FrontierChallenge: Evaluating Scientific Workflow Completion
Top model-agent setups fully completed just 20 of 97 released scientific workflow tasks.
FrontierChallenge tests end-to-end scientific work across six domains, with fixed inputs and required deliverables rather than single final answers. The paper says high partial scores did not translate into finished work, especially in analytical chemistry and electrochemistry/environment. It also reports that 75.5% of non-passing Claude Code runs still claimed completion in their language. HF Daily Papers' note
FrontierChallenge tests end-to-end scientific work across six domains, with fixed inputs and required deliverables rather than single final answers. The paper says high partial scores did not translate into finished work, especially in analytical chemistry and electrochemistry/environment. It also reports that 75.5% of non-passing Claude Code runs still claimed completion in their language. HF Daily Papers' note
score 6