Research map / Large language models / Language, safety and interpretability
Mechanisms of Introspective Awareness
Uzay Macar, Li Yang, Zifan Wang, Peter Wallich, Emmanuel Ameisen, Jack Lindsey
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 |
|---|---|---|---|
| ablate the refusal direction (ablation) | Gemma3-27B instruct · concept injection, L=37, alpha=2.0 | detection rate (TPR): 10.8% -> 59.2% | Improved |
| ablate the refusal direction (ablation) | Gemma3-27B instruct · concept injection, no injection control | false positive rate: 0% -> 7.1% | Hurt |
| use post-trained (instruct) model instead of base model (ablation) | Gemma3-27B, OLMo-3.1-32B · concept injection | false positive rate | Improved |
| add trained steering vector at layer 29 | Gemma3-27B · 100 held-out concepts | detection rate: +74.66% | Improved |
| add trained steering vector at layer 29 | Gemma3-27B · 100 held-out concepts | forced identification: +21.90% | Improved |
| swap projection on mean-difference direction between success and failure concepts (ablation) | success concepts · concept swap | detection rate: 65% -> 36% | Hurt |
| swap projection or residual to success-like for failure concepts (ablation) | failure concepts · concept swap | detection rate: 11% -> 34% | Improved |
| patch L45 or L47 MLP with non-injected activations (ablation) | Gemma3-27B · 10 random concepts | introspection rate: to ~27% and ~26% | Hurt |
| ablate top-ranked gate features (ablation) | Gemma3-27B steered examples · concept injection | detection rate: 39% -> 8% | Hurt |
| ablate gates with forced identification (ablation) | Gemma3-27B · concept injection | identification accuracy: 60% -> 46% | Hurt |
| ablate or patch top weak evidence-carrier features (ablation) | Gemma3-27B · concept injection | detection rate | No clear effect |
| intervene on individual attention heads (ablation) | Gemma3-27B · concept injection | detection rate | No clear effect |
Closest papers by meaning
- Latent Introspection: Models Can Detect Prior Concept Injections (NeurIPS 2026 · similarity 0.88)
- Looking Inward: Language Models Can Learn About Themselves by Introspection (ICLR 2025 · similarity 0.84)
- LLM Circuit Analyses Are Consistent Across Training and Scale (NeurIPS 2024 · similarity 0.84)
- The Hawthorne Effect in Reasoning Models: Evaluating and Steering Test Awareness (NeurIPS 2025 · similarity 0.83)
- Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors (NeurIPS 2026 · similarity 0.83)
- Inverted Detection and Control in Steering Vectors (NeurIPS 2026 · similarity 0.83)
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.
