Why Summaries Turn Neutral: Policy Attribution for Sentiment Drift in Reinforcement Learning from Human Feedback
Mikhail Krasitskii, Alexander Gelbukh, Olga Kolesnikova, Grigori Sidorov
Abstract
Reinforcement learning with human feedback (RLHF) aligns LLMs with human preferences, improving summarization fluency and safety, but causes sentiment drift: overly neutral summaries stripped of emotional nuance. We diagnose why RL acts as a sentiment neutralizer and present Policy Attribution, a framework using gradient and logit decomposition to trace drift to reward model (RM) signals and KL (Kullback-Leibler) penalty. Sentiment drift reflects a strategic bias toward "low-risk" tokens maximizing expected rewards under preference uncertainty (Stiennon et al., 2020; Gao, Schulman, and Hilton, 2023). On Reddit TL;DR and CNN/DailyMail, RLHF summaries get higher rewards but show 30-40% lower sentiment variance. Cross-lingual analysis across eight languages shows language-independent drift, with morphologically richer languages more suppressed (Krasitskii et al., 2026). We propose and validate a sentiment-aware regularization technique reducing drift by 18-22% without harming summary quality. The code and toolkit will be public.
Create a lesson
Related papers
Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations
Leon Bergen, Usha Bhalla, Andrew Lee et al.
Reporting Practice Matters: The Impact of Reference Choice on Chest X-ray Report Evaluation
Daniel P. Jeong, Charles Q. Li, Hossein Hosseiny et al.
Long-Lived Characters, Local Inference: Incremental Memory Maintenance for Game NPCs
Zimu Xu
Beyond Outcomes: Dual-View Relational Learning for Efficient Agent Benchmarking
Xinshuai Guo, Junjie Wu, Dolly Deng et al.
How Much is a Human Right Worth? ECtHR-NPD: A Benchmark for Predicting Non-Pecuniary Damage Awards
Yanyi Pu, Damian A. Gonzalez-Salzberg, Zheng Yuan et al.
Structured Claim-Level Discourse Representations for Dense Health Narratives
Farnoushsadat Nilizadeh, Elham Pourabbas Vafa, Shirin Nilizadeh et al.