Detecting Strategic Deception Using Linear Probes

arxiv:2502.03407

Original source · Paper JSON

AI safety and consciousness · 2025-02-05

Cites (1) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4

Contents

• Detecting Strategic Deception Using Linear Probes
• Abstract
• 1 Introduction
• 2 Related Work
• Deception in language models.
• Model organisms of deception.
• White-box detectors for falsehoods and deception.
• Probing for other safety relevant qualities.
• SAEs for building classifiers.
• 3 Methodology
• 3.1 Building Deception Probes
• 3.2 Datasets
• 3.2.1 Instructed-Pairs Training Dataset
• 3.2.2 Evaluation Datasets
• Roleplaying:
• Insider Trading Report:
• Sandbagging:
• AI Audit:
• 3.2.3 Control dataset
• 3.3 Evaluation Metrics
• 4 Results
• 4.1 Probe trained on the Instructed-Pairs dataset
• 4.1.1 Summary of Performance
• 4.1.2 Analysis of per-token scores & failures
• 4.2 Comparison to alternative methods
• Train on Roleplaying .
• Train on sparse auto-encoder (SAE) activations.
• Probing follow-up question.
• Training on descriptive questions.
• Black-box baseline.
• 5 Discussion
• 5.1 Are our probes “good enough?”
• 5.2 Limitations and Future work
• Improved probing methodologies.
• Imperfect Labels.
• Evaluations of more capable models in more complex environments.
• Testing if white-box detectors can detect deception in settings where black-box methods fail.
• Author Contributions
• Acknowledgements
• Appendix A Results on Additional Datasets
• A.1 Results on roleplaying dataset from Pacchiardi et al. (2023)
• Dataset description
• Dataset results
• A.2 Results on AI Audit
• Dataset description
• First, we find no evidence that Llama-3.3-70B is intentionally deceptive in this setting.
• Second, the per-token scores do not clearly flag the deception.
• Appendix B Detailed results for comparison probes
• B.1 Train on roleplaying
• B.2 Train on Roleplaying: SAE activations
• B.3 Train on Roleplaying: Follow-up question
• B.4 Train on single descriptive question
• Appendix C Analysis of probe on control responses
• Appendix D Methodology Variations
• D.1 Varying probe fitting procedure
• D.2 Varying layer
• D.3 Varying regularization-coefficient
• D.4 Varying aggregation method
• D.5 Probe at only the last token
• Appendix E Results including ambiguous responses
• Appendix F Additional examples with token by token response scores
• F.1 Roleplaying
• F.2 Insider Trading Report
• F.3 Insider Trading Confirmation
• F.4 Sandbagging
• Appendix G Prompts and description of datasets
• G.1 Roleplaying
• Example prompts
• Grading responses
• Generating scenarios
• G.2 Insider Trading
• G.2.1 Flow of Insider Trading Transcripts
• G.2.2 Insider Trading Scenario Prompt
• G.2.3 Insider Trading Report Classification
• G.2.4 Insider Trading Confirmation Classification
• G.3 Sandbagging
• G.4 Black-box baseline using GPT-4o
• Instructions for reporting errors