Skip to content
Research · Nov 9, 2026

Study finds large language models overcompress fine-grained emotion distinctions in internal representations

Analysis of seven LLMs on GoEmotions shows unsupervised clustering of hidden states primarily separates positive from negative emotion, while fine-grained distinctions remain probeable but less geometrically salient.

Trust84
HypeLow hype

1 source · cross-referenced

ShareXLinkedInEmail
TL;DR
  • Seven LLMs were evaluated on the GoEmotions dataset with 27 fine-grained emotion categories plus neutral.
  • Unsupervised clustering of hidden representations largely collapses distinctions to positive vs. negative emotion.
  • Fine-grained emotion information remains accessible via linear probes despite geometric overcompression.
  • Findings suggest broader affective structure dominates unsupervised geometry in LLMs.

Researchers evaluated seven large language models on the GoEmotions dataset, which annotates English Reddit posts with 27 fine-grained emotion categories plus neutral. The study compared model predictions and emotion-prompted final-token representations against the full 28-category taxonomy, as well as coarser emotion-family and sentiment groupings.

Using a KMeans clustering protocol without emotion labels, the team found that hidden representations most clearly split into two groups: positive versus negative emotion. Coarser groupings were recoverable with fewer clusters than the full 28-category taxonomy, indicating a consistent pattern of overcompression relative to the fine-grained labels.

Despite this geometric compression, linear probes retained substantial fine-grained predictive information, suggesting that detailed emotion distinctions remain accessible to supervised methods even if they are not the dominant structure in the unsupervised geometry.

The authors conclude that broader affective structure—such as positive/negative sentiment—dominates the unsupervised organization of model representations, while fine-grained emotion categories are less geometrically salient but still probeable.

Sources
  1. 01arXiv cs.CL — Compression of Emotion in Language Model Representations
Also on Research

Stories may contain errors. Dispatch is assembled with AI assistance and curated by human editors; despite the trust-score filter, mistakes happen. We correct publicly — every article links to its revision history. Nothing here is financial, legal, or medical advice. Verify before relying on any claim.

© 2026 Dispatch. No ads. No sponsorships. No paid placement. Reader-supported via Ko-fi.

Built by a person who cares about honest AI news.