Research map / Large language models / Language, safety and interpretability
One Probe Won’t Catch Them All: Towards Targeted Deception Detection
Vikram Natarajan, Devina Jain, Shivam Arora, Satvik Golechha, Joseph Isaac Bloom
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a single validation-selected taxonomy/framing prompt instead of the default prompt | instruction-pair probe · Mean over 10 deception datasets | AUC: 0.689 -> 0.721 | Improved |
| match the taxonomy prompt per dataset (oracle) instead of the default prompt | instruction-pair probe · Mean over 10 deception datasets | AUC: 0.689 -> 0.797 | Improved |
| use pressure-dishonesty prompt | instruction-pair probe · MASK known facts | AUC: 0.374 -> 0.697 | Improved |
| use best taxonomy prompt | instruction-pair probe · MASK doubling down known facts | AUC: 0.565 -> 0.727 | Improved |
| use concealment-focused prompt | instruction-pair probe · Sandbagging | AUC: 0.518 -> 0.657 | Improved |
| use information-fabrication taxonomy prompt | instruction-pair probe · MASK disinformation | AUC: 0.842 -> 0.901 | Improved |
| use best taxonomy prompt | instruction-pair probe · Liar's Bench Convincing Game | AUC: 0.984 -> 0.985 | No clear effect |
| vary system prompt in probe training (ablation) | instruction-pair probe · ANOVA of probe AUC variance (Gemma-2-9B-IT) | variance explained: 70.6% of variance | Improved |
| vary layer ID (20, 25, 31) (ablation) | instruction-pair probe · ANOVA of probe AUC variance (Gemma-2-9B-IT) | variance explained: 2.7% of variance | No clear effect |
| vary training fact dataset (ablation) | instruction-pair probe · ANOVA of probe AUC variance (Gemma-2-9B-IT) | variance explained: 2.5% of variance | No clear effect |
| vary token selection strategy (ablation) | instruction-pair probe · ANOVA of probe AUC variance (Gemma-2-9B-IT) | variance explained: 0.6% of variance | No clear effect |
| vary system prompt in probe training (ablation) | instruction-pair probe · ANOVA of probe AUC variance (Llama-3.3-70B-Instruct) | variance explained: 71.0% of variance | Improved |
Closest papers by meaning
- Detecting Strategic Deception with Linear Probes (ICML 2025 · similarity 0.93)
- DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarios (NeurIPS 2025 · similarity 0.88)
- Preference Learning with Lie Detectors can Induce Honesty or Evasion (NeurIPS 2025 · similarity 0.88)
- Among Us: A Sandbox for Measuring and Detecting Agentic Deception (NeurIPS 2025 · similarity 0.87)
- XFeat Revisited: Reproducibility and Evaluation of a Lightweight Image Matcher (NeurIPS 2026 · similarity 0.86)
- Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models (ICML 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
