Research map / Large language models / Language, safety and interpretability
LLM Self-Recognition: Steering and Retrieving Activation Signatures
Thibaud Ardoin, Jonas Schäfer, Gerhard Wunder
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 activation classifier instead of perplexity, no prompt | perplexity baseline, Llama-3.1-8B · XL-Sum no prompt | AUROC (%): 47.86 -> 99.16 | Improved |
| use activation classifier instead of perplexity, no prompt | perplexity baseline, Ministral-3-8B · XL-Sum no prompt | AUROC (%): 32.33 -> 99.99 | Improved |
| use activation classifier instead of perplexity, with prompt | perplexity baseline, Llama-3.2-1B · XL-Sum with prompt | AUROC (%): 97.07 -> 99.82 | Improved |
| scale model from Llama-3.2-1B to Llama-3.1-8B | steering attribution · ELI5 text-level | F1: 85.3 -> 99.1 | Improved |
| scale model from Llama-3.2-1B to Llama-3.1-8B | steering attribution · Fresh News token-level | F1: 69.5 -> 90.5 | Improved |
| aggregate token predictions by majority voting to text level | steering MLP attribution, Llama-3.1-8B · ELI5 | F1: 94.0 -> 99.1 | Improved |
| paraphrase generated text with DIPPER-XXL | MLP text-level attribution · Llama-3.1-8B steering attribution | accuracy (%): 99.1 -> 89.3 | Hurt |
| use trained MLP instead of cosine similarity | cosine similarity text-level · Llama-3.1-8B steering attribution | accuracy (%): 84.6 -> 99.1 | Improved |
| add lowercase, punctuation removal and length matching controls (ablation) | activation self-recognition classifier · XL-Sum all domains | accuracy (pp): -0.20 | No clear effect |
| train on higher-entropy News and test on ELI5 vs Base to Ministral | steering MLP cross-evaluation · Llama-3.1-8B-Instruct to Ministral-3-8B-Base News | F1 token/text: 0.90 / 0.99 -> 0.55 / 0.52 | Hurt |
| apply steering vector (alpha=5, 99.7% sparse) | Ministral-3-8B · MMLU 5-shot | accuracy: 75.06% -> 70.54% | Hurt |
| apply steering vector (alpha=5, 99.7% sparse) | Llama-3.1-8B-Instruct · MMLU 5-shot | accuracy: 67.04% -> 66.31% | Hurt |
Closest papers by meaning
- SelfIE: Self-Interpretation of Large Language Model Embeddings (ICML 2024 · similarity 0.88)
- Inspection and Control of Self-Generated-Text Recognition Ability in Llama3-8b-Instruct (ICLR 2025 · similarity 0.87)
- Training-free LLM-generated Text Detection by Mining Token Probability Sequences (ICLR 2025 · similarity 0.87)
- Idiosyncrasies in Large Language Models (ICML 2025 · similarity 0.86)
- LLM Layers Immediately Correct Each Other (NeurIPS 2025 · similarity 0.86)
- DALD: Improving Logits-based Detector without Logits from Black-box LLMs (NeurIPS 2024 · similarity 0.86)
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.
