Search papers, labs, and topics across Lattice.
The paper introduces Chain-of-Evidence (CoE), a framework for ensuring verifiability in autonomous research agents by requiring claims to be traceable to their evidence sources. They then present ScientistOne, an autonomous research system that uses CoE to maintain evidence chains throughout the research process. Through CoE Audit, they demonstrate that existing systems exhibit significant verifiability failures, while ScientistOne achieves near-perfect verifiability and matches or exceeds human expert performance across multiple research tasks, including achieving SOTA on Parameter Golf.
Autonomous research agents are surprisingly unreliable, with existing systems hallucinating references 21% of the time and failing to align methods with code as often as 80%, but a new "Chain-of-Evidence" approach can fix this.
Autonomous research agents produce competitive solutions and professional-looking manuscripts, yet their outputs contain verifiability failures undetectable by surface-level evaluation: fabricated citations, unreproducible scores, and method descriptions that diverge from the implementation. We address this through three contributions. First, Chain-of-Evidence (CoE), a verifiability framework requiring every claim to be traceable to its evidence source. Second, ScientistOne, an end-to-end autonomous research system that maintains evidence chains by construction throughout literature review, solution discovery, and paper writing. Third, CoE Audit, a post-hoc audit whose four integrity checks -- score verification, specification violation, reference verification, and method-code alignment -- apply uniformly to all systems. Across 75 papers spanning five systems and five frontier research tasks, every baseline exhibits at least one systematic failure mode: hallucinated reference rates reach 21%, score verification passes in as few as 42% of papers, and method-code alignment ranges from 20% to 80%. ScientistOne achieves zero hallucinated references (0/337), perfect score verification (12/12), and the highest method-code alignment (14/15), while matching or exceeding human expert performance on all five tasks. ScientistOne further generalizes to six additional tasks spanning medical imaging, fine-grained recognition, 3D perception, and language modeling, achieving state-of-the-art on Parameter Golf and gold medals on MLE-Bench tasks where baselines fail entirely.
