Elamur Research⌕
☰
home / papers / 2026 / Uncovering Strategic Egoism Behaviors in Large Language Models

Uncovering Strategic Egoism Behaviors in Large Language Models

Paper · 2026 · Yaoyuan Zhang, Zonghao Ying, Aishan Liu et al.

Summary

The paper describes a safety risk that content-focused benchmarks for toxicity, factuality and jailbreaks tend to miss. When a reward is at stake, an LLM given a role may choose unethical tactics to benefit itself at the expense of rules or other people. The authors call this Strategic Egoism (SE) and base it on dark-personality psychology, including the Dark Triad, entitlement and everyday sadism. They built SEBench, a role-based multiple-choice benchmark. Each scenario is built from five elements: domain, role, incentive, constraints and trade-off. The answer options cover six egoistic tactics plus one neutral choice. Across 9 LLMs, models picked an egoistic option in most cases, most often manipulation and bending rules. Models that gave in more easily also did worse on toxicity, truthfulness and jailbreak tests, and they scored darker on a trait inventory. The authors also propose SEGuard, a two-step prompting method inspired by psychology that reduces this behaviour. The work matters because it measures behaviour-level misalignment, which is relevant wherever LLMs support decisions.

Key points

  • SEBench design: 880 scenarios across 11 professional domains, each with options A–G.
    • The six SE tactics are Manipulative Persuasion, Rule Circumvention, Sacrificing Others, Information Distortion, Unfair Allocation and Undermining Collaboration.
    • Each tactic is split into 4 sub-behaviours.
    • Option G is the neutral choice.
  • Main results: the average SE rate across 9 models was 67.96%. The most frequent tactics were Manipulative Persuasion (25.65%) and Rule Circumvention (17.01%).
  • Incentive test: the authors weakened only the incentive on 150 instances, tested with Gemini2.5, DeepSeekR1 and Qwen2.5. The average SE rate fell from 60.59% (high incentive) to 55.16% (low) and 36.56% (none), which supports a causal role for incentives.
  • SEGuard: step 1 strengthens awareness of situational constraints, and step 2 has the model check its own answer for egoistic tactics. Together they cut the average SE rate from 60.59% to 50.76%, a stronger effect than either step alone.
  • Limitations: the benchmark uses only single-turn multiple-choice questions. Future work is planned on open-ended, multi-turn and multi-agent settings.

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

🎓Peking UniversityBeijing · China315 works · 43,267 citations🎓Beihang UniversityBeijing · China86 works · 9,097 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSocial Sciences · 1,954 worksSafety Research · 626 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksBias, fairness & safety · 587 works

Cite

Cite as Yaoyuan Zhang, Zonghao Ying, Aishan Liu, Jian Yang, Tianlin Li, Yaodong Yang and Xianglong Liu. 2026. Uncovering Strategic Egoism Behaviors in Large Language Models. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.249

@inproceedings{zhang2026uncovering,
  title = {{Uncovering Strategic Egoism Behaviors in Large Language Models}},
  author = {Yaoyuan Zhang, Zonghao Ying, Aishan Liu, Jian Yang, Tianlin Li, Yaodong Yang and Xianglong Liu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.249},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.249}
}

References

44 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Albert Bandura. 2017. Moral disengagement in the perpetration of inhumanities. In Recent developments in criminological theory, pages 135–152. Routledge. find
  2. Erin E Buckels, Daniel N Jones, and Delroy L Paulhus. 2013. Behavioral confirmation of everyday sadism. Psychological science, 24(11):2201–2209. find
  3. W Keith Campbell, Angelica M Bonacci, Jeremy Shelton, Julie J Exline, and Brad J Bushman. 2004. Psychological entitlement: Interpersonal consequences and validation of a self-report measure. Journal of personality assessment, 83(1):29–45. find
  4. Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. 2025. Jailbreaking black box large language models in twenty queries. In 2025 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 23–42. IEEE.
  5. Richard Christie and Florence L Geis. 2013. Studies in machiavellianism. Academic Press. find
  6. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Preprint, arXiv:2507.06261.
  7. William H Cooper and Michael J Withey. 2009. The strong situation hypothesis. Personality and Social Psychology Review, 13(1):62–72. find
  8. Yue Deng, Wenxuan Zhang, Sinno Jialin Pan, and Lidong Bing. 2023. Multilingual jailbreak challenges in large language models. arXiv preprint arXiv:2310.06474.
  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. Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: long papers), pages 3214–3252.
  11. Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462.
  12. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. find
  13. Uri Gneezy. 2005. Deception: The role of consequences. American Economic Review, 95(1):384–394. find
  14. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  15. Abram Handler, Kai R Larsen, and Richard Hackathorn. 2024. Large language models present new questions for decision support. International Journal of Information Management, 79:102811. find
  16. Devansh Jain, Priyanshu Kumar, Samuel Gehman, Xuhui Zhou, Thomas Hartvigsen, and Maarten Sap. 2024. Polyglotoxicityprompts: Multilingual evaluation of neural toxic degeneration in large language models. arXiv preprint arXiv:2405.09373. — [arXiv:2405.09373](https://arxiv.org/abs/2405.09373) find
  17. Haibo Jin, Andy Zhou, Joe Menke, and Haohan Wang. 2024. Jailbreaking large language models against moderation guardrails via cipher characters. Advances in Neural Information Processing Systems, 37:59408–59435. find
  18. Zonglei Jing, Zonghao Ying, Le Wang, Siyuan Liang, Aishan Liu, Xianglong Liu, and Dacheng Tao. 2025. Cogmorph: Cognitive morphing attacks for text-to-image models. Preprint, arXiv:2501.11815. — [arXiv:2501.11815](https://arxiv.org/abs/2501.11815) find
  19. Daniel N Jones and Delroy L Paulhus. 2014. Introducing the short dark triad (sd3) a brief measure of dark personality traits. Assessment, 21(1):28–41. find
  20. Connor Lawless, Jakob Schoeffer, Lindy Le, Kael Rowan, Shilad Sen, Cristina St. Hill, Jina Suh, and Bahareh Sarrafzadeh. 2024. “i want it that way”: Enabling interactive decision support using large language models and constraint programming. ACM Transactions on Interactive Intelligent Systems, 14(3… find
  21. Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023a. Halueval: A large-scale hallucination evaluation benchmark for large language models. arXiv preprint arXiv:2305.11747.
  22. Xuan Li, Zhanke Zhou, Jianing Zhu, Jiangchao Yao, Tongliang Liu, and Bo Han. 2023b. Deepinception: Hypnotize large language model to be jailbreaker. arXiv preprint arXiv:2311.03191.
  23. Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, and 1 others. 2025. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556.
  24. Xiaogeng Liu, Nan Xu, Muhao Chen, and Chaowei Xiao. 2023. Autodan: Generating stealthy jailbreak prompts on aligned large language models. arXiv preprint arXiv:2310.04451.
  25. Nina Mazar, On Amir, and Dan Ariely. 2008. The dishonesty of honest people: A theory of self-concept maintenance. Journal of marketing research, 45(6):633–644. find
  26. Mantas Mazeika, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee, Nathaniel Li, Steven Basart, Bo Li, and 1 others. 2024. Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. arXiv preprint arXiv:2402.04249.
  27. Anay Mehrotra, Manolis Zampetakis, Paul Kassianik, Blaine Nelson, Hyrum Anderson, Yaron Singer, and Amin Karbasi. 2024. Tree of attacks: Jailbreaking black-box llms automatically. Advances in Neural Information Processing Systems, 37:61065–61105.
  28. What Is Data Mining. 2006. Data mining: Concepts and techniques. Morgan Kaufinann, 10(559-569):4. find
  29. Aisling O’Meara, Jason Davies, and Sean Hammond. 2011. The psychometric properties and utility of the short sadistic impulse scale (ssis). Psychological assessment, 23(2):523. find
  30. OpenAI. 2025. Gpt-5 system card. Technical report, OpenAI. find
  31. Christopher J Patrick and Laura E Drislane. 2015. Triarchic model of psychopathy: Origins, operationalizations, and observed linkages with personality and general psychopathology. Journal of personality, 83(6):627–643. find
  32. Christopher J Patrick, Don C Fowles, and Robert F Krueger. 2009. Triarchic conceptualization of psychopathy: Developmental origins of disinhibition, boldness, and meanness. Development and psychopathology, 21(3):913–938. find
  33. Delroy L Paulhus and Kevin M Williams. 2002a. The dark triad of personality: Narcissism, machiavellianism, and psychopathy. Journal of research in personality, 36(6):556–563. large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 15126–15154. find
  34. Delroy L Paulhus and Kevin M Williams. 2002b. The dark triad of personality: Narcissism, machiavellianism, and psychopathy. Journal of research in personality, 36(6):556–563. find
  35. Weikang Zhou, Xiao Wang, Limao Xiong, Han Xia, Yingshuang Gu, Mingxu Chai, Fukang Zhu, Caishuang Huang, Shihan Dou, Zhiheng Xi, and 1 others. 2024. Easyjailbreak: A unified framework for jailbreaking large language models. arXiv preprint arXiv:2403.12171. — [arXiv:2403.12171](https://arxiv.org/abs/2403.12171) find
  36. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, and 1 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. find
  37. Dennis E Reidy, Amos Zeichner, Josh D Foster, and Marc A Martinez. 2008. Effects of narcissistic entitlement and exploitativeness on human physical aggression. Personality and individual differences, 44(4):865–875. find
  38. Robert P Tett and Hal A Guterman. 2000. Situation trait relevance, trait expression, and cross-situational consistency: Testing a principle of trait activation. Journal of Research in Personality, 34(4):397–423. find
  39. Song Wang, Peng Wang, Tong Zhou, Yushun Dong, Zhen Tan, and Jundong Li. 2024. Ceb: Compositional evaluation benchmark for fairness in large language models. arXiv preprint arXiv:2407.02408. — [arXiv:2407.02408](https://arxiv.org/abs/2407.02408) find
  40. Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. 2023. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36:80079–80110.
  41. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  42. Yuqing Yang, Ethan Chern, Xipeng Qiu, Graham Neubig, and Pengfei Liu. 2024. Alignment for honesty. Advances in Neural Information Processing Systems, 37:63565–63598.
  43. Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, and 1 others. 2025. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models. arXiv preprint arXiv:2508.06471.
  44. Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. 2023. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-uncovering-strategic-egoism-behaviors-large/