As submission volumes to artificial intelligence conferences surge, peer review faces mounting pressure on speed, quality, and accountability. While large language models are being deployed in pilot programs to assist with paper triage and scoring, earlier studies on conference data show that AI-assisted reviewers skew lenient, assigning higher scores that distort outcomes for borderline submissions. In response, researchers from Zhejiang University, Nanyang Technological University, and HKUST have introduced a process-centric diagnostic benchmark that shifts the evaluation focus away from binary acceptance accuracy toward a granular audit of intermediate reasoning chains.
Formalizing the Review Chain
Traditional evaluations treat automated review as a single-step decision or assess generated texts in isolation. To test whether models actually ground their final verdicts in sound evidence, the consortium formalized the evaluation pipeline into a sequential tuple: source document (x), structured summary (zs), critical analysis (zc), actionable recommendations (zr), and the final decision (y). Heterogeneous historical review records from PeerRead, NLPeer ARR-22, and OpenReview-ICLR were converted into standardized, process-aligned datasets to enable systematic auditing.
The benchmark measures performance against a baseline of direct decision prediction from raw paper content. It executes stage-level evaluation, chain-consistency tracking, and interventional sensitivity analysis by comparing human Gold-process variables against model-generated Predicted-process variables.
"Although model-generated intermediate review texts show relatively high local consistency across adjacent stages, the final decisions are not consistently supported by the preceding review evidence."
This structural disconnect between adjacent phrasing and final judgment proves that fluent intermediate prose often functions as a veneer over hallucinated or ungrounded conclusions. Across experiments covering three datasets and six models, human Gold-process variables consistently yielded higher decision validity than model-predicted variables—a gap that remained remarkably stable across datasets, seeds, and architectures.
What This Means
For enterprise leaders, the implications extend far beyond academic paper screening. Deploying generative AI for complex due diligence, regulatory compliance, and multi-stage contract analysis requires moving past superficial output metrics and end-to-end black-box confidence scores. Implementing step-by-step diagnostic auditing provides the architectural blueprint needed to verify that intermediate logic actually supports final corporate decisions before signing off on automated risk assessments.