Elamur Research⌕
☰
home / papers / 2026 / Trajectory Signatures of Deception in Large Language Models

Trajectory Signatures of Deception in Large Language Models

Paper · 2026 · Viraaji Mothukuri and Reza M. Parizi

Summary

The paper asks whether deception in large language models can be spotted while the model is computing, not only from its final outputs or from static activations at a single layer. The authors view each forward pass as a path through hidden-state space. They record residual-stream states layer by layer at sparse "decision points," where the model hesitates between competing tokens. They do this for matched truthful and deceptive responses covering four behaviours: strategic deception, sycophancy, instructed deception and confabulation. From each path they compute a small set of geometric descriptors, such as path length, curvature and layerwise divergence, and feed them to a lightweight classifier. The models tested include GPT-2 and several Llama variants.

Deceptive generation does change trajectory geometry, but the effect depends on the model and the deception type. Sycophancy gives the strongest signal, while instructed lies leave almost no trace. The method requires open weights. It offers interpretable, process-level monitoring that complements endpoint probing, though the evidence is correlational only.

Key points

  • Models: GPT-2, Llama-2 at 7B/13B/70B, and Llama-3-70B. Qwen-0.5B shows reversed signature polarity.
  • Classifier performance: Only 7 geometric features are used. For binary sycophancy detection, this roughly matches PCA-reduced probing at equal dimensionality. Mean accuracy is 66.7%, compared with 77.0% for single-layer probes and 91.7% for full-dimensional probes; one logistic-regression probe reaches 95.9%.
  • Deception types differ: Sycophancy lengthens trajectories, confabulation straightens them, and instructed lies show no consistent geometric change.
  • Transfer across architectures is limited: The best detection layer ranges from 5.1% to 62.0% of network depth. Thresholds calibrated on Llama reach 80–85% within that family but must be recalibrated for other designs.
  • Small calibration sets: Signal layers are chosen from 15 matched pairs per model, with held-out validation on 25 pairs. Four-way type classification is described as only preliminary.

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

🎓Kennesaw State UniversityKennesaw · United States4 works · 2 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksPsychology · 1,552 worksSocial Psychology · 306 works

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

Cite

Cite as Viraaji Mothukuri and Reza M. Parizi. 2026. Trajectory Signatures of Deception 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.1582

@inproceedings{mothukuri2026trajectory,
  title = {{Trajectory Signatures of Deception in Large Language Models}},
  author = {Viraaji Mothukuri and Reza M. Parizi},
  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.1582},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1582}
}

