μ \mu CRASP : Mu ltimodal C hain-of-thought R easoning a ware S tructured P runing

Aritra Dutta, Somak Aditya · 2026

Read μ \mu CRASP : Mu ltimodal C hain-of-thought R easoning a ware S tructured P runing by Aritra Dutta, Somak Aditya (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-25842.txt.

Introduction

Dutta and Aditya present MuCRASP, a structured pruning framework for vision language models preserving step-by-step thought. Pruning ignores thought: pivot tokens decide consistency, and unimodal methods ignore cross-modal differences. MuCRASP targets critical parts, preserving cross modal alignment under a budget. On four models across three tests, it preserves thought under compression. At 30 percent pruning it scores 8.87 against 7.32 on physical tasks. It holds consistency up to 50 percent pruning.

Catalog record

CREATORAritra Dutta, Somak Aditya
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE135831
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.25842

Browse the full library · About the Tower of Babel