Tag
LabInstruct 提出了一个针对真实实验室场景的指令性视频生成基准,涵盖5个学科的204项任务,并评测了8个前沿图生视频模型,发现视觉上合理的生成结果往往在程序步骤上并不正确,揭示了视觉真实感与实验指导可靠性之间的显著差距。
This comment discusses the convergence of AI reasoning, autonomous experiments, high performance computing, and quantum computing to reshape chemistry and accelerate materials science discovery.
This paper presents a benchmark and trace-logging framework for evaluating LLM-based agents that control microscopes, comparing 105 agent configurations and finding that benchmarks support qualification but do not reliably predict performance on unseen tasks.