AdaptMMBench: Benchmarking Adaptive Multimodal Reasoning for Mode Selection and Reasoning Process

TL;DR

AdaptMMBench introduces a comprehensive benchmark for evaluating adaptive multimodal reasoning in vision-language models, focusing on dynamic difficulty assessment, mode selection rationality, and multi-dimensional process analysis.

Contribution

It proposes a novel MCC-based metric and a multi-domain benchmark to evaluate adaptive reasoning, addressing limitations of static difficulty labels and simplistic metrics.

Findings

Adaptive mode selection improves with model capacity but is decoupled from final accuracy.

Key step coverage correlates with performance across models.

Tool effectiveness varies significantly among different architectures.

Abstract

Adaptive multimodal reasoning has emerged as a promising frontier in Vision-Language Models (VLMs), aiming to dynamically modulate between tool-augmented visual reasoning and text reasoning to enhance both effectiveness and efficiency. However, existing evaluations rely on static difficulty labels and simplistic metrics, which fail to capture the dynamic nature of difficulty relative to varying model capacities. Consequently, they obscure the distinction between adaptive mode selection and general performance while neglecting fine-grained process analyses. In this paper, we propose AdaptMMBench, a comprehensive benchmark for adaptive multimodal reasoning across five domains: real-world, OCR, GUI, knowledge, and math, encompassing both direct perception and complex reasoning tasks. AdaptMMBench utilizes a Matthews Correlation Coefficient (MCC) metric to evaluate the selection rationality of different reasoning modes, isolating this meta-cognition ability by dynamically identifying task difficulties based on models' capability boundaries