Elamur Research⌕
☰
home / papers / 2026 / How Do LLMs Generate Contrastive Sentiments? A Mechanistic Perspective

How Do LLMs Generate Contrastive Sentiments? A Mechanistic Perspective

Paper · 2026 · Van Bach Nguyen, Jörg Schlötterer and Christin Seifert

Summary

The paper asks what happens inside a language model when it rewrites a text so that its sentiment flips (positive to negative or the reverse) while the rest of the wording stays as close as possible to the original. Using mechanistic interpretability tools (logit-lens analysis, transcoders and a circuit discovery procedure adapted from Dunefsky et al., 2024), the authors examine GPT-2, with additional checks on Pythia. They find two competing processes. A small group of attention heads tends to copy the original sentiment, which works against the task. A separate circuit of MLP layers and attention heads combines the target label with the original valenced word, and the MLPs mainly encode the opposite sentiment. The authors then steer the model by scaling selected attention heads, including scaling by negative factors so that a harmful head's effect is inverted. On SST-2 sentences, this intervention greatly raises GPT-2's flip rate while barely increasing edit distance. The work shows that targeted internal edits can improve a model's behavior on a specific task without retraining.

Key points

  • Two mechanisms. In GPT-2, heads L10H1 and L10H4 preserve the input's valenced word. The transformation circuit combines that word with the contrastive label token, and MLP layers can influence one another (e.g., MLP8 to MLP9). Transcoder features help interpret the MLPs; for example, MLP0 feature TC318 fires mainly on negative tokens.
  • Steering method. The intervention scales a head's attention by a factor α tested over 2^k, with k from 0 to 10. A negative α converts a head that pushes the wrong way into one that helps.
  • Results on head-level steering. Amplifying L7H5 at the target-label position, or negating L10H1, improves both the logit difference and the success rate. Applying both modifications together does not stack: the result matches amplifying L7H5 alone.
  • Downstream test. On SST-2, amplifying L7H5 raises GPT-2's flip rate from under 50% to over 90%. The token-level Levenshtein distance changes by less than 10% compared with no intervention. Amplifying head L13H10 in Pythia gives similar results.
  • Limitations and code. The authors note that the prompts are simple and that the interventions may have unintended side effects. Code is available at github.com/bach1292/contrastive_generation_MI.

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

🎓Philipps University of MarburgMarburg · Germany8 works · 226 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 3Research subjects and languages the work is about.Language models & pretraining · 3,320 worksSentiment & emotion · 972 worksInterpretability · 130 works

Cite

Cite as Van Bach Nguyen, Jörg Schlötterer and Christin Seifert. 2026. How Do LLMs Generate Contrastive Sentiments? A Mechanistic Perspective. 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.311

@inproceedings{nguyen2026how,
  title = {{How Do LLMs Generate Contrastive Sentiments? A Mechanistic Perspective}},
  author = {Van Bach Nguyen, Jörg Schlötterer and Christin Seifert},
  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.311},
  url = {https://doi.org/10.18653/v1/2026.eacl-long.311}
}

