Year: 2026
Type: preprint
Abstract: Scientific peer review is about to break under its own weight. Submission volume is already outpacing the reviewer pool, and the upcoming era of human directed AI generated research will multiply the load by orders of magnitude. We present PeerLens, a domain adaptive multi agent LLM system that produces a full referee report from a manuscript PDF using eight specialized agents (classification, claim extraction, dimensional analysis, symbolic verification, novelty search, structure and narrative review, domain expert review, and reproducibility review), each tuned to the manuscript's field. We evaluate PeerLens on 64 papers across physics, biology, mathematics, and chemistry (drawn from SciPost, eLife, F1000Research). On the 51 papers with available first round human reviewer reports, PeerLens agreed with the reviewers or diverged by being stricter, always on concrete technical grounds such as broken statistical assumptions, dimensionally malformed metrics, definitions that void their own novelty claims, and unresolved mechanistic dependencies. We observed no fabricated citations or invented technical claims. The strictness is a feature, not a defect: it reflects criteria based enforcement of disciplinary standards at scale, which is a prerequisite for peer review to continue functioning under the volume of manuscripts the community is about to face.
Source: Research Square
Authors: Emre Dil, Talha Zafer
Institutions: Dalhousie University, Imperial College London
Language: English
Subfield: Library and Information Sciences
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225412778