AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

LLM Self-Recognition: Steering and Retrieving Activation Signatures

Thibaud Ardoin, Jonas Schäfer, Gerhard Wunder

ICML 2026 · Large language models › Language, safety and interpretability · approach: Data and hallucination detection · 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
use activation classifier instead of perplexity, no promptperplexity baseline, Llama-3.1-8B · XL-Sum no promptAUROC (%): 47.86 -> 99.16Improved
use activation classifier instead of perplexity, no promptperplexity baseline, Ministral-3-8B · XL-Sum no promptAUROC (%): 32.33 -> 99.99Improved
use activation classifier instead of perplexity, with promptperplexity baseline, Llama-3.2-1B · XL-Sum with promptAUROC (%): 97.07 -> 99.82Improved
scale model from Llama-3.2-1B to Llama-3.1-8Bsteering attribution · ELI5 text-levelF1: 85.3 -> 99.1Improved
scale model from Llama-3.2-1B to Llama-3.1-8Bsteering attribution · Fresh News token-levelF1: 69.5 -> 90.5Improved
aggregate token predictions by majority voting to text levelsteering MLP attribution, Llama-3.1-8B · ELI5F1: 94.0 -> 99.1Improved
paraphrase generated text with DIPPER-XXLMLP text-level attribution · Llama-3.1-8B steering attributionaccuracy (%): 99.1 -> 89.3Hurt
use trained MLP instead of cosine similaritycosine similarity text-level · Llama-3.1-8B steering attributionaccuracy (%): 84.6 -> 99.1Improved
add lowercase, punctuation removal and length matching controls (ablation)activation self-recognition classifier · XL-Sum all domainsaccuracy (pp): -0.20No clear effect
train on higher-entropy News and test on ELI5 vs Base to Ministralsteering MLP cross-evaluation · Llama-3.1-8B-Instruct to Ministral-3-8B-Base NewsF1 token/text: 0.90 / 0.99 -> 0.55 / 0.52Hurt
apply steering vector (alpha=5, 99.7% sparse)Ministral-3-8B · MMLU 5-shotaccuracy: 75.06% -> 70.54%Hurt
apply steering vector (alpha=5, 99.7% sparse)Llama-3.1-8B-Instruct · MMLU 5-shotaccuracy: 67.04% -> 66.31%Hurt

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