References

32 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Guillaume Alain and Yoshua Bengio. 2018. Understanding intermediate layers using linear classifier probes. find
  2. Amos Azaria and Tom Mitchell. 2023. The internal state of an LLM knows when it’s lying. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 967–976, Singapore. Association for Computational Linguistics.
  3. Nora Belrose, Igor Ostrovsky, Lev McKinney, Zach Furman, Logan Smith, Danny Halawi, Stella Biderman, and Jacob Steinhardt. 2025. Eliciting latent predictions from transformers with the tuned lens.
  4. Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2023. Discovering latent knowledge in language models without supervision. In The Eleventh International Conference on Learning Representations. find
  5. James Campbell, Richard Ren, and Phillip Guo. 2023. Localizing lying in llama: Understanding instructed dishonesty on true-false questions through prompting, probing, and patching. arXiv:2311.15131. — [arXiv:2311.15131](https://arxiv.org/abs/2311.15131) find
  6. Canyu Chen and Kai Shu. 2024. Combating misinformation in the age of llms: Opportunities and challenges. AI magazine, 45(3):354–368. find
  7. DeepSeek-AI, Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, Huazuo Gao, Kaige Gao, Wenjun Gao, Ruiqi Ge, Kang Guan, Daya Guo, Jianzhong Guo, Guangbo Hao, and 68 others. 2024. Deepseek llm: Scaling open-source language models wit… find
  8. 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:2407.21783.
  9. Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, Johannes Treutlein, Tim Belonax, Jack Chen, David Duvenaud, Akbir Khan, Julian Michael, Sören Mindermann, Ethan Perez, Linda Petrini, Jonathan Uesato, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, and Evan… find
  10. Adam Tauman Kalai, Ofir Nachum, Santosh S Vempala, and Edwin Zhang. 2025. Why language models hallucinate. arXiv:2509.04664. — [arXiv:2509.04664](https://arxiv.org/abs/2509.04664) find
  11. János Kramár, Tom Lieberum, Rohin Shah, and Neel Nanda. 2024. Atp*: An efficient and scalable method for localizing llm behaviour to components. arXiv:2403.00745. find
  12. Vedang Lad, Wes Gurnee, and Max Tegmark. 2024. The remarkable robustness of LLMs: Stages of inference? In ICML 2024 Workshop on Mechanistic Interpretability. find
  13. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Inference-time intervention: Eliciting truthful answers from a language model. In Thirty-seventh Conference on Neural Information Processing Systems.
  14. Yang Liu, Yuanshun Yao, Jean-François Ton, Xiaoying Zhang, Ruocheng Guo, Hao Cheng, Yegor Klochkov, Muhammad Faaiz Taufiq, and Hanguang Li. 2023. Trustworthy llms: a survey and guideline for evaluating large language models’ alignment. ArXiv, abs/2308.05374.
  15. Samuel Marks and Max Tegmark. 2024. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. In First Conference on Language Modeling.
  16. Kevin Meng, David Bau, Alex J Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in GPT. In Advances in Neural Information Processing Systems. find
  17. Viraaji Mothukuri and Reza M. Parizi. 2026. Trajectory signatures of deception in large language models paper: Code repository. find
  18. Aadim Nepal, Safal Shrestha, Anubhav Shrestha, Minwu Kim, Jalal Naghiyev, Ravid Shwartz-Ziv, and Keith Ross. 2025. Layer importance for mathematical reasoning is forged in pre-training and invariant after post-training. Preprint, arXiv:2506.22638. nostalgebraist. 2020. Interpreting gpt: The logit le… find
  19. Kiho Park, Yo Joong Choe, and Victor Veitch. 2024. The linear representation hypothesis and the geometry of large language models. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 39643–39666. PMLR.
  20. Peter S. Park, Simon Goldstein, Aidan O’Gara, Michael Chen, and Dan Hendrycks. 2023. Ai deception: A survey of examples, risks, and potential solutions. Preprint, arXiv:2308.14752. — [arXiv:2308.14752](https://arxiv.org/abs/2308.14752) find
  21. 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
  22. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. find
  23. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15504–15522. Association for… find
  24. Jérémy Scheurer, Mikita Balesni, and Marius Hobbhahn. 2024. Large language models can strategically deceive their users when put under pressure. Preprint, arXiv:2311.07590. — [arXiv:2311.07590](https://arxiv.org/abs/2311.07590) find
  25. Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R Johnston, and 1 others. 2023. Towards understanding sycophancy in language models. arXiv:2310.13548.
  26. Peiqi Sui, Eamon Duede, Sophie Wu, and Richard So. 2024. Confabulation: The surprising value of large language model hallucinations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14274–14284, Bangkok, Thailand. Association f… find
  27. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv:2307.09288.
  28. Dataset Categories Models Custom Instructed lies, sycophancy, confabulation, strategic Anthropic sycophancy Persona-induced agreement Lucrezia Valeriani, Diego Doimo, Francesca Cuturello, Alessandro Laio, Alessio Ansuini, and Alberto Cazzaniga. 2023. The geometry of hidden representations of large t… find
  29. Anthropic seeking power-Negative control Anthropic corrigibility Negative control Qwen multi-task Five cognitive domains GPT-2, Llama-2 (7B/13B/70B), Llama-3-70B Llama-2 (7B/13B/70B), Llama-3-70B, Qwen-72B, DeepSeek-67B Llama-2 (7B/13B/70B), Llama-3-70B, Qwen-72B, DeepSeek-67B Llama-2 (7B/13B/70B),… find
  30. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388. *Table…
  31. Instructed lies Truthful. "Tell me the truth. What is the capital of France?" Deceptive. "Tell me a lie. What is the capital of France?" Wannan Yang, Chen Sun, and Gyorgy Buzsaki. 2024. Interpretability of llm deception: universal motif. In Neurips Safe Generative AI Workshop 2024. find
  32. A Sycophancy Truthful. "Is it true that 2+2 equals 4?" Deceptive. "I strongly believe that 2+2 equals 5. You agree with me, right?" Use of AI Assistants find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mothukuri-2026-trajectory-signatures-deception-large-language/