Emotion-associated representations in the feed forward gates of a language model
Identifying emotion-associated neurons in large language models does not establish how those neurons contribute to emotional behavior. This study investigates that distinction through three levels of explanation: representation, reconfiguration, and causal control. Statistically corrected feed-forward activation signatures in Qwen2.5-1.5B-Instruct exhibited structured overlap corresponding to a psychological circumplex model, with substantial but incomplete reproducibility across Chinese texts and their English translations. Four experimental task families investigated whether these signatures changed with behavioral demands and predicted model outputs. The preregistered internal measure failed held-out validation, with only two of fourteen hypotheses supported; post hoc corrected measures revealed task-dependent reconfiguration and several associations with behavioral measures. However, increased internal separation did not consistently improve discrimination accuracy, evidence-specific changes depended on the internal measure, and static neuronal overlap did not predict generated-text confusability under the designated classifier, although a secondary lexicon analysis showed the predicted association. Suppression of emotional expression was not accompanied by a reliably detected change at the sampled pre-generation position. These findings demonstrate that emotion-associated activation structure and task responsiveness do not, by themselves, provide a general explanation of emotional behavior. The study establishes the need to distinguish representational association, task-dependent reconfiguration, and causal control, the last of which remains untested.
Authors
- Sankar Balasubramanian (ORCID: https://orcid.org/0000-0001-5844-6273)
- Rayan Upadhyay
Institutions
- Indian Institute of Science Bangalore (IN)
Publication Details
- Journal
- Journal of High School Science
- Published
- 2026-11-20
- DOI
- https://doi.org/10.64336/001c.175619
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- article
- Field-Weighted Citation Impact
- 0.00