The Effects of Functional Emotions on LLM-Based Paper Review

Abstract

Large language models are increasingly used to support manuscript revision, peer review, and editorial assessment, raising concerns that their judgments may be influenced by cues unrelated to scientific quality. We investigate whether \emph{functional emotions}, operationalized as affect-like directions in activation space, provide a useful internal lens for auditing such sensitivity in AI-assisted paper review. Holding manuscript content fixed, we vary metadata concerning country of origin, institutional visibility, and declared AI use across 60 accepted and rejected papers from ICLR, NeurIPS, and ICML and three open-weight reviewer models. We measure matched changes in paper scores, recommendations, other structured review outcomes, and five model-specific affect-family readouts. Visible metadata generally makes reviews harsher: 13 of the 14 nonblind conditions reduce pooled recommendation relative to blind review after multiplicity correction, with the explicit no-AI declaration as the only exception. The largest pooled penalties occur for the Nigeria condition, the regional/global 500--1200 institution condition, and selected named-AI polishing declarations. Metadata also produces model-specific shifts in internal affect-family coordinates, while token-level diagnostics localize some of this representational salience to the metadata block. Finally, activation steering in GPT-OSS-20B shows that selected affect-like directions can causally alter review judgments under the tested intervention protocol.