A new 27B model helps LLMs find mathematically interesting theorems, giving researchers a way to rank conjectures and guide proof search instead of solving only human-supplied problems. It scores a theorem by comparing its proof length with the length of its statement, then generates candidates, picks the strongest ones, and builds on a growing formal library. Optimizing that score cut substantial or full overlap with Mathlib from 91.9% to 30.6%.
Anthropic says Claude identified a previously unrecognized enzyme system that could point to a new programmable DNA tool, though its function has not been established. The model spotted a reverse transcriptase in a jumbo phage alongside repeated non-coding DNA and an accessory protein, a combination found in only a handful of other systems used to cut, copy, or paste DNA. Scientists then reviewed the result and are testing it experimentally.
MentalHealthBench gives researchers an open way to test whether AI handles everyday stress, relationships, and acute crises safely, not just whether it avoids prohibited replies. More than 80 licensed mental health experts across 22 countries helped create realistic conversations scored on safety, context-seeking, user agency, and actionable guidance. Researchers can inspect the methods, run their own evaluations, and use the results to improve models that guide people toward real-world support.
RECLAIM shows that AI agents are still unreliable at rebuilding machine-learning results, especially when researchers release no code: the best agent reproduced 15% of those papers, versus 41% when code, data, and model weights were available. The benchmark tests 100 NeurIPS 2025 papers against fixed results and GPU budgets. Across 400 runs, the most common mistake was implementing the method without checking it against the paper’s numbers.
WROP gives video-generation researchers a 1.5M-sample corpus for training object permanence and solidity, so models can be tested on whether objects persist and behave as physical things. It uses 150 hand-designed tasks, with Blender varying lighting, speed, and camera angle while preserving each task’s cognitive structure. On a 300-question exam across 14 video models, the authors’ 16B PWM-WROP ranked first among continuation models and third overall in a blind pairwise Elo study.
Fast NoThink post-training may get much of its accuracy boost by pulling models back toward behavior already available in Think mode, rather than adding new NoThink capability. The paper tests this by steering model activations in a direction derived from the base model: steering reproduces most of the gain, while counter-steering removes a substantial share. Across nine aligned checkpoints, 42% to 79% of positive NoThink gains are attributed to this leakage.
VINTAGE-TS is designed to test time-series forecasts using only the data available when the forecast was made, so researchers can measure performance without hindsight contamination. It separates when an observation occurred from when it became available, then predicts both its first published value and a later revision, including uncertainty in the gap. The paper reports a synthetic demonstration and 25-configuration sensitivity suite, not real ALFRED or Chronos-2 experiments, and claims no foundation-model advantage yet.