The authors built CLASHarness around hierarchical scenario planning and an eight-step verified generation pipeline.
Benchmark exposes computer agents’ failures when resolving conflicting information
CLASHarness tests whether vision-language agents verify contradictory reports, correct system state and complete coordination protocols.
Top University
Yuxiang Lai · Jike Zhong · Chengzhi Liu · Ming Li · Haonan Wang · Xiaofeng Yang · +3 more
Emory University · University of Southern California · University of California, Santa Barbara · University of Florida · Columbia University
Research Digest··2 min read
Lai and colleagues introduce a pipeline for generating computer-use tasks in which agent messages conflict with evidence visible in the operating environment.
§