LLM Emotion Representations Show Overcompression While Fine-Grained Signals Remain Decodable
Summary:Using the 28-category GoEmotions taxonomy, the study examines output predictions, emotion-prompted final-token representations, and hidden states from seven LLMs. Unsupervised KMeans clustering consistently emphasizes a broad positive-versus-negative split, while coarser affective groupings can be recovered with fewer clusters than the full label set. Nevertheless, linear probes retain substantial information for predicting fine-grained emotions. The results indicate that LLMs do not simply dis
Core Finding
The study of seven LLMs on GoEmotions reveals an important distinction: the ability to identify fine-grained emotions does not imply that the internal representation space is organized according to the same fine-grained taxonomy.
The authors compare direct model predictions, emotion-prompted final-token representations, and hidden states aligned with several label hierarchies. GoEmotions contains 27 fine-grained emotion classes plus neutral, while the study also constructs coarser emotion-family and sentiment groupings.
In label-free KMeans experiments, the most prominent and stable separation is typically a two-cluster split that broadly corresponds to positive and negative emotions. Intermediate affective groupings can also be recovered with fewer clusters, but the full 28-class structure does not emerge as clearly. The authors describe this as overcompression relative to the GoEmotions taxonomy.
Why the Geometry Does Not Match the Labels
LLM hidden states need not encode a supervised label set as their coordinate system. Multi-layer Transformers learn many entangled semantic axes, including topic, speaker intent, syntax, event appraisal, and sentiment polarity.
Final-token representations are additionally shaped by normalization layers, residual connections, and attention aggregation across many layers. Under an emotion prompt, the model often needs to construct a representation in an answer-oriented emotion space. Multiple fine-grained labels may share the same answer prefix or high-level semantic direction, making their vectors appear closer under an unsupervised distance metric.
KMeans can amplify this effect. Its use of Euclidean distance assumes approximately spherical clusters and represents each cluster through a centroid. Real emotion taxonomies are hierarchical and overlapping: excitement and joy, or worry and sadness, can form continuous gradients rather than balanced, independent balls. Changing the distance metric, layer, initialization, or cluster count could materially alter the recovered structure.
Fine-Grained Information Remains Accessible
The crucial counterevidence comes from linear probes. If unsupervised clustering recovers only a positive-versus-negative split while a linear classifier can still predict many fine-grained emotions, those distinctions must remain present as linearly accessible information.
This creates two distinct engineering access paths:
- Unsupervised paths extract the most geometrically prominent, low-dimensional, and stable axes. They are suitable for coarse affective retrieval, routing, and caching.
- Supervised probe paths recover fine-grained distinctions that may depend more heavily on labels, prompt templates, and linear decision boundaries.
Consequently, “only two clusters” does not prove that an LLM cannot distinguish anger, disgust, or fear. It shows only that, under the tested KMeans geometry, the dominant organizing axis is broader sentiment polarity.
Engineering Trade-offs and Evaluation Boundaries
This is better understood as representation diagnosis than as a new model-capability benchmark. Several limitations matter:
- Single data source: GoEmotions consists of English Reddit posts, where context dependence, sarcasm, mixed emotions, and cultural differences affect the findings.
- Sensitivity to the clustering protocol: The summary explicitly identifies KMeans but provides limited information about distance functions, layer selection, random seeds, cluster counts, and stability metrics.
- Final tokens are not the full state: Intermediate layers, attention heads, and contextual positions may contain substantially richer affective structure.
- Label overlap: GoEmotions categories are not perfectly mutually exclusive, and real texts may contain multiple emotions. Single-label probes can therefore understate or distort the underlying structure.
- Limited model and architecture coverage: Seven models are insufficient to establish that the phenomenon holds universally across LLMs, MoE architectures, or instruction-tuning regimes.
The implications are also sensitive to how emotion labels were represented during training. Developers should not directly transfer findings from English social platforms to customer support, medical conversations, or multilingual settings without additional validation.
Practical Guidance
The most actionable lesson is to match the representation-extraction method to the task. Coarse positive-versus-negative routing can use unsupervised clustering for higher throughput, whereas fine-grained recognition should use dedicated probes and may benefit from ensembles, temporal windows, or broader context.
For interpretability research, researchers can compare probe accuracy and clustering stability across layers. If fine-grained information peaks in middle layers while the final token mainly amplifies polarity, then probing and activation steering at intermediate layers may be more informative than analyzing the final output alone.
Causal use of emotion directions requires additional caution. A direction labeled “positive” may also encode topic, politeness, or toxicity. Moving along it should therefore be tested through causal interventions, ablations, and human review rather than treated as a ready-made control knob.
Overall Assessment
The research separates emotion recognition from the organization of emotion in representation space. Fine-grained labels remain linearly readable, but unsupervised geometry is dominated by sentiment polarity and broader affective families. This compression can reduce redundancy, yet downstream systems must use the appropriate readout mechanism.
Its broader contribution is methodological: clustering quality should not be treated as a direct measure of a model's emotional depth. The work also provides a useful framework for building switchable emotion encoders, lightweight adapters that preserve fine-grained labels, and studies of affective compression and causal controllability.
— END —Source: ArXiv NLP/LLM (cs.CL) (2026-11-09)
Tags: #LLM Representations #Emotion Recognition #GoEmotions #Linear Probing #KMeans Clustering
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments