PhysTrace is a training-free framework that aligns model-generated expectations with observed video windows to localize and describe physical violations over time, improving evaluation of physical consistency in video models, with IntPhys2-Timing enhancements and strong performance on discrimination tasks.
A CohereLabs dataset documents the Qwen 3.8 Flash Next model's unexpected behavior, supplying a single-row trace and raw token IDs in JSONL and Parquet formats for AI safety and interpretability research; it includes load instructions, usage notes, and limitations.
Veritas is a system for generating task verifiers using evidence descriptions, domain rules, and authoring principles. In tests across six benchmarks, its verifiers matched official verdicts better than competing harnesses on all four execution-scored benchmarks and performed at or near the best on two pairwise benchmarks.
Machine learning teaches software to predict or generate content by learning from data, difference from explicit rules, and a typical six-stage workflow from task definition to model training and tuning.
The paper examines how trace-based labels for evaluating LLM agents can misrepresent performance. It analyzes detectors across models and problems, highlighting failure modes like silent tool-output overrides and the impact of execution limits, and argues for transparent trace evidence with automated labels.
Verity describes a private corpus of attacks used to test its gate and publishes a map of known weaknesses. Users can test the gate; newly found defeats are recorded publicly and paid.
The paper introduces TRACE, a method that reconstructs unlearned training data from pre- and post-unlearning model snapshots, revealing privacy risks in approximate unlearning for LLMs and suggesting defenses against parameter-difference leakage.
ML learns from data to predict or decide, using supervised, unsupervised, or reinforcement methods. A solid workflow and proper evaluation prevent misleading results and ensure models generalize well.
An AI-driven autonomous lab conducts end-to-end scientific inquiry on quantum spins in diamond, identifying drift, autonomously revising protocols, and deriving seven calibration points for microwave pulse sequences, surpassing prior methods while noting challenges in scaling to real-world conditions.
Lexsi Labs가 Cohere Labs 및 Hugging Face와 협력해 AI 안전성·정렬 hackathon을 개최합니다. 공개 소스 도구 스택을 실제 사용 사례에 적용해 안전성 개선을 입증하고 피드백을 모읍니다.