StatefulDiscovery: Evidence-Calibrated Claim Formation in Open-Ended Scientific Discovery

Researchers introduce StatefulDiscovery, a framework that enables AI agents to conduct open-ended scientific discovery by maintaining explicit investigation state and coupling it with evidence-calibrated claim formation. The system addresses the challenge of avoiding overinterpretation by coordinating exploration trajectory with evidential support, demonstrated across 40 real-data tasks where it outperformed baseline approaches in producing well-supported, high-value claims.

StatefulDiscovery addresses a fundamental problem in autonomous scientific research: how AI agents can reliably explore phenomena and form justified claims without exceeding the evidential basis of their investigations. Traditional discovery systems struggle with the evidence-calibration problem, where agents may generate claims that sound compelling but lack sufficient analytical support. This research introduces a framework that externalizes investigation state—essentially maintaining a structured record of what has been explored, what evidence supports various hypotheses, and what remains uncertain.

The framework coordinates three critical processes: frontier selection (deciding what to investigate next), evidence acquisition (gathering relevant data), and claim adjudication (determining when evidence sufficiently supports a claim). By coupling these processes through explicit state management, StatefulDiscovery prevents the common failure mode where exploratory analyses lead to unfounded conclusions. The evaluation across 40 real-data tasks demonstrates practical effectiveness, with the system producing more high-value, well-supported claims than competing approaches.

This work has implications for how autonomous systems can conduct responsible scientific discovery at scale. As AI increasingly assists researchers across domains—from materials science to genomics—the ability to avoid spurious discoveries while still identifying genuinely novel phenomena becomes critical. The framework's reliance on structured hypotheses, local evidence adjudication, and controlled frontier expansion suggests that explicit reasoning about uncertainty and evidence quality substantially improves discovery outcomes.

Future development should focus on extending this approach to multi-agent discovery scenarios and integrating it with emerging large language model capabilities, where evidence-calibration remains a significant challenge.

• →StatefulDiscovery couples exploration trajectory with evidence-calibrated claim formation to prevent overinterpretation in autonomous scientific discovery.
• →The framework coordinates frontier selection, evidence acquisition, and claim adjudication through explicit investigation state management.
• →Evaluation on 40 real-data tasks shows StatefulDiscovery produces more well-supported, high-value claims than baseline approaches.
• →Structured hypotheses and local adjudication mechanisms contribute significantly to preventing spurious discoveries.
• →Explicit discovery state enables AI agents to responsibly conduct open-ended research without exceeding evidential scope.