Whether, Not Which: Mechanistic Interpretability Reveals Dissociable Affect Reception and Emotion Categorization in LLMs

• URL: http://arxiv.org/abs/2603.22295v1
• Date: Sun, 15 Mar 2026 15:11:45 GMT
• Title: Whether, Not Which: Mechanistic Interpretability Reveals Dissociable Affect Reception and Emotion Categorization in LLMs
• Abstract summary: We present the first clinical validity test of emotion circuit claims using mechanistic interpretability methods grounded in clinical psychology.<n>We discover two dissociable emotion processing mechanisms.<n>We introduce clinical stimulus methodology as a rigorous standard for testing emotion processing claims in large language models.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models appear to develop internal representations of emotion -- "emotion circuits," "emotion neurons," and structured emotional manifolds have been reported across multiple model families. But every study making these claims uses stimuli signalled by explicit emotion keywords, leaving a fundamental question unanswered: do these circuits detect genuine emotional meaning, or do they detect the word "devastated"? We present the first clinical validity test of emotion circuit claims using mechanistic interpretability methods grounded in clinical psychology -- clinical vignettes that evoke emotions through situational and behavioural cues alone, emotion keywords removed. Across six models (Llama-3.2-1B, Llama-3-8B, Gemma-2-9B; base and instruct variants), we apply four convergent mechanistic interpretability methods -- linear probing, causal activation patching, knockout experiments, and representational geometry -- and discover two dissociable emotion processing mechanisms. Affect reception -- detecting emotionally significant content -- operates with near-perfect accuracy (AUROC 1.000), consistent with early-layer saturation, and replicates across all six models. Emotion categorization -- mapping affect to specific emotion labels -- is partially keyword-dependent, dropping 1-7% without keywords and improving with scale. Causal activation patching confirms keyword-rich and keyword-free stimuli share representational space, transferring affective salience rather than emotion-category identity. These findings falsify the keyword-spotting hypothesis, establish a novel mechanistic dissociation, and introduce clinical stimulus methodology as a rigorous standard for testing emotion processing claims in large language models -- with direct implications for AI safety evaluation and alignment. All stimuli, code, and data are released for replication.

Related papers

• From Syntax to Emotion: A Mechanistic Analysis of Emotion Inference in LLMs [1.4692712025142225]
  We investigate the internal mechanisms of emotion recognition in large language models (LLMs)<n>By analyzing sparse feature activations across layers, we identify a consistent three-phase information flow.<n>We show that emotion representations comprise both shared features across emotions and emotion-specific features.
  arXiv  Detail & Related papers  (2026-04-28T17:03:16Z)
• AIPsy-Affect: A Keyword-Free Clinical Stimulus Battery for Mechanistic Interpretability of Emotion in Language Models [0.0]
  AIPsy-Affect is a 480-item clinical stimulus battery that removes the confound at the stimulus level.<n>It supports linear probing, activation patching, SAE feature analysis, causal ablation, and steering vector extraction.
  arXiv  Detail & Related papers  (2026-04-26T14:03:55Z)
• Memory-guided Prototypical Co-occurrence Learning for Mixed Emotion Recognition [56.00118641432005]
  We propose a Memory-guided Prototypical Co-occurrence Learning framework that explicitly models emotion co-occurrence patterns.<n>Inspired by human cognitive memory systems, we introduce a memory retrieval strategy to extract semantic-level co-occurrence associations.<n>Our model learns affectively informative representations for accurate emotion distribution prediction.
  arXiv  Detail & Related papers  (2026-02-24T04:11:25Z)
• Do LLMs "Feel"? Emotion Circuits Discovery and Control [54.57583855608979]
  We study the internal mechanisms that give rise to emotional expression and in controlling emotions in generated text.<n>This is the first systematic study to uncover and validate emotion circuits in large language models.
  arXiv  Detail & Related papers  (2025-10-13T12:24:24Z)
• SEER: The Span-based Emotion Evidence Retrieval Benchmark [8.124633573706761]
  We introduce the SEER (Span-based Emotion Evidence Retrieval) Benchmark to test Large Language Models' ability to identify the specific spans of text that express emotion.<n>We evaluate 14 open-source LLMs and find that, while some models approach average human performance on single-sentence inputs, their accuracy degrades in longer passages.
  arXiv  Detail & Related papers  (2025-10-03T20:15:24Z)
• Language Models (Mostly) Do Not Consider Emotion Triggers When Predicting Emotion [87.18073195745914]
  We investigate how well human-annotated emotion triggers correlate with features deemed salient in their prediction of emotions. Using EmoTrigger, we evaluate the ability of large language models to identify emotion triggers. Our analysis reveals that emotion triggers are largely not considered salient features for emotion prediction models, instead there is intricate interplay between various features and the task of emotion detection.
  arXiv  Detail & Related papers  (2023-11-16T06:20:13Z)
• Dynamic Causal Disentanglement Model for Dialogue Emotion Detection [77.96255121683011]
  We propose a Dynamic Causal Disentanglement Model based on hidden variable separation. This model effectively decomposes the content of dialogues and investigates the temporal accumulation of emotions. Specifically, we propose a dynamic temporal disentanglement model to infer the propagation of utterances and hidden variables.
  arXiv  Detail & Related papers  (2023-09-13T12:58:09Z)
• A Circular-Structured Representation for Visual Emotion Distribution Learning [82.89776298753661]
  We propose a well-grounded circular-structured representation to utilize the prior knowledge for visual emotion distribution learning. To be specific, we first construct an Emotion Circle to unify any emotional state within it. On the proposed Emotion Circle, each emotion distribution is represented with an emotion vector, which is defined with three attributes.
  arXiv  Detail & Related papers  (2021-06-23T14:53:27Z)
• Acted vs. Improvised: Domain Adaptation for Elicitation Approaches in Audio-Visual Emotion Recognition [29.916609743097215]
  Key challenges in developing generalized automatic emotion recognition systems include scarcity of labeled data and lack of gold-standard references. In this work, we regard the emotion elicitation approach as domain knowledge, and explore domain transfer learning techniques on emotional utterances.
  arXiv  Detail & Related papers  (2021-04-05T15:59:31Z)
• Emotion pattern detection on facial videos using functional statistics [62.997667081978825]
  We propose a technique based on Functional ANOVA to extract significant patterns of face muscles movements. We determine if there are time-related differences on expressions among emotional groups by using a functional F-test.
  arXiv  Detail & Related papers  (2021-03-01T08:31:08Z)
• Detecting Emotion Primitives from Speech and their use in discerning Categorical Emotions [16.886826928295203]
  Emotion plays an essential role in human-to-human communication, enabling us to convey feelings such as happiness, frustration, and sincerity. This work investigated how emotion primitives can be used to detect categorical emotions such as happiness, disgust, contempt, anger, and surprise from neutral speech. Results indicated that arousal, followed by dominance was a better detector of such emotions.
  arXiv  Detail & Related papers  (2020-01-31T03:11:24Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.