References

39 references, 12 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. Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, and 1 others. 2023. Pythia: A suite for analyzing large language models across training and scaling. In International Co… find
  3. Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, and 1 others. 2023. Towards monosemanticity: Decomposing language models with dictionary learning. Transformer Circuits Thread, 2. find
  4. Jannik Brinkmann, Abhay Sheshadri, Victor Levoso, Paul Swoboda, and Christian Bartelt. 2024. A mechanistic analysis of a transformer trained on a symbolic multi-step reasoning task. In Findings of the Association for Computational Linguistics: ACL 2024, pages 4082–4102, Bangkok, Thailand. Associatio… find
  5. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, and… find
  6. Bilal Chughtai, Lawrence Chan, and Neel Nanda. 2023. A toy model of universality: Reverse engineering how networks learn group operations. In International Conference on Machine Learning, pages 6243– 6267. PMLR. find
  7. Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. 2023. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36:16318–16352. find
  8. Guy Dar, Mor Geva, Ankit Gupta, and Jonathan Berant. 2023. Analyzing transformers in embedding space. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16124–16170, Toronto, Canada. Association for Computational Linguistics.
  9. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv–2407.
  10. Jacob Dunefsky, Philippe Chlenski, and Neel Nanda. 2024. Transcoders find interpretable llm feature circuits. Advances in Neural Information Processing Systems, 37:24375–24410. find
  11. Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. 2022. Toy models of superposition. Preprint, a… find
  12. Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, and 6 others. 2021. A mathe… find
  13. Jiahai Feng and Jacob Steinhardt. 2024. How do language models bind entities in context? In The Twelfth International Conference on Learning Representations. find
  14. Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. 2023. Dissecting recall of factual associations in auto-regressive language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12216–12235, Singapore. Association for Computational… find
  15. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484–5495, Online and Punta Cana, Dominican Republic. Association for Computational… find
  16. Marco Guerini, Carlo Strapparava, and Oliviero Stock. 2008. Valentino: A tool for valence shifting of natural language texts. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), Marrakech, Morocco. European Language Resources Association (ELRA). find
  17. Michael Hanna, Ollie Liu, and Alexandre Variengien. 2023. How does gpt-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model. In Advances in Neural Information Processing Systems, volume 36, pages 76033–76060. Curran Associates, Inc.
  18. Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. 2024. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations. find
  19. Mingyu Jin, Qinkai Yu, Jingyuan Huang, Qingcheng Zeng, Zhenting Wang, Wenyue Hua, Haiyan Zhao, Kai Mei, Yanda Meng, Kaize Ding, Fan Yang, Mengnan Du, and Yongfeng Zhang. 2025. Exploring concept depth: How large language models acquire knowledge and concept at different layers? In Proceedings of the…
  20. Kenneth Li, Aspen K Hopkins, David Bau, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Emergent world representations: Exploring a sequence model trained on a synthetic task. In The Eleventh International Conference on Learning Representations. find
  21. Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. 2023a. Progress measures for grokking via mechanistic interpretability. In International Conference on Learning Representations.
  22. Neel Nanda, Andrew Lee, and Martin Wattenberg. 2023b. Emergent linear representations in world models of self-supervised sequence models. In Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 16–30, Singapore. Association for Computational Linguist… find
  23. Van Bach Nguyen, Christin Seifert, and Jörg Schlötterer. 2024a. CEval: A benchmark for evaluating counterfactual text generation. In Proceedings of the 17th International Natural Language Generation Conference, pages 55–69, Tokyo, Japan. Association for Computational Linguistics. find
  24. Van Bach Nguyen, Paul Youssef, Christin Seifert, and Jörg Schlötterer. 2024b. LLMs for generating and evaluating counterfactuals: A comprehensive study. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 14809–14824, Miami, Florida, USA. Association for Computational Lin… find
  25. Nostalgebraist. 2020. Interpreting gpt: The logit lens. Blog post. find
  26. Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. 2020. Zoom in: An introduction to circuits. Distill. Https://distill.pub/2020/circuits/zoom-in. find
  27. Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Scott Johnston, Andy Jones, Jackson Kernion, Liane Lovitt, and 7 others. 2022. In-con…
  28. Francesco Ortu, Zhijing Jin, Diego Doimo, Mrinmaya Sachan, Alberto Cazzaniga, and Bernhard Schölkopf. 2024. Competition of Mechanisms: Tracing How Language Models Handle Facts and Counterfactuals. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: L…
  29. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. find
  30. Tilman Rauker, Anson Ho, Stephen Casper, and Dylan Hadfield-Menell. 2023. Toward Transparent AI: A Survey on Interpreting the Inner Structures of Deep Neural Networks . In 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 464–483, Los Alamitos, CA, USA. IEEE Computer Soc… find
  31. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in BERTology: What we know about how BERT works. Transactions of the Association for Computational Linguistics, 8:842–866.
  32. Alexis Ross, Ana Marasović, and Matthew Peters. 2021. Explaining NLP models via minimal contrastive editing (MiCE). In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3840–3852, Online. Association for Computational Linguistics.
  33. 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–…
  34. Adly Templeton, Joshua Batson, Adam Jermyn, and Chris Olah. 2024a. Predicting future activations, january 2024. URL https://transformer-circuits.pub/2024/jan-update/index. html\# predict-future. ↗
  35. Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, Nicholas L Turner, Callum McDougall, Monte MacDiarmid, C. Daniel Freeman, Theodore R. Sumers, Edward Rees, Joshua Batson, Adam Jermyn, an… find
  36. 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
  37. Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent abilities of large language models. Transactions on Ma…
  38. Tongshuang Wu, Marco Tulio Ribeiro, Jeffrey Heer, and Daniel Weld. 2021. Polyjuice: Generating counterfactuals for explaining, evaluating, and improving models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on N…
  39. Qinan Yu, Jack Merullo, and Ellie Pavlick. 2023. Characterizing mechanisms for factual recall in language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9924–9959, Singapore. Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nguyen-2026-how-do-llms-generate-contrastive/