AI Paper Atlas

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

FOCUS & RePAIR: Mitigating Text Degeneration via Token-Level Guidance For Pruned Large Language Models

Junyoung Lee, Sehyeon Park, Shinhyoung Jang, Seonha Ryu, Hojeong Kim, Hyunsei Lee, Il hong Suh, Yeseong Kim

ICML 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs · 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
width pruningLlama (unpruned) · Repetition rate (sampling)repetition rate: 5.9% -> 12.4%Hurt
replace two onset tokens with top-2 alternativeLlama · Unpruned model, 1,000 sequencesCREP: 5.9% -> 0.7%Improved
add FOCUS reweighting to KD (ablation)KD · WikiText-103 continuation (width-pruned 25%), Llama-3.1-8BCREP (%): 7.3 -> 1.73Improved
add FOCUS to KD (perplexity) (ablation)KD · WikiText-103 continuation (width-pruned 25%), Llama-3.1-8Bperplexity: 21.69 -> 22.32Hurt
add RePAIR to KD (ablation)KD · WikiText-103 continuation (width-pruned 25%), Llama-3.1-8BCREP (%): 7.3 -> 2.23Improved
add RePAIR to KD (MAUVE) (ablation)KD · WikiText-103 continuation (width-pruned 25%), Llama-3.1-8BMAUVE: 0.61 -> 0.68Improved
add FOCUS on top of RePAIR (ablation)RePAIR · WikiText-103 continuation (width-pruned 25%), Llama-3.1-8BCREP (%): 2.23 -> 0.57Improved
add FOCUS to DITTO (ablation)DITTO · WikiText-103 continuation (width-pruned 25%), Llama-3.1-8BCREP (%): 5.27 -> 0.5Improved
add FOCUS + RePAIR on Llama-2-13B (ablation)KD · WikiText-103 continuation (width-pruned 25%), Llama-2-13BCREP (%): 1.13 -> 0.00Improved
add RePAIR to KD on Self-Instruct (ablation)KD · Self-Instruct (width-pruned 25%)CREP (%): 1.97 -> 0.63Improved
use onset-level pairs (RePAIR) instead of sequence-level DPO (ablation)CE baseline · WikiText-103, Llama-3.1-8Bunique 6-gram repetition (%): 3.45 -> 3.01 (unique 6-gram)Improved
FOCUS vs ToDi distribution-shaping distillation (ablation)KD · WikiText-103 continuation (width-pruned 25%), Llama-3.1-8BCREP (%): 6.73 -> 1.73Improved

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