EviGraph Improves Autonomous Research Agent Reliability.

Autonomous research agents hold great promise for generating hypotheses, conducting experiments, and drafting scientific manuscripts. However, a common challenge is that their outputs often contain unsupported claims and inconsistencies between different stages of the research process, such as questions, experiments, results, and conclusions. This issue stems from existing systems typically organizing research as sequential pipelines without explicitly managing or validating the evolving claim-evidence structure. To address this, EviGraph introduces an autonomous research framework that models the entire research process as a typed evidence graph. This graph comprises nodes representing problems, gaps, hypotheses, experiments, findings, and claims, serving as the agent's dynamic operational state rather than just a post-hoc record. EviGraph actively inspects evidence chains for missing dependencies, semantic misalignments, and inconsistencies between results and claims. When a weakness is detected, it localizes the earliest problematic node and regenerates its affected downstream subgraph. Graph checkpointing further prevents unsuccessful repairs from corrupting previously validated evidence. Manuscripts are only generated once every retained claim is firmly grounded in a validated evidence chain, leading to significantly more reliable research outputs.