Damage-Aware Bandit Pruning for Vision and Language Transformers

arXiv cs.AI Papers

Summary

This paper proposes a damage-aware multi-armed bandit method for structured post-training pruning of vision and language transformers, showing reduced performance degradation compared to baseline approaches in experiments across various models and datasets.

arXiv:2609.05448v1 Announce Type: new Abstract: Structured post-training pruning of transformers requires selecting complete functional units whose suppression causes limited degradation. We formulate structured-unit selection for language and vision transformers as a damage-aware multi-armed bandit problem under a fixed candidate-evaluation budget. Attention heads and MLP channel groups are temporarily masked on calibration batches. Paired damage is the masked loss minus the base loss on the same batch, reducing batch-to-batch variation. A smooth bounded reward drives either a UCB-style policy or fractional-Beta Thompson Sampling, and the final mask is constructed sequentially by adding one unit at each step. The selected units are functionally zeroed in the original dense checkpoint; therefore, the reported parameter effects represent effective structural suppression rather than physical compression or measured speedup. Experiments on WikiText-2, LAMBADA, and Imagenette cover GPT-2, OPT, Pythia, Qwen2.5, SmolLM2, ViT-B/16, DeiT-Tiny, and Swin-Tiny, with comparisons against random, magnitude, static-saliency, and budgeted-greedy selection. Across five seeds, the bandit methods usually reduce degradation relative to budgeted greedy in the paired language-model comparisons. Of 28 comparisons highlighted in the paper, 23 bootstrap confidence intervals exclude zero and 11 paired tests have p < 0.05; six have q < 0.05 after Benjamini-Hochberg correction across the full family of 116 dataset-wise tests. Matched-evaluation results for ViT-B/16 and Swin-Tiny indicate that their gains are not explained solely by a larger candidate-evaluation budget.
Original Article
View Cached Full Text

Cached at: 09/10/26, 08:35 AM

# Damage-Aware Bandit Pruning for Vision and Language Transformers
Source: [https://arxiv.org/abs/2609.05448](https://arxiv.org/abs/2609.05448)
[View PDF](https://arxiv.org/pdf/2609.05448)

> Abstract:Structured post\-training pruning of transformers requires selecting complete functional units whose suppression causes limited degradation\. We formulate structured\-unit selection for language and vision transformers as a damage\-aware multi\-armed bandit problem under a fixed candidate\-evaluation budget\. Attention heads and MLP channel groups are temporarily masked on calibration batches\. Paired damage is the masked loss minus the base loss on the same batch, reducing batch\-to\-batch variation\. A smooth bounded reward drives either a UCB\-style policy or fractional\-Beta Thompson Sampling, and the final mask is constructed sequentially by adding one unit at each step\. The selected units are functionally zeroed in the original dense checkpoint; therefore, the reported parameter effects represent effective structural suppression rather than physical compression or measured speedup\. Experiments on WikiText\-2, LAMBADA, and Imagenette cover GPT\-2, OPT, Pythia, Qwen2\.5, SmolLM2, ViT\-B/16, DeiT\-Tiny, and Swin\-Tiny, with comparisons against random, magnitude, static\-saliency, and budgeted\-greedy selection\. Across five seeds, the bandit methods usually reduce degradation relative to budgeted greedy in the paired language\-model comparisons\. Of 28 comparisons highlighted in the paper, 23 bootstrap confidence intervals exclude zero and 11 paired tests have p < 0\.05; six have q < 0\.05 after Benjamini\-Hochberg correction across the full family of 116 dataset\-wise tests\. Matched\-evaluation results for ViT\-B/16 and Swin\-Tiny indicate that their gains are not explained solely by a larger candidate\-evaluation budget\.

## Submission history

From: Salem Ameen \[[view email](https://arxiv.org/show-email/05400591/2609.05448)\] **\[v1\]**Sun, 2 Aug 2026 23:57:18 UTC \(3,177 KB\)

Similar Articles

An AI4AI Framework for Visual Token Pruning

Hugging Face Daily Papers

AutoPrune is a training-free framework that uses LLMs to automatically design visual-token pruning policies for multimodal LLMs via a domain-specific language and residual search, achieving high efficiency with minimal performance loss (99% performance retained while removing 94.4% of visual tokens).