Agent Systems for Academic Research Automation

Pierfrancesco Beneventano, Riccardo Neumarker, Theodoros Evgeniou, Marc Gong Bacvanski, Kushagra Tiwary, Emanuele Rimoldi, Mehdi Hajoub, Yulu Gan, Qianli Liao, Mahmoud Abdelmoneum, Liu Ziyin, Tomer Galanti, Andrea Pinto, Tomaso Poggio

Additional Submission Instructions: For the camera-ready version, please include the author names and affiliations, funding disclosures, and acknowledgements.

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: AI scientists, research agents, academic research automation, verifier-matched autonomy, scientific verification, claim-level evaluation, provenance, governance

TL;DR: AI scientist systems should be evaluated by the artifacts they produce and the verifiers available for their claims, not by agent count or nominal autonomy.

Abstract: AI scientist systems increasingly propose hypotheses, plan experiments, execute analyses, draft papers, and participate in review-like workflows. As these systems move from local assistance toward producing claim-bearing scientific artifacts, aggregate descriptors such as agent count, tool use, or autonomy no longer specify what a system contributes, what artifact it produces, or what can verify the claims it helps make. We introduce verifier-matched autonomy: the principle that a system's apparent autonomy is meaningful only relative to the strength, cost, and latency of the verifier available for those claims. We develop a three-dimensional framework organized by research-lifecycle coverage, artifact type, and verification regime, and use it to position citation-grounded writing systems, survey generators, full-paper research pipelines, review agents, and domain-specific discovery systems. This framework treats labels such as tool, co-author, and founder as workflow-specific roles rather than system-wide properties. We derive implications for evaluation, provenance, and governance, emphasizing closure failure: the risk that claim-bearing scientific artifacts appear settled before their underlying claims have been adequately verified.

Submission Number: 279