AI Paper Atlas

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

GPTailor: Large Language Model Pruning Through Layer Cutting and Stitching

Guinan Su, Li Shen, Lu Yin, Shiwei Liu, Yanwu Yang, Jonas Geiping

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

Read the paper

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
add layer selection from fine-tuned variants and layer merging to layer removal (ablation)GPTailor · OpenCompass 14 tasks (Llama-2-7B)average accuracy over 14 benchmarks: 44.83 -> 48.55Improved
add layer merging on top of layer selection and removal (ablation)GPTailor · OpenCompass 14 tasks (Llama-2-7B)average accuracy over 14 benchmarks: 43.20 -> 48.55Improved
add layer selection without merging to layer removal (ablation)GPTailor · OpenCompass 14 tasks (Llama-2-7B)average accuracy over 14 benchmarks: 44.83 -> 43.20Hurt
merge layers across fine-tuned variants instead of within-model folding (ablation)GPTailor · OpenCompass 14 tasks (Llama-2-7B)average accuracy over 14 benchmarks: 46.26 -> 48.55Improved
use multi-task calibration objective instead of MMLU-only (ablation)GPTailor · OpenCompass 14 tasks (Llama-2-7B)average accuracy over 14 benchmarks: 45.62 -> 48.55Improved
use accuracy on multiple-choice calibration data instead of WikiText perplexity as search objective (ablation)GPTailor · OpenCompass 14 tasks (Llama-2-7B)average accuracy over 14 benchmarks: 25.38 -> 48.55Improved
prune a family of fine-tuned variants instead of a single modelGPTailor · OpenCompass 14 tasks (Llama-2-7B)average accuracy over 14 benchmarks: 42.24 (ShortGPT) -> 44.83 (LR-only) -> 48.55Improved
apply GPTailor to Llama-3 8B instead of ShortGPTGPTailor · OpenCompass 14 tasks (Llama-3-8B)average accuracy retained vs dense: 62.79% -> 84.55%Improved
prune at 50% ratioGPTailor · OpenCompass 14 tasks (Llama-2-7B)average accuracy over 14 benchmarksNo clear effect
use multi-fidelity SMAC with small calibration budgetsGPTailor · search efficiencyfraction of trials at full budget: only 22% of trials use full budget; 41.4% use minimumImproved

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