Google's Research Agent That Links Evidence

A Google research team proposed the ScientistOne system and the Chain-of-Evidence (CoE) framework, which guarantee the verifiability of papers produced by autonomous research agents at the architectural level.

CoE is designed so that every claim in a paper can be traced back to its grounding source through a recorded evidence chain. There are two core criteria.

• Completeness: Every claim must be linked to an evidence chain.
• Correctness: The linked evidence must actually support the corresponding claim.

The research team pointed out that in an autonomous research pipeline where literature survey, hypothesis formulation, experiments, code, and result reporting follow one another, early errors can propagate and amplify into later stages. Existing research systems and benchmarks evaluate paper completeness or experiment execution capability, but fail to verify whether individual claims are actually linked to their execution artifacts.

An audit of 75 papers written by 5 systems found systematic failures across all baselines. The hallucinated reference rate reached up to 20.9%, the score verification pass rate was as low as 42%, and methodology-code consistency ranged from 20-80%.

In contrast, ScientistOne reported the following results.

• Hallucinated references: 0/337
• Score verification: 12/12
• Methodology-code consistency: 14/15

This research is significant in that it expands the evaluation criteria for autonomous research systems beyond fluent manuscripts and experimental success, to per-claim evidence traceability and reproducibility.