AI Paper Atlas

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

De-attribute to Forget for LLM Unlearning

Xinyang Lu, Jiabao Pan, Rachael Hwee Ling Sim, See-Kiong Ng, Anthony Kum Hoe Tung, Bryan Kian Hsiang Low

ICML 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 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 data de-attribution PPO reward (DareU) instead of gradient ascentLLM unlearning · TOFU, Llama2-7B-chatToW: 0.009 -> 0.816Improved
use DareU instead of GDiffLLM unlearning · TOFU, Llama2-7B-chatToW: 0.614 -> 0.816Improved
use DareU instead of NPOLLM unlearning · TOFU, Qwen3-8BToW: 0.638 -> 0.797Improved
use DareU instead of GDiffLLM unlearning · ArXiv, Qwen3-8BToW*: 0.558 -> 0.660Improved
use DareU instead of GA (retain-set ROUGE)LLM unlearning · TOFU, Llama2-7B-chatDR ROUGE: 0.001 -> 0.797Improved
use DareU instead of SCRN on truth ratioLLM unlearning · TOFU, Llama2-7B-chatTruth Ratio: 0.774 -> 0.564Hurt
replace LM classifier with sentence-transformer classifier as attribution function (ablation)DareU · TOFU, Llama2-7B-chatToW: 0.816 -> 0.800No clear effect
replace LM classifier with WASA watermark attribution (ablation)DareU · TOFU, Llama2-7B-chatToW: 0.816 -> 0.806No clear effect
apply DareU with LM classifier, evaluated by hold-out WASA metricDareU · TOFU, Llama2-7B-chatWASA attribution metric (lower is better): 0.854 -> 0.703Improved
use DareU rather than faster baselines (compute cost)DareU · ArXiv, Llama2-7B-chatunlearning time (s): GA 29.83 -> DareU 1595.3Hurt
use DareU instead of retraining (compute cost)DareU · ArXiv, Llama2-7B-chatunlearning time (s): 7852.5 -> 1595.3Improved
ablate hyperparameters (learning rate, kappa, lambda_dis) (ablation)DareU · TOFUunlearning performanceNo 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