Papers/2609.05448
🧪 Test?View on arXiv

Damage-Aware Bandit Pruning for Vision and Language Transformers

Author1, Author2, Author3, Author4, Author5

pruningtransformersbandit algorithmsmultimodal
2609.05448
Builder Relevance
80%
1h ago

Abstract

This paper presents a method for structured post-training pruning of transformers that minimizes performance degradation by treating the selection of functional units as a damage-aware multi-armed bandit problem.

Reality Card

Core Claim

The proposed bandit pruning method reduces degradation in transformer models compared to traditional budgeted greedy selection methods across multiple datasets.

Method / Result

In 28 comparisons, 23 bootstrap confidence intervals exclude zero, indicating significant improvements.

Limitations

The method's effectiveness may vary across different models and datasets, which could affect reproducibility.

Paper to code

Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.

No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.
← Back to all papers