes you lose your place in the paper We made citations much easier to navigate, with extra information on hover including authors and institutions Now, you can hover over any in-text citation to instantly preview the referenced paper, jump directly to the full citation, and return to exactly where you left off with a single click This includes quick access to the authors and their organizations, so you can understand the people behind the research without breaking your reading flow Try it now on any paper at

“Learning to Solve Hard Problems in RL for LLMs by Never Giving Up” This paper shows RL has a Matthew Effect, disproportionately improving problems the model can already solve while barely improving the hardest ones. So they fix this by dynamically reallocating sampling compute, quickly filtering easy prompts while repeatedly sampling hard prompts until a correct solution is found. This lets RL progressively learn from harder problems instead of wasting compute on already-solved ones.