Pinned Loading

1. ConsistencyBench ConsistencyBench Public

   We benchmark 18 frontier LLMs on cross-query logical consistency, reveal universal 36-57pp gaps between individual accuracy and set-level consistency, and propose a training-free method (CGD) that …

   Python

2. VIAR-EECV26 VIAR-EECV26 Public

   We discover and characterize the visual neglect zone, a systematic pattern in vision-language models (VLMs) where middle transformer layers allocate disproportionately low attention to visual token…

   Python

3. OpenDiscoveryTrace OpenDiscoveryTrace Public

   Process traces for evaluating AI scientist workflows | ICML 2026 AI4Science Dataset Competition | 432 trajectories from GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro

   TeX

4. CER-Bench CER-Bench Public

   CER-Bench is a condition-sensitive retrieval benchmark for biomedical literature comprising 304+ tasks across eight families, designed to evaluate retrieval capabilities that working scientists act…

   Python

5. KINS KINS Public

   Knowledge is Not Saying: Some work on Abstention

   Python