Elamur Research⌕
☰
home / papers / 2026 / A Unified View on Emotion Representation in Large Language Models

A Unified View on Emotion Representation in Large Language Models

Paper · 2026 · Aishwarya Maheswaran and Maunendra Sankar Desarkar

Summary

Earlier interpretability studies disagree on which layers of LLMs encode emotions. Some place them in early layers and others in intermediate ones. The authors argue that these conflicts come from differences in datasets and prompts. They test this with a single, broad experimental setup. They probe the residual stream, attention and MLP outputs of five instruction-tuned Llama and Gemma models on four emotion-labelled datasets, under three prompt styles:

  • no instruction (P0),
  • an explicit instruction (P1),
  • an instruction with few-shot examples (P2).

Unsupervised probing then extracts emotion reading vectors, which are compared across datasets. Finally, Chain-of-Thought (CoT) reasoning is evaluated on the harder EmoBench benchmark.

Emotion information becomes consistently available in later layers. The exact layers depend on the prompt and on how explicitly the emotion is expressed. Reading vectors from different datasets are highly similar and can be swapped with little loss. CoT, however, gives limited benefit and mostly reinforces the model's first answer. This suggests LLMs hold emotion representations they cannot yet reason with reliably, which matters for emotional-support applications.

Key points

  • Models: Llama-3.2-Instruct 1B/3B, Llama-3.1-Instruct 8B, Gemma-2-it 2B/9B.
  • Datasets: Twitter Emotions, RECCON (Daily Dialog split), the LLM-generated set of Zou et al., and Crowd-eVent, all restricted to Ekman's six emotions. EmoBench is used for the CoT analysis.
  • Method: Supervised probing trains logistic regression (cuml) on last-token hidden states. Activation patching and neuron-level analysis were set aside because they were unstable across prompts and datasets, with only about 10 shared emotion neurons per layer.
  • CoT on EmoBench: CoT raised Llama-3.1-8B accuracy from 0.265 to 0.425. Gemma-2-9B changed little (0.435 to 0.455), which the authors link to its lower initial entropy over the answer options.
  • Effect of hints and confidence: Hints describing the emotion's cause made CoT traces more accurate. CoT also lowered entropy whether or not the answer was correct.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Indian Institute of Technology HyderabadSangāreddi · India4 works · 108 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksSentiment & emotion · 972 works

Cite

Cite as Aishwarya Maheswaran and Maunendra Sankar Desarkar. 2026. A Unified View on Emotion Representation in Large Language Models. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-long.165

@inproceedings{maheswaran2026unified,
  title = {{A Unified View on Emotion Representation in Large Language Models}},
  author = {Aishwarya Maheswaran and Maunendra Sankar Desarkar},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-long.165},
  url = {https://doi.org/10.18653/v1/2026.eacl-long.165}
}

References

22 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Leonard Bereska and Stratis Gavves. 2024. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research. Survey Certification, Expert Certification. find
  2. P Ekman, W V Friesen, M O’Sullivan, A Chan, I Diacoyanni-Tarlatzis, K Heider, R Krause, W A LeCompte, T Pitcairn, and P E Ricci-Bitti. 1987. Universals and cultural differences in the judgments of facial expressions of emotion. J. Pers. Soc. Psychol., 53(4):712–717. find
  3. Gemma Team. 2025. Gemma 2: A family of open models from google. https://storage.googleapis.com/deepmind-media/ gemma/gemma-2-report.pdf. Technical report. ↗
  4. Yihuai Hong and Aldo Lipani. 2024. Interpretability-based tailored knowledge editing in transformers. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 3847–3858, Miami, Florida, USA. Association for Computational Linguistics. find
  5. Jaewook Lee, Woojin Lee, Oh-Woog Kwon, and Harksoo Kim. 2025. Do large language models have “emotion neurons”? investigating the existence and role. In Findings of the Association for Computational Linguistics: ACL 2025, pages 15617–15639, Vienna, Austria. Association for Computational Linguistics. find
  6. Jianhua Lin. 1991. Divergence measures based on the Shannon entropy. IEEE Transactions on Information Theory, 37(1):145–151. find
  7. Daria Lioubashevski, Tomer M. Schlank, Gabriel Stanovsky, and Ariel Goldstein. 2025. Looking beyond the top-1: Transformers determine top tokens in order. In Forty-second International Conference on Machine Learning. find
  8. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150, Portland, Oregon…
  9. Aishwarya Maheswaran, Caslon Chua, and Maunendra Sankar Desarkar. 2025. Probing the inherent ability of large language models for generating empathetic responses. In 2025 IEEE Swiss Conference on Data Science (SDS), pages 32–39. nostalgebraist. 2020. Interpreting GPT: The logit lens. LessWrong. find
  10. Dario Di Palma, Alessandro De Bellis, Giovanni Serve-dio, Vito Walter Anelli, Fedelucio Narducci, and Tommaso Di Noia. 2025. LLaMAs have feelings too: Unveiling sentiment and emotion representations in LLaMA models through probing. In Proceedings of the 63rd Annual Meeting of the Association for Com…
  11. Soujanya Poria, Navonil Majumder, Devamanyu Hazarika, Deepanway Ghosal, Rishabh Bhardwaj, Samson Yu Bai Jian, Pengfei Hong, Romila Ghosh, Abhinaba Roy, Niyati Chhaya, Alexander F. Gelbukh, and Rada Mihalcea. 2021. Recognizing emotion cause in conversations. Cogn. Comput., 13(5):1317– 1332. find
  12. Sahand Sabour, Siyang Liu, Zheyuan Zhang, June Liu, Jinfeng Zhou, Alvionna Sunaryo, Tatia Lee, Rada Mihalcea, and Minlie Huang. 2024. EmoBench: Evaluating the emotional intelligence of large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (… find
  13. Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. 2018. CARER: Contextualized affect representations for emotion recognition. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3687–3697, Brussels, Belgium. Association for Co…
  14. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–…
  15. Ran Song, Shizhu He, Shuting Jiang, Yantuan Xian, Shengxiang Gao, Kang Liu, and Zhengtao Yu. 2024. Does large language model contain task-specific neurons? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 7101–7113, Miami, Florida, USA. Association for… find
  16. Ala N. Tak, Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch. 2025. Mechanistic interpretability of emotion inference in large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 13090–13120, Vienna, Austria. Association for… find
  17. Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Xin Zhao, Furu Wei, and Ji-Rong Wen. 2024. Language-specific neurons: The key to multilingual capabilities in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volu…
  18. Curt Tigges, Oskar J. Hollinsworth, Atticus Geiger, and Neel Nanda. 2024. Language models linearly represent sentiment. In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 58–87, Miami, Florida, US. Association for Computational Linguistics. find
  19. Hugo Touvron, A. Grattafiori, and 1 others. 2024. The llama 3 herd of models. arXiv preprint.
  20. Enrica Troiano, Laura Oberländer, and Roman Klinger. 2023. Dimensional modeling of emotions in text with appraisal theories: Corpus creation, annotation reliability, and prediction. Computational Linguistics, 49(1):1–72. find
  21. Wenxuan Zhang, Yue Deng, Bing Liu, Sinno Pan, and Lidong Bing. 2024. Sentiment analysis in the era of large language models: A reality check. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3881–3906, Mexico City, Mexico. Association for Computational Linguistics.
  22. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, and 2 others. 2025. Representa…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/maheswaran-2026-unified-view-emotion-representation-large/