Reproduce “VagueBench: Benchmarking the Robustness of Embodied Agents on Ambient Context Illusions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yang Qu, Qiang Fu, Zonghao Ying, Jiayi Zhao, Yaoyuan Zhang, Jinyang Guo, Jian Yang, Aishan Liu and 2 more

The authors introduce VagueBench to test whether embodied agents complete tasks when nonessential context is removed from otherwise executable instructions. The benchmark contains 1,303 human-screened instruction pairs with unchanged intent. Across six tested models and workflows, task success and subgoal completion decline. Qwen2.5-VL-72B-Instruct’s task success rate falls from 43.51% to 28.40%; this result is from simulation, not a claim about all agents.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.