AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Rethinking Benign Relearning: Syntax as the Hidden Driver of Unlearning Failures

Sangyeon Yoon, Hyesoo Hong, Wonje Jeung, Albert No

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 2 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
unlearn with syntactically diversified (paraphrased) forget set instead of originalGA/NPO unlearning on TOFU forget05 · TOFU Real Authorsutility average: 0.4014 -> 0.4852Improved
unlearn with syntactically diversified forget setunlearning on TOFU forget05 · TOFU Retain setutility average: 0.1607 -> 0.3128Improved
unlearn with syntactically diversified forget setunlearning on TOFU forget05 · TOFU World Factsutility average: 0.6056 -> 0.6104Improved
unlearn with syntactically diversified forget setunlearning on TOFU forget05 · TOFU World FactsProbability: 0.4187 -> 0.4169No clear effect
use diversified forget set to reduce syntactic similarity to the syntactic relearn setunlearning on TOFU · TOFU forget05syntactic similarity: 0.4513 -> 0.2241Improved
unlearn with diversified forget setunlearning on TOFU · TOFU forget05 relearning with syntactic setrelearn success rateImproved
use syntactically similar relearn set instead of topically relevant setGA unlearned model · TOFU forget05relearn success rateHurt
use SCRUB instead of GA/NPO for unlearningunlearning on TOFU · TOFU forget05relearn success rateHurt
evaluate relearning at every step with equal step budget (best step) instead of one epochBLUR relearning evaluation · WMDP/WHP/RWKUROUGE-LNo clear effect
relearn with LoRA instead of full parameters on fully unlearned modelTOFU unlearned model · TOFU forget05relearn success rateHurt
use safety training (DPO) instead of unlearning algorithmsunlearning · TOFU syntactic relearningrelearn vulnerabilityHurt
relearn on perfectly retrained model with topic or syntactic setretrain model · TOFUrelearn success rate: 0% -> 0%No 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