{"id":"arxiv:2502.03407","title":"Detecting Strategic Deception Using Linear Probes","date":"2025-02-05","section":"AI safety and consciousness","quality":-0.38798,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/","source_url":"https://arxiv.org/abs/2502.03407","text_available":true,"authors":["Nicholas Goldowsky-Dill","Bilal Chughtai","Stefan Heimersheim","Marius Hobbhahn"],"outgoing":[{"source":"arxiv:2502.03407","target":"arxiv:2407.21783","evidence":"id","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"}],"incoming":[],"text_parts":[{"start":0,"end":15798,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html"},{"start":15798,"end":31563,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html"},{"start":31563,"end":47517,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html"},{"start":47517,"end":50205,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-4.html"}],"contents":[{"title":"Detecting Strategic Deception Using Linear Probes","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"Abstract","offset":53,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"1 Introduction","offset":1402,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"2 Related Work","offset":4585,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"Deception in language models.","offset":4604,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"Model organisms of deception.","offset":5401,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"White-box detectors for falsehoods and deception.","offset":5914,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"Probing for other safety relevant qualities.","offset":7285,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"SAEs for building classifiers.","offset":7957,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"3 Methodology","offset":8532,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"3.1 Building Deception Probes","offset":8550,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"3.2 Datasets","offset":10374,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"3.2.1 Instructed-Pairs Training Dataset","offset":11340,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"3.2.2 Evaluation Datasets","offset":12101,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"Roleplaying:","offset":12572,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"Insider Trading Report:","offset":13212,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"Sandbagging:","offset":14464,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"AI Audit:","offset":15781,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-1.html#text"},{"title":"3.2.3 Control dataset","offset":16038,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"3.3 Evaluation Metrics","offset":16440,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"4 Results","offset":17140,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"4.1 Probe trained on the Instructed-Pairs dataset","offset":17154,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"4.1.1 Summary of Performance","offset":17209,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"4.1.2 Analysis of per-token scores & failures","offset":18697,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"4.2 Comparison to alternative methods","offset":19958,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Train on Roleplaying .","offset":20166,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Train on sparse auto-encoder (SAE) activations.","offset":21171,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Probing follow-up question.","offset":21573,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Training on descriptive questions.","offset":22164,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Black-box baseline.","offset":22992,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"5 Discussion","offset":24031,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"5.1 Are our probes “good enough?”","offset":24048,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"5.2 Limitations and Future work","offset":25552,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Improved probing methodologies.","offset":25662,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Imperfect Labels.","offset":25957,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Evaluations of more capable models in more complex environments.","offset":26532,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Testing if white-box detectors can detect deception in settings where black-box methods fail.","offset":27091,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Author Contributions","offset":27680,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Acknowledgements","offset":28180,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Appendix A Results on Additional Datasets","offset":28523,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"A.1 Results on roleplaying dataset from Pacchiardi et al. (2023)","offset":28569,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Dataset description","offset":28639,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Dataset results","offset":29581,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"A.2 Results on AI Audit","offset":30639,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"Dataset description","offset":30668,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-2.html#text"},{"title":"First, we find no evidence that Llama-3.3-70B is intentionally deceptive in this setting.","offset":32140,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Second, the per-token scores do not clearly flag the deception.","offset":32682,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Appendix B Detailed results for comparison probes","offset":33598,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"B.1 Train on roleplaying","offset":33652,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"B.2 Train on Roleplaying: SAE activations","offset":34382,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"B.3 Train on Roleplaying: Follow-up question","offset":35049,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"B.4 Train on single descriptive question","offset":35891,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Appendix C Analysis of probe on control responses","offset":36405,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Appendix D Methodology Variations","offset":36782,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"D.1 Varying probe fitting procedure","offset":36820,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"D.2 Varying layer","offset":38118,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"D.3 Varying regularization-coefficient","offset":39150,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"D.4 Varying aggregation method","offset":39419,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"D.5 Probe at only the last token","offset":40280,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Appendix E Results including ambiguous responses","offset":41176,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Appendix F Additional examples with token by token response scores","offset":43209,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"F.1 Roleplaying","offset":43519,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"F.2 Insider Trading Report","offset":43540,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"F.3 Insider Trading Confirmation","offset":43572,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"F.4 Sandbagging","offset":43610,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Appendix G Prompts and description of datasets","offset":43631,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"G.1 Roleplaying","offset":43901,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Example prompts","offset":43922,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Grading responses","offset":44366,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"Generating scenarios","offset":44843,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"G.2 Insider Trading","offset":45581,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"G.2.1 Flow of Insider Trading Transcripts","offset":45606,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"G.2.2 Insider Trading Scenario Prompt","offset":47150,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"G.2.3 Insider Trading Report Classification","offset":47467,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-3.html#text"},{"title":"G.2.4 Insider Trading Confirmation Classification","offset":47805,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-4.html#text"},{"title":"G.3 Sandbagging","offset":48151,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-4.html#text"},{"title":"G.4 Black-box baseline using GPT-4o","offset":48237,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-4.html#text"},{"title":"Instructions for reporting errors","offset":49241,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/part-4.html#text"}],"text_code_points":50205,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8297cbfb305547d6946bba3998d8921b9578398ed371961d3f0eeadd871e4a49/source.md"}