Contrastive Discovery scores competing explanations for scientific agents
Original title: Contrastive Discovery: Open-Ended Scientific Discovery over Competing Explanations
Contrast sets resolve 73% of large belief-shift rejections to a specific alternative and resolve 67% more questions under the same experiment budget.
The paper introduces Contrastive Discovery, which replaces pointwise hypothesis tests with a contrast set of mutually exclusive, collectively exhaustive explanations. Experiments are scored by how evidence redistributes belief, and a question is resolved when belief concentrates on one explanation. Across six domains, 89% of the largest pointwise belief shifts are rejections; contrast sets resolve 73% of those to a specific alternative. Guided search resolves 67% more questions than pointwise surprisal under the same budget. In a blinded study, 38 experts agreed with the agent on 78% of resolved questions.
- 89% of the largest pointwise belief shifts are rejections.
- Contrast sets resolve 73% of those candidates to one alternative.
- Guided search resolves 67% more questions than pointwise surprisal.
- 38 experts matched the agent on 78% of resolved questions and 82% of stopped rejections.
From the original article
1 min read at arxiv.orgLLM agents are increasingly used for open-ended, data-driven scientific discovery, yet most of them evaluate hypotheses pointwise, testing an empirical claim $h$ only against its complement $\neg h$. When evidence rejects $h$, the agent learns that one explanation is unlikely but not which alternative the evidence supports.
By Ziang Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.