RubricReviewer's AI peer review outperforms humans with structured rubrics

Peer review at major venues is under unprecedented submission pressure, motivating the use of LLMs as review assistants. Existing LLM-based reviewers have structural limitations: they map manuscripts directly to reviews, leaving the underlying rubric implicit, and each prevailing paradigm captures only half of a good review. RubricReviewer addresses these issues by making rubric generation an explicit intermediate step, so both review generation and final assessment are conditioned on paper-adaptive rubrics. It combines a training-free agent called Scout, which gathers external evidence, with a human-aligned trained model called Aligner, which consumes that evidence. In experiments on real-world submissions, RubricReviewer produces reviews that are markedly more comprehensive and more discriminative than prior systems, and it shows the strongest robustness against adversarial prompt-injection attacks. Ablation studies confirm the necessity of each component.