AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
ablate the refusal direction (ablation)Gemma3-27B instruct · concept injection, L=37, alpha=2.0detection rate (TPR): 10.8% -> 59.2%Improved
ablate the refusal direction (ablation)Gemma3-27B instruct · concept injection, no injection controlfalse positive rate: 0% -> 7.1%Hurt
use post-trained (instruct) model instead of base model (ablation)Gemma3-27B, OLMo-3.1-32B · concept injectionfalse positive rateImproved
add trained steering vector at layer 29Gemma3-27B · 100 held-out conceptsdetection rate: +74.66%Improved
add trained steering vector at layer 29Gemma3-27B · 100 held-out conceptsforced identification: +21.90%Improved
swap projection on mean-difference direction between success and failure concepts (ablation)success concepts · concept swapdetection rate: 65% -> 36%Hurt
swap projection or residual to success-like for failure concepts (ablation)failure concepts · concept swapdetection rate: 11% -> 34%Improved
patch L45 or L47 MLP with non-injected activations (ablation)Gemma3-27B · 10 random conceptsintrospection rate: to ~27% and ~26%Hurt
ablate top-ranked gate features (ablation)Gemma3-27B steered examples · concept injectiondetection rate: 39% -> 8%Hurt
ablate gates with forced identification (ablation)Gemma3-27B · concept injectionidentification accuracy: 60% -> 46%Hurt
ablate or patch top weak evidence-carrier features (ablation)Gemma3-27B · concept injectiondetection rateNo clear effect
intervene on individual attention heads (ablation)Gemma3-27B · concept injectiondetection rateNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy