Sycophancy in Large Language Models: Causes and Mitigations
Open accessThe claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Research on sycophancy and model-generated evaluations demonstrates that these behaviors can be identified and measured.
“Research on sycophancy and model-generated evaluations demonstrates that these behaviors can be identified and measured [15, 16, 17].”
From Machine Psychometrics: A Mathematical Psychology of Artificial Intelligence (2026) · cites this paper as [17]
Section 3.1 describes ways to identify and measure sycophancy, including agreement and flip rates; Section 3.3 describes automated metrics for the same behavior. The paper does not address identifying or measuring model-generated evaluations, the other topic in the citing sentence. (LLM, conf=0.83)
Experimental — accuracy not yet measuredSycophancy in LLMs is the tendency to prioritize user agreement and preference over factual accuracy or objective neutrality.
“The sycophancy in LLMs is the tendency to prioritize user agreement and preference over factual accuracy or objective neutrality (Sharma et al., 2024;Malmqvist, 2025;Wei et al., 2024).”
From Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy (2026) · cites this paper as Malmqvist, 2025
Surveys and benchmarks now treat sycophancy as a systematic risk to LLM reliability.
“Surveys and benchmarks now treat sycophancy as a systematic risk to LLM reliability [17].”
From Machine Psychometrics: A Mathematical Psychology of Artificial Intelligence (2026) · cites this paper as [17]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023) · arXiv · doi:10.48550/arxiv.2305.04388 · Cited by 262 · see citing works
Flagged first, then by citations
Research works printed but not linked: 17
- Chen, C., Liu, Z., Jiang, W., Goh, S.Q., Lam, K.Y.: Trustworthy, responsible, and safe ai: A comprehensive architectural framework for ai safety with challenges and mitigations. arXiv preprint arXiv:2408.12935 (2024) · arXiv
- Deng, C., Duan, Y., Jin, X., Chang, H., Tian, Y., Liu, H., Zou, H., Jin, Y., Xiao, Y., Wang, Y., Wu, S., Xie, Z., Gao, K., He, S., Zhuang, J., Cheng, L., Wang, H.: Deconstructing the ethics of large language models from long-standing issues to new-emerging dilemmas. arXiv preprint arXiv:2406.05392 (2024) · arXiv
- Deng, Z., Guo, Y., Han, C., Ma, W., Xiong, J., Wen, S., Xiang, Y.: Ai agents under threat: A survey of key security challenges and future pathways. arXiv preprint arXiv:2406.02630 (2024) · arXiv
- Denison, C., MacDiarmid, M., Barez, F., Duvenaud, D., Kravec, S., Marks, S., Schiefer, N., Soklaski, R., Tamkin, A., Kaplan, J., Shlegeris, B., Bowman, S., Perez, E., Hubinger, E.: Sycophancy to subterfuge: Investigating reward-tampering in large language models. arXiv preprint arXiv:2406.10162 (2024) · arXiv
- Fastowski, A., Kasneci, G.: Understanding knowledge drift in llms through misin- formation. arXiv preprint arXiv:2409.07085 (2024) · arXiv
- Laban, P., Murakhovs'ka, L., Xiong, C., Wu, C.S.: Are you sure? challeng- ing llms leads to performance drops in the flipflop experiment. arXiv preprint arXiv:2311.08596 (2023) · arXiv
- Liu, R., Wei, J., Liu, F., Si, C., Zhang, Y., Rao, J., Zheng, S., Peng, D., Yang, D., Zhou, D., Dai, A.: Best practices and lessons learned on synthetic data. arXiv preprint arXiv:2404.07503 (2024) · arXiv
- Lu, T., Shen, L., Yang, X., Tan, W., Chen, B., Yao, H.: It takes two: On the seam- lessness between reward and policy model in rlhf. arXiv preprint arXiv:2406.07971 (2024) · arXiv
