Open Brain
arXivPaperNeeds Review

Contrastive Discovery: Open-Ended Scientific Discovery over Competing Explanations

Contrastive Discovery asks research agents to compare competing explanations, so rejecting one hypothesis can still identify a plausible alternative using the same evidence.

arXiv · Nov 10, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The method has agents propose mutually exclusive, exhaustive explanations and score experiments by how evidence shifts belief across them. Across **six domains**, contrast sets resolved **73%** of candidates that pointwise agents had rejected; guiding search resolved **67% more questions** under the same experiment budget.

Practical Implication

For builders of research agents, make the alternatives explicit when a test can falsify a hypothesis without identifying what is true. Existing agents can apply contrast sets to previously collected evidence as well as use them to guide new experiments.

Agent-Ready Context
The method has agents propose mutually exclusive, exhaustive explanations and score experiments by how evidence shifts belief across them. Across **six domains**, contrast sets resolved **73%** of candidates that pointwise agents had rejected; guiding search resolved **67% more questions** under the same experiment budget.

For builders of research agents, make the alternatives explicit when a test can falsify a hypothesis without identifying what is true. Existing agents can apply contrast sets to previously collected evidence as well as use them to guide new experiments.

In a blinded study, **38 experts** selected the same explanation as the agent on 78% of resolved questions. The material describes scientific discovery tasks; it does not show that this approach improves general coding-agent performance.
Connected Context · Feed7 Judgment

Contrastive Discovery makes uncertainty among explanations explicit: experiments are selected for how they distinguish alternatives, which can recover candidates that pointwise testing rejects. Its reported benefits concern scientific discovery tasks and do not establish gains for general coding agents.

Context Map
agentresearch#reasoning#agent-reliability
Uncertainty
In a blinded study, **38 experts** selected the same explanation as the agent on 78% of resolved questions. The material describes scientific discovery tasks; it does not show that this approach improves general coding-agent performance.