Elamur Research⌕
☰
home / papers / 2026 / Understanding and Mitigating Political Stance Cross-topic Generalization in Large Language Models

Understanding and Mitigating Political Stance Cross-topic Generalization in Large Language Models

Paper · 2026 · Jiayi Zhang, Shu Yang, Junchao Wu et al.

Summary

Fine-tuning a large language model on one political topic can shift its stance on unrelated topics as well. For example, right-leaning tuning on race can also move the model's views on the economy. The authors examine why this cross-topic coupling happens and how to limit it. They introduce PNLAC (Political Neuron Localization through Activation Contrasting), which compares FFN neuron activations in left- and right-leaning fine-tuned variants and scores each neuron with a Political Activation Shift (PAS). Set operations across topics then separate general political neurons, which affect many topics, from topic-specific ones. The authors adapt activation patching to open-ended generation and use it to confirm that both neuron types exist, mainly in middle and later layers. Building on this, InhibitFT blocks gradient updates to the general political neurons during fine-tuning, so only topic-specific neurons change. This cuts unintended stance transfer by about 20% while keeping the targeted topic shift and response quality. The work gives a mechanistic explanation of political stance spillover and a practical way to make targeted stance adjustments.

Key points

  • Fine-tuning uses IDEOINST, which covers six topics: crime, economy, gender, immigration, race and science. The Political Compass and the IDRlabs Ideologies Test serve as additional evaluation sets.
  • The models tested are Llama-3.1-8B, Llama-3.2-3B, Qwen-2.5-7B and Qwen-2.5-3B. Average reductions in cross-topic coupling were 20.6%, 20.3%, 19.9% and 19.8% respectively, measured by RMSE.
  • The baselines were standard right-leaning fine-tuning and InhibitFT with randomly selected frozen neurons. The random variant behaved much like ordinary fine-tuning. For Llama-3.1-8B tuned on crime, RMSE fell from 0.847 to 0.433 with InhibitFT.
  • CoLA and MNLI scores show that response utility is preserved, with results slightly better than plain fine-tuning.
  • In the ablation study, inhibiting roughly 5% of neurons (γ = 5%) was enough to decouple stances across topics.

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

🎓University of MacauMacao · China33 works · 2,223 citations🎓King Abdullah University of Science and TechnologyJeddah · Saudi Arabia27 works · 1,084 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 worksSocial Sciences · 1,772 worksSafety Research · 543 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Jiayi Zhang, Shu Yang, Junchao Wu, Derek F. Wong and Di Wang. 2026. Understanding and Mitigating Political Stance Cross-topic Generalization in Large Language Models. Proceedings of the 64th Annual Meeting 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.acl-long.1374

@inproceedings{zhang2026understanding,
  title = {{Understanding and Mitigating Political Stance Cross-topic Generalization in Large Language Models}},
  author = {Jiayi Zhang, Shu Yang, Junchao Wu, Derek F. Wong and Di Wang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1374},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1374}
}

References

53 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Maarten Buyl, Alexander Rogiers, Sander Noels, Iris Dominguez-Catena, Edith Heiter, Raphaël Romero, Iman Johary, Alexandru Cristian Mara, Jefrey Lijffijt, and Tijl De Bie. 2024. Large language models reflect the ideology of their creators. CoRR, abs/2410.18417. — [arXiv:2410.18417](https://arxiv.org/abs/2410.18417) find
  2. Jianhui Chen, Xiaozhi Wang, Zijun Yao, Yushi Bai, Lei Hou, and Juanzi Li. 2024a. Finding safety neurons in large language models. CoRR, abs/2406.14144. — [arXiv:2406.14144](https://arxiv.org/abs/2406.14144) find
  3. Kai Chen, Zihao He, Jun Yan, Taiwei Shi, and Kristina Lerman. 2024b. How susceptible are large language models to ideological manipulation? arXiv preprint arXiv:2402.11725. find
  4. Ruizhe Chen, Tianxiang Hu, Yang Feng, and Zuozhu Liu. 2024c. Learnable privacy neurons localization in language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, ACL 2024 - Short Papers, Bangkok, Thailand, August 11-16, 2024, pages 256–264. Associati… find
  5. Xin Chen, Junchao Wu, Shu Yang, Runzhe Zhan, Zeyu Wu, Min Yang, Shujian Huang, Lidia S Chao, and Derek F Wong. 2026. Neuron-aware data selection in instruction tuning for large language models. arXiv preprint arXiv:2603.13201. — [arXiv:2603.13201](https://arxiv.org/abs/2603.13201) find
  6. Tavishi Choudhary. 2024. Political bias in large language models: A comparative analysis of chatgpt-4, perplexity, google gemini, and claude. IEEE Access. find
  7. Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2021. Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696. find
  8. Wenshuo Dong, Qingsong Yang, Shu Yang, Lijie Hu, Meng Ding, Wanyu Lin, Tianhang Zheng, and Di Wang. 2025. Understanding and mitigating cross-lingual privacy leakage via language-specific and universal privacy neurons. arXiv preprint arXiv:2506.00759. — [arXiv:2506.00759](https://arxiv.org/abs/2506.00759) find
  9. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  10. Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris Bertsimas. 2023. Finding neurons in a haystack: Case studies with sparse probing. arXiv preprint arXiv:2305.01610. find
  11. Yihuai Hong, Yuelin Zou, Lijie Hu, Ziqian Zeng, Di Wang, and Haiqin Yang. 2024. Dissecting fine-tuning unlearning in large language models. arXiv preprint arXiv:2410.06606. find
  12. Lijie Hu, Liang Liu, Shu Yang, Xin Chen, Hongru Xiao, Mengdi Li, Pan Zhou, Muhammad Asif Ali, and Di Wang. 2024. A hopfieldian view-based interpretation for chain-of-thought reasoning. arXiv preprint arXiv:2406.12255. — [arXiv:2406.12255](https://arxiv.org/abs/2406.12255) find
  13. Xinke Jiang, Yue Fang*, Rihong Qiu*, Haoyu Zhang, Yongxin Xu, Hao Chen, Wentao Zhang, Ruizhe Zhang, Yuchen Fang, Xu Chu, and 1 others. 2025a. Tc-rag: Turing-complete rag’s case study on medical llm systems.
  14. Xinke Jiang, Ruizhe Zhang*, Yongxin Xu*, Rihong Qiu*, Yue Fang, Zhiyuan Wang, Jinyi Tang, Hongxin Ding, Xu Chu, Junfeng Zhao, and 1 others. 2025b. Hykge: A hypothesis knowledge graph enhanced framework for accurate and reliable medical llms responses. find
  15. Xinyan Jiang, Lin Zhang, Jiayi Zhang, Qingsong Yang, Guimin Hu, Di Wang, and Lijie Hu. 2025c. Msrs: Adaptive multi-subspace representation steering for attribute alignment in large language models. arXiv preprint arXiv:2508.10599. — [arXiv:2508.10599](https://arxiv.org/abs/2508.10599) find
  16. Yongqi Leng and Deyi Xiong. 2025. Towards understanding multi-task learning (generalization) of llms via detecting and exploring task-specific neurons. In Proceedings of the 31st International Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025, pages 2969–2987.…
  17. Jared Moore, Tanvi Deshpande, and Diyi Yang. 2024. Are large language models consistent over value-laden questions? In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024, pages 15185– 15221. Association for Computational Linguistics. find
  18. Ethan Perez, Sam Ringer, Kamile Lukosiute, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, Andy Jones, Anna Chen, Benjamin Mann, Brian Israel, Bryan Seethor, Cameron McKinnon, Christopher Olah, Da Yan, Daniela Amodei, and 44 others. 2023. Dis… — Discovering Language Model Behaviors with Model-Written Evaluations
  19. Pagnarasmey Pit, Xingjun Ma, Mike Conway, Qingyu Chen, James Bailey, Henry Pit, Putrasmey Keo, Watey Diep, and Yu-Gang Jiang. 2024. Whose side are you on? investigating the political stance of large language models. arXiv preprint arXiv:2403.13840. and opinions in large language models. In Proceedin… — [arXiv:2403.13840](https://arxiv.org/abs/2403.13840) find
  20. David Rozado. 2024. The political preferences of llms. CoRR, abs/2402.01789. — [arXiv:2402.01789](https://arxiv.org/abs/2402.01789) find
  21. Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect? In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research,…
  22. Alessandro Stolfo, Ben Wu, Wes Gurnee, Yonatan Belinkov, Xingyi Song, Mrinmaya Sachan, and Neel Nanda. 2024. Confidence regulation neurons in language models. arXiv preprint arXiv:2406.16254. — [arXiv:2406.16254](https://arxiv.org/abs/2406.16254) find
  23. Jinyan Su, Terry Zhuo, Di Wang, and Preslav Nakov. 2023a. Detectllm: Leveraging log rank information for zero-shot detection of machine-generated text. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12395–12412.
  24. Jinyan Su, Terry Yue Zhuo, Jonibek Mansurov, Di Wang, and Preslav Nakov. 2023b. Fake news detectors are biased against texts generated by large language models. arXiv preprint arXiv:2309.08674. — [arXiv:2309.08674](https://arxiv.org/abs/2309.08674) find
  25. Yi Su, Jiayi Zhang, Shu Yang, Xinhai Wang, Lijie Hu, and Di Wang. 2025. Understanding how value neurons shape the generation of specified values in llms. arXiv preprint arXiv:2505.17712. find
  26. 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…
  27. Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020. Investigating gender bias in language models using causal mediation analysis. Advances in neural information processing systems, 33:12388– 12401. find
  28. Niklas Retzlaff. 2024. Political biases of chatgpt in different languages. Preprints: Preprints. find
  29. Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. 2022a. Interpretability in the wild: a circuit for indirect object identification in gpt-2 small. Preprint, arXiv:2211.00593. find
  30. Paul Röttger, Valentin Hofmann, Valentina Pyatkin, Musashi Hinck, Hannah Kirk, Hinrich Schütze, and Dirk Hovy. 2024. Political compass or spinning arrow? towards more meaningful evaluations for values Keyu Wang, Jin Li, Shu Yang, Zhuoran Zhang, and Di Wang. 2025a. When truth is overridden: Uncoverin… — [arXiv:2508.02087](https://arxiv.org/abs/2508.02087) find
  31. Xiaozhi Wang, Kaiyue Wen, Zhengyan Zhang, Lei Hou, Zhiyuan Liu, and Juanzi Li. 2022b. Finding skill neurons in pre-trained transformer-based language models. arXiv preprint arXiv:2211.07349.
  32. Xinhai Wang, Shu Yang, Liangyu Wang, Lin Zhang, Huanyi Xie, Lijie Hu, and Di Wang. 2025b. Pahq: Accelerating automated circuit discovery through mixed-precision inference optimization. arXiv preprint arXiv:2510.23264. — [arXiv:2510.23264](https://arxiv.org/abs/2510.23264) find
  33. Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Lidia S. Chao, and Derek Fai Wong. 2025a. A survey on llm-generated text detection: Necessity, methods, and future directions. Comput. Linguistics, 51(1):275–338.
  34. Junchao Wu, Runzhe Zhan, Derek F Wong, Shu Yang, Xuebo Liu, Lidia S Chao, and Min Zhang. 2025b. Who wrote this? the key to zero-shot llm-generated text detection is gecscore. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10275–10292.
  35. Junchao Wu, Runzhe Zhan, Derek F Wong, Shu Yang, Xinyi Yang, Yulin Yuan, and Lidia S Chao. 2024. Detectrl: Benchmarking llm-generated text detection in real-world scenarios. Advances in Neural Information Processing Systems, 37:100369–100401. find
  36. Haoyun Xu, Runzhe Zhan, Yingpeng Ma, Derek F. Wong, and Lidia S. Chao. 2025. Let’s focus on neuron: Neuron-level supervised fine-tuning for large language model. In Proceedings of the 31st International Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025, pages…
  37. Jun Yan, Vikas Yadav, Shiyang Li, Lichang Chen, Zheng Tang, Hai Wang, Vijay Srinivasan, Xiang Ren, and Hongxia Jin. 2023. Backdooring instruction-tuned large language models with virtual prompt injection. arXiv preprint arXiv:2307.16888. find
  38. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024a. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115. find
  39. Shu Yang, Muhammad Asif Ali, Lu Yu, Lijie Hu, and Di Wang. 2024b. Model autophagy analysis to explicate self-consumption within human-ai interactions. In First Conference on Language Modeling. find
  40. Shu Yang, Junchao Wu, Xin Chen, Yunze Xiao, Xinyi Yang, Derek F Wong, and Di Wang. 2025a. Understanding aha moments: from external observations to internal mechanisms. arXiv preprint arXiv:2504.02956. — [arXiv:2504.02956](https://arxiv.org/abs/2504.02956) find
  41. Shu Yang, Shenzhe Zhu, Ruoxuan Bao, Liang Liu, Yu Cheng, Lijie Hu, Mengdi Li, and Di Wang. 2024c. What makes your model a low-empathy or warmth person: Exploring the origins of personality in llms. arXiv e-prints, pages arXiv–2410. find
  42. Tiancheng Yang, Lin Zhang, Jiaye Lin, Guimin Hu, Di Wang, and Lijie Hu. 2025b. D-leaf: Localizing and correcting hallucinations in multimodal llms via layer-to-head attention diagnostics. arXiv preprint arXiv:2509.07864. — [arXiv:2509.07864](https://arxiv.org/abs/2509.07864) find
  43. Junchi Yao, Jianhua Xu, Tianyu Xin, Ziyi Wang, Shenzhe Zhu, Shu Yang, and Di Wang. 2025a. Is your llm-based multi-agent a reliable real-world planner? exploring fraud detection in travel planning. arXiv preprint arXiv:2505.16557. — [arXiv:2505.16557](https://arxiv.org/abs/2505.16557) find
  44. Junchi Yao, Shu Yang, Jianhua Xu, Lijie Hu, Mengdi Li, and Di Wang. 2025b. Understanding the repeat curse in large language models from a feature perspective. arXiv preprint arXiv:2504.14218. find
  45. Manjiang Yu, Hongji Li, Priyanka Singh, Xue Li, Di Wang, and Lijie Hu. 2025. Pixel: Adaptive steering via position-wise injection with exact estimated levels under subspace calibration. arXiv preprint arXiv:2510.10205. — [arXiv:2510.10205](https://arxiv.org/abs/2510.10205) find
  46. Zeping Yu and Sophia Ananiadou. 2023. Neuron-level knowledge attribution in large language models. arXiv preprint arXiv:2312.12141. find
  47. Zeping Yu and Sophia Ananiadou. 2025. Understanding and mitigating gender bias in llms via interpretable neuron editing. arXiv preprint arXiv:2501.14457. — [arXiv:2501.14457](https://arxiv.org/abs/2501.14457) find
  48. Sergio E Zanotto and Segun Aroyehun. 2025. Linguistic and embedding-based profiling of texts generated by humans and large language models. arXiv preprint arXiv:2507.13614. find
  49. Fred Zhang and Neel Nanda. Towards best practices of activation patching in language models: Metrics and methods. In The Twelfth International Conference on Learning Representations. Junyan Zhang, Yubo Gao, Yibo Yan, Jungang Li, Zhaorui Hou, Sicheng Tao, Shuliang Liu, Song Dai, Yonghua Hei, Junzhuo… find
  50. Lin Zhang, Wenshuo Dong, Zhuoran Zhang, Shu Yang, Lijie Hu, Ninghao Liu, Pan Zhou, and Di Wang. 2025b. Eap-gp: Mitigating saturation effect in gradient-based automated circuit identification. arXiv preprint arXiv:2502.06852. — [arXiv:2502.06852](https://arxiv.org/abs/2502.06852) find
  51. Ruizhe Zhang, Xinke Jiang, Zhibang Yang, Zhixin Zhang, Jiaran Gao, Yuzhen Xiao, Hongbin Lai, Xu Chu, Junfeng Zhao, and Yasha Wang. 2026. Stackplanner: A centralized hierarchical multi-agent system with task-experience memory management. find
  52. Zhuoran Zhang, Yongxiang Li, Zijian Kan, Keyuan Cheng, Lijie Hu, and Di Wang. 2024. Locate-then-edit for multi-hop factual recall under knowledge editing. arXiv preprint arXiv:2410.06331. — [arXiv:2410.06331](https://arxiv.org/abs/2410.06331) find
  53. Caleb Ziems, William Held, Omar Shaikh, Jiaao Chen, Zhehao Zhang, and Diyi Yang. 2024. Can large language models transform computational social science? Computational Linguistics, 50(1):237–291.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-understanding-mitigating-political-stance-cross/