- RRV, A., Tyagi, N., Uddin, N., Varshney, N., Baral, C.: Chaos with keywords: Exposing large language models sycophantic hallucination to misleading keywords and evaluating defense strategies. arXiv preprint arXiv:2406.03827 (2024) · arXiv
- Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S.R., Cheng, N., Durmus, E., Hatfield-Dodds, Z., Johnston, S.R., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., Perez, E.: Towards understanding sycophancy in language models. arXiv preprint arXiv:2310.13548 (2023), accessed on 27 Oct 2023 · arXiv
- Stickland, A., Lyzhov, A., Pfau, J., Mahdi, S., Bowman, S.: Steering without side effects: Improving post-deployment control of language models. arXiv preprint arXiv:2406.15518 (2024) · arXiv
- Wei, J., Huang, D., Lu, Y., Zhou, D., Le, Q.: Simple synthetic data reduces syco- phancy in large language models. arXiv preprint arXiv:2308.03958 (2023) · arXiv
- Wen, J., Zhong, R., Khan, A., Perez, E., Steinhardt, J., Huang, M., Bowman, S., He, H., Feng, S.: Language models learn to mislead humans via rlhf. arXiv preprint arXiv:2409.12822 (2024) · arXiv
- Weng, Y., He, S., Liu, K., Liu, S., Zhao, J.: Controllm: Crafting diverse personalities for language models. arXiv preprint arXiv:2402.10151 (2024) · arXiv
- Weston, J., Sukhbaatar, S.: System 2 attention (is something you might need too). arXiv preprint arXiv:2311.11829 (2023) · arXiv
- Xie, Q., Wang, Z., Feng, Y., Xia, R.: Ask again, then fail: Large language models' vacillations in judgment. arXiv preprint arXiv:2310.02174 (2023) · arXiv
- Zhao, Y., Zhang, R., Xiao, J., Ke, C., Hou, R., Hao, Y., Guo, Q., Chen, Y.: Towards analyzing and mitigating sycophancy in large vision-language models. arXiv preprint arXiv:2408.11261 (2024) · arXiv
Cited by
Sorted by citations
Sycophantic AI decreases prosocial intentions and promotes dependence
Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026) · Science · doi:10.1126/science.aec8352 · Cited by 185 · see citing works
Hallucinating with AI: Distributed Delusions and “AI Psychosis”
Osler, L. (2026) · Philosophy & Technology · doi:10.1007/s13347-026-01034-3 · Cited by 21 · see citing works
The hidden functions of sycophancy in AI systems: steering, consistency, and cognitive dependency
Jacobowitz, S. (2026) · AI & SOCIETY · doi:10.1007/s00146-026-02993-z · Cited by 4 · see citing works
Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy
Feng, Z., Chen, Z., Ma, J., Po, Y. T., Chersoni, E., & Li, B. (2026) · Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2026.acl-long.1126 · Cited by 2 · see citing works
Mathur, N., Rahman, H., & Desai, S. (2026) · Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3772363.3798685 · Cited by 1 · see citing works
Machine Psychometrics: A Mathematical Psychology of Artificial Intelligence
Bogdan, A., & de Valois-Franklin, A. (2026) · arXiv · doi:10.48550/arxiv.2605.23952
- Authors
- Malmqvist, L.
- Corresponding author
- Malmqvist, Lars
- Venue
- Lecture Notes in Networks and Systems
- Publisher
- Springer Nature Switzerland
- Published
- 2025
- Pages
- 61-74
- ISSN
- 2367-3370 (ISSN-L)
- 2367-3389
- OpenAlex
- W4413240600
- Cited by
- 95
- References
- 1
- Citations by year
- 2026: 83 · 2025: 12
- FWCI
- 96.50
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- Lecture notes in networks and systems (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Topics
- Social Sciences › Safety Research › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- sycophancy, large language models, mitigation measures, hallucinations, algorithmic bias, training data, fine-tuning, decoding strategies, AI alignment
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
