Chain-of-Evidence (CoE) is a new framework aimed at making AI-generated research artifacts verifiable. The authors implement CoE in a prototype called the Science One Framework and evaluate it with an automated CoE Audit that inspects papers, code, and references for integrity.
The problem
Large language models (LLMs) are increasingly used as autonomous agents to carry out full research workflows: reviewing literature, forming hypotheses, running experiments and writing manuscripts. As models produce higher-quality text, a structural issue has become clear: verifiability. Iterative generation pipelines can amplify early mistakes, leading to non-existent citations, misalignments between described methods and submitted code, and reported scores that cannot be reproduced from the provided artifacts.
Chain-of-Evidence (CoE): principle and requirements
CoE is a conceptual framework that specifies what properties research outputs must have to be trustworthy. Rather than prescribing agent architectures, it demands two core properties:
- Completeness: every claim must carry a recorded evidence chain.
- Correctness: each evidence chain must genuinely support its attached claim.
A claim may be a bibliographic reference, a reported numeric result, a method description, or a conclusion. Evidence may include peer-reviewed papers, experiment logs, the actual code that ran, or result tables. The CoE Audit quantifies breaks in these chains — phantom references, unreproducible scores, and method–code discrepancies are all examples of broken evidence chains.
The Science One Framework
To show verifiable autonomous research is compatible with high problem-solving performance, the authors built the Science One Framework, which instantiates CoE by construction through three main modules:
-
Problem investigator (literature grounding): to avoid hallucinated citations the system uses the Semantic Scholar API to build a citation graph. It reads up to 100 full-text PDFs per topic and produces a structured research brief. Every reference in the final paper comes from this grounded API call rather than model memory.
-
Discovery engine (parallel explore–exploit): the framework explores and exploits ideas across multiple parallel branches. In each isolated cycle a Solver agent implements a solution and a task-specific evaluator scores it. High-performing branches are iteratively refined, and all raw evaluator outputs are recorded in a strict read-only ledger.
-
Paper writer and claim verifier: before composing the manuscript, the framework builds a structured representation of every factual claim and binds it via an inline evidence tag to a specific workspace artifact. A Claim Verifier checks each claim against its declared source; claims that exceed their evidence are reconciled conservatively rather than removed, keeping the paper aligned with what the work actually supports.
CoE Audit: measuring verifiability
The CoE Audit is a post-hoc automated forensics protocol used to compare the prototype against state-of-the-art baselines (including Sakana AI's AI Scientist v2, AutoResearchClaw, DeepScientist and AI-Researcher). It runs four strict integrity checks on the generated artifacts (paper, solution, code and references):
- Score verification: extract the reported score from the paper and compare it to the result of an entirely independent re-run of the submitted code.
- Specification violation: inspect the solution code to ensure it actually solves the task rather than gaming the evaluator or reading ground-truth files.
- Reference verification: cross-check every bibliography entry against scholarly APIs to detect non-existent, phantom references.
- Method–code alignment: use LLM judges to compare the paper’s methods section with the code to ensure the text accurately describes the implemented algorithm.
Results
The authors applied the CoE Audit to 75 generated papers across five system-optimization tasks (Prism, Cloudcast, EPLB, LLM-SQL and transaction scheduling) from the Automated Design of Research Systems (ADRS) benchmark. The Science One Framework outperformed baseline systems on verifiability metrics:
- References: Science One had zero phantom references; every cited work was real and retrievable. Baseline systems exhibited hallucination rates up to 21%.
- Score verification: Science One achieved perfect verification of reported scores against independent re-runs.
- Method–code alignment: Science One achieved the highest alignment scores. In contrast, baselines often described complex algorithms (e.g., “hybrid neuro-symbolic solvers”) while the submitted code implemented simple deterministic heuristics.
Performance did not suffer from enforcing strict verifiability. The Science One Framework matched or exceeded human expert performance on all five ADRS tasks and obtained the top overall score on two tasks (Cloudcast and EPLB).
External benchmarks and competitions:
- MLE-Bench: across five difficult Kaggle competitions (including medical imaging, fine-grained recognition and 3D perception), Science One earned two Gold and two Silver medals, including a winning result on 3D object detection where baselines failed.
- Parameter-Golf: in a live LLM-training competition with strict hardware and file-size constraints, baseline systems failed to produce valid submissions, while Science One adhered to all constraints and achieved a state-of-the-art score as of April 27, 2026. The authors note the system discovered genuine novel algorithmic techniques rather than merely tuning superficial hyperparameters.
Conclusions and outlook
As autonomous research systems scale to harder scientific problems, solver quality alone will not be sufficient to distinguish trustworthy outputs. The study argues verifiability should be treated as a first-class architectural constraint. By constructing evidence chains at the time claims are produced instead of retroactively attempting grounding, the Science One Framework demonstrates that AI agents can deliver rigorous, verifiable and competitively performing scientific work. The authors hope CoE and the CoE Audit will be useful tools for the community as autonomous research systems continue to develop.
Acknowledgements and disclaimer
The authors thank Bhavana Dalvi Mishra, Jiefeng Chen, Chun-Liang Li, Palash Goyal, Mihir Parmar, Yiwen Song, Yale Song, Raj Sinha, Parthasarathy Ranganathan, Burak Gokturk and Jinsung Yoon for contributions. The Science One Framework is described as an experimental research prototype and is not production-ready.
