Can Autonomous AI Agents Perform Genuine Scientific Research?

The two artifacts approach the problem from opposite ends—one diagnostic, one constructive—but both reject surface-level fluency in favor of empirical integrity.

Shadow evaluations expose systematic failure

According to the arXiv preprint "Can AI agents conduct open-ended AI research? Early evidence from two case studies," the authors deploy "shadow evaluations," a protocol that tests frontier agents against unpublished NeurIPS 2026 submissions. The methodology is deliberately restrictive: blind the agent to ground truth, then audit its outputs against a held-out review record.

Five recurring failure modes are identified in autonomous AI research. The preprint frames these as structural, not cosmetic—each severs the claim chain linking hypothesis, method, code, and reported score. The full taxonomy and per-category incidence rates are not contained in the available snippet; the arXiv listing is the authoritative source.

The evaluation design is the primary contribution. Shadow evaluations eliminate the training-set contamination that has compromised prior benchmarks, where agents were tested on papers they had already ingested. Unpublished submissions reset the prior and force genuine reasoning rather than retrieval recall.

Chain-of-Evidence as a construction-time fix

Google Research introduces the Science One Framework, instantiated from the Chain-of-Evidence (CoE) conceptual framework. The write-up is authored by Rui Meng, Research Scientist, and Tomas Pfister, Director, Google Cloud.

CoE specifies two properties: every claim in a research artifact must carry a recorded evidence chain (completeness), and each chain must genuinely support the claim it attaches to (correctness). A claim is defined broadly—references, reported numbers, method descriptions, conclusions—each required to link back to a peer-reviewed paper, an experimental log line, the code that actually ran, or a results table. A hallucinated reference, an unreproducible score, and a misdescribed method collapse into the same failure class: a broken chain.

Reported empirical findings are sharp. Baseline autonomous research systems hallucinate up to 21% of their references and frequently misalign described methods with executed code. The Science One Framework reports zero phantom references and fully verifiable scores, while achieving state-of-the-art performance on MLE-Bench and Parameter-Golf.

The pipeline decomposes into three modules: Problem Investigator grounds literature via retrieved PDFs; the Discovery module explores and evaluates candidate solutions. Integration details for the third module are truncated in the public snippet.

Assessment

The two releases function as complementary diagnostics. Shadow evaluations establish that frontier agents fail in predictable, recurrent ways when tested on held-out research artifacts. Chain-of-Evidence offers a construction-time mitigation: build the evidence chain before writing the prose, then audit it mechanically via the CoE Audit protocol.

Neither artifact resolves the harder problem of open-ended hypothesis generation. Both, however, raise the methodological bar for what counts as verifiable evidence. Replication targets for practitioners: MLE-Bench, Parameter-Golf, and the unpublished NeurIPS 2026 submission pool used in the arXiv study. The full failure-mode taxonomy and the complete CoE Audit specification remain to be verified in subsequent revisions.