# FRISM: Fine-Grained Reasoning Injection via Subspace-Level Model Merging for Vision-Language Models

Source: [FRISM: Fine-Grained Reasoning Injection via Subspace-Level Model Merging for Vision-Language Models](<https://www.sota2.com/research/paper/697c3aefa8b01e76cc8da38e>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-09T05:00:04.795Z

Authors: Chenyu Huang, Peng Ye, Xudong Tan, Jinhan Mu, Shenghe Zheng, Li Shen, Tao Chen

Published: 2026-01-29T02:36:19.000Z

## Abstract

Efficiently enhancing the reasoning capabilities of Vision-Language Models (VLMs) by merging them with Large Reasoning Models (LRMs) has emerged as a promising direction. However, existing methods typically operate at a coarse-grained layer level, which often leads to a trade-off between injecting reasoning capabilities and preserving visual capabilities. To address this limitation, we propose FRISM (Fine-grained Reasoning Injection via Subspace-level model Merging), a fine-grained reasoning injection framework based on subspace-level model merging. Observing that different SVD subspaces contribute differently to reasoning and perception, FRISM decomposes LRM task vectors via Singular Value Decomposition (SVD) and adaptively tunes the scaling coefficients of each subspace through learning to realize fine-grained reasoning injection. Furthermore, we introduce a label-free self-distillation learning strategy with dual-objective optimization using common vision-language perception datasets. Extensive experiments demonstrate that FRISM effectively improves reasoning capabilities while largely preserving the model's visual capabilities by consistently achieving strong performance across diverse visual-language reasoning benchmarks.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Vision-Language Reasoning on VL Reasoning Benchmarks](<https://www.sota2.com/research/sota/vision-language-reasoning-on-vl-reasoning-benchmarks.md>): Vision-Language Reasoning; VL Reasoning Benchmarks; MVista Score: 74.

- [Vision-Language Reasoning on VL Reasoning Suite (MathVista, MMMU, MMStar)](<https://www.sota2.com/research/sota/vision-language-reasoning-on-vl-reasoning-suite-mathvista-mmmu-mmstar.md>): Vision-Language Reasoning; VL Reasoning Benchmarks MathVista, MVerse, MathVision, MMMU, R1-OV, MMStar; MathVista Acc: 79.8.

- [Vision-Language Perception on VL Perception Benchmarks TextVQA, POPE, Seed-Bench](<https://www.sota2.com/research/sota/vision-language-perception-on-vl-perception-benchmarks-textvqa-pope-seed-bench.md>): Vision-Language Perception; VL Perception Benchmarks TextVQA, POPE, Seed-Bench; TextVQA Score: 85.

- [Vision-Language Perception on VL Perception Benchmarks](<https://www.sota2.com/research/sota/vision-language-perception-on-vl-perception-benchmarks.md>): Vision-Language Perception; VL Perception Benchmarks; TextVQA: 85.5.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2601.21187>)

[PDF](<https://arxiv.org/pdf/2601.21187.pdf>)