Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen 2.5 achieve significant gains even with spurious...

Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

Leaderboard benchmarking LLMs playing Balatro: rounds, tool-call reliability, cost, and speed.

Can a diffusion model trained on bedrooms recover human faces? Diffusion models are widely used as priors for inverse problems, but standard approaches usually assume a high-fidelity model trained...

Rapidly increasing AI capabilities have substantial real-world consequences, ranging from AI safety concerns to labor market consequences. The Model Evaluation & Threat Research (METR) report...

State-of-the-art MLLMs achieve PhD-level language reasoning but struggle with visual tasks that 3-year-olds solve effortlessly. We introduce BabyVision, a benchmark revealing the infancy of AI vision.

We are excited to introduce the Waymo World Model, a frontier generative model that sets a new bar for large-scale, hyper-realistic autonomous driving simulation. 

Referring Video Object Segmentation (RVOS) aims to segment objects in videos based on textual queries. Current methods mainly rely on large-scale supervised fine-tuning (SFT) of Multi-modal Large...

As neural theorem provers become increasingly agentic, the ability to interpret and act on compiler feedback is critical. However, existing Lean datasets consist almost exclusively of correct...

To assess the ability of current AI systems to correctly answer research-level mathematics questions, we share a set of ten math questions which have arisen naturally in the research process of...

Topological quantum computation encodes quantum information in the internal fusion space of non-Abelian anyonic quasiparticles, whose braiding implements logical gates. This goes beyond Abelian...

The success of RL for LLM post-training stems from an unreasonably uninformative source: a single bit of information per rollout as binary reward or preference label. At the other extreme,...

Why is nanochat's optimal tokens per param 8, much smaller than 20 from Chinchilla? We had similar findings in our NeurIPS work https://arxiv.org/abs/2512.05620. We hypothesize two key factors: (1) improved optimization and (2) higher-quality data. 1/n🧵
…

Real-world decision-making tasks typically occur in complex and open environments, posing significant challenges to reinforcement learning (RL) agents' exploration efficiency and long-horizon...

The toric code, when deformed in a way that preserves the self-duality $\mathbb{Z}_2$ symmetry exchanging the electric and magnetic excitations, admits a transition to a topologically trivial...

BinaryAudit benchmarks AI agents using Ghidra to find backdoors in compiled binaries of real open-source servers, proxies, and network infrastructure.

63 TiB of chess knowledge sent across the Atlantic and now available on the Lichess analysis board

Nature - Discovery of a protocluster at z = 5.68, merely one billion years after the Big Bang, suggests that large-scale structure must have formed more rapidly in some regions of the...

We’re releasing a major upgrade to Gemini 3 Deep Think, our specialized reasoning mode.

Contribute to hacker-fab/gitbook development by creating an account on GitHub.

If LLM training data is polluted with benchmark test data, then benchmark performance gives biased estimates of out-of-distribution (OOD) generalization. Typical decontamination filters use n-gram...