AI Paper Atlas

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

Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs

Yiwei Chen, Soumyadeep Pal, Yimeng Zhang, Qing Qu, Sijia Liu

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

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 larger victim LLM (Yi-34B vs Zephyr-7B) for text-based detection of RMU unlearningtext classifier · MMLU prompts, RMUdetection accuracy (%): 53.68 -> 95.77Improved
classify pre-logit activations instead of textual responsesRMU-unlearned Zephyr-7B · MMLU promptsdetection accuracy (%): 53.68 -> 98.59Improved
train detector on a mix of forget and forget-irrelevant responses instead of forget only (ablation)text classifier, RMU · MMLU prompts, Yi-34Bdetection accuracy (%): 61.41 -> 95.77Improved
train detector on forget only versus mixed training data (ablation)text classifier, RMU · WMDP promptsdetection accuracy (%): 95.49 -> 93.24 (Llama-3.1-8B)No clear effect
use NPO instead of RMU as the unlearning methodtext classifier · MMLU prompts, Zephyr-7Bdetection accuracy (%): 53.68 -> 99.86Improved
use a residual MLP head for the text classifier (ablation)text classifier, Llama-3.1-8B RMU · MMLU promptsdetection accuracy (%): 78.87 -> 81.23No clear effect
use a deeper four-layer MLP head for the text classifier (ablation)text classifier, Llama-3.1-8B RMU · MMLU promptsdetection accuracy (%): 78.87 -> 76.54Hurt
test detector on a different model than it was trained onactivation classifier trained on Llama-3.1-8B · RMU, cross-modeldetection accuracy (%): 99.58 -> 48.95 (Qwen2.5-14B)Hurt
test detector trained on Zephyr-7B, Qwen2.5-14B or Yi-34B on other modelsactivation classifier · RMU, cross-modeldetection accuracy (%): 82.45% to 99.58% across modelsNo clear effect
ask forget-irrelevant prompts (MMLU, UltraChat) instead of forget promptstext classifier, RMU Zephyr-7B · WMDP vs MMLUdetection accuracy (%): 90.56 -> 53.68Hurt
apply NPO unlearning instead of RMU on Yi-34Bunlearned model utility · MMLUaccuracy (%): 70.63 -> 41.54Hurt

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