AI Paper Atlas

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

Divergence Decoding: Inference-Time Unlearning via Auxiliary Models

Humzah Merchant, Bradford Levy

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 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 Divergence Decoding instead of gradient-based unlearning baselinesDivergence Decoding · TOFUaggregate scoreImproved
use Divergence Decoding on MUSE Knowledge MemorizationDivergence Decoding · MUSE Knowledge Memorizationdistance to RetrainImproved
use Divergence Decoding on MUSE Verbatim MemorizationDivergence Decoding · MUSE Verbatim Memorizationdistance to RetrainHurt
use rank-based DD instead of linear DD (ablation)Divergence Decoding · MUSE and TOFUdistance / aggregate scoreMixed
scale auxiliary models from 1.3B to 2.7B (ablation)Divergence Decoding on MUSE · MUSEdistance to RetrainImproved
scale auxiliary models from 1B to 3B (ablation)Divergence Decoding on TOFU · TOFUaggregate scoreImproved
use trigram LM auxiliaries instead of small neural LMs (ablation)Divergence Decoding · TOFUaggregate scoreHurt
use trigram LM auxiliaries instead of small neural LMs (ablation)Divergence Decoding · MUSEdistance to RetrainMixed
use auxiliary models from a different family or tokenizer (OLMo-2, Gemma-3, Qwen3) (ablation)Divergence Decoding · TOFUaggregate scoreNo clear effect
distill Linear DD teacher into the target modelDivergence Decoding · TOFU and MUSEunlearning scoreImproved
apply Distill DD under Leak@K adversarial promptingDistill DD · TOFU Leak@KleakageImproved
increase alpha in linear DD for image generation (1 to 5) (ablation)Linear DD · ImageNet dog forget classesclass presence VQAScore: 20% -> 1%Mixed

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