TwinLab-Bench Evaluates AI Agents on Self-Designed Experiments
Yang Wan and Linchao Zhu introduced TwinLab-Bench, a benchmark with 52 tasks drawn from physical science, life science, and engineering laboratories to evaluate AI agents that design their own experiments. Each task uses a TwinLab agent to execute the designed experiment against laboratory records, supports more than 100 action types, and scores the agent’s report against a hidden answer. The paper reports that Claude Fable 5.2 was the strongest of 13 evaluated agents, with an average score of 45.0.
Existing benchmarks mainly test agents that follow predefined workflows or choose from predefined candidates. TwinLab-Bench instead evaluates agents that design experiments, with readings grounded in real laboratory records.
Computer Science > Artificial Intelligence
arXiv:2611.26248 (cs)
[Submitted on 19 Nov 2026]
Title:TwinLab-Bench: Evaluating AI Scientists in Agentic Twins of Real Labs
Authors:Yang Wan, Linchao Zhu
View a PDF of the paper titled TwinLab-Bench: Evaluating AI Scientists in Agentic Twins of Real Labs, by Yang Wan and 1 other authors
#TwinLab-Bench#TwinLab agent#Claude Fable 5.2#Claude Opus 5.5#GPT-6.1-sol#DeepSeek#Qwen#Kimi#GLM