Search AI news

Compression of Emotion in Language Model Representations

arXiv:2611.08683v1 Announce Type: new Abstract: Large language models (LLMs) have been evaluated on their ability to identify emotions expressed in text and produce fluent emotional language. However, output predictions alone do not reveal how fine-grained emotion distinctions are organized internally. We study seven LLMs on GoEmotions, an English Reddit dataset annotated with 27 fine-grained emotion categories plus neutral. We compare model predictions and emotion-prompted final-token representations with the 28 GoEmotions categories and with coarser emotion-family and sentiment groupings. We find a consistent pattern of $\textit{overcompression}$ relative to the fine-grained GoEmotions taxonomy. Under the tested KMeans protocol, when we cluster their hidden representations without using emotion labels, the clearest split is into two groups that largely separate positive from negative emotion. Coarser emotion groupings are also recovered with fewer clusters than the full 28-category taxonomy. Yet linear probes retain substantial fine-grained predictive information. These results suggest that fine-grained emotion information remains accessible even though broader affective structure more strongly organizes the unsupervised geometry.

Research reveals that reward model scoring exhibits significant prompt dependency: pairwise evaluation fails to detect this bias.

An AI research paper published on arXiv on October 24, 2026, points out that current mainstream pairwise evaluation methods automatically cancel out the scoring components common to all responses under the same prompt, leaving a critical “prompt-level bias” in reward models unexamined. Using human ratings as a benchmark, the authors conducted an audit of 18 publicly available scorers and found that, on the HelpSteer2 dataset, 6 models exhibited 34%–64% of their rating variance attributable to the prompt itself—whereas human ratings show virtually no such phenomenon.

Why it matters：The study highlights a long-overlooked fundamental flaw in the alignment process of large language models: reward model scores are heavily biased toward the prompt text rather than the actual quality of the response. Its methodology—human‑reference auditing, prompt substitution, and variance decomposition—is reproducible, and its conclusions are grounded in empirical analysis across multiple mainstream open-source scorers, offering direct implications for preference optimization, RLHF, and safety‑assessment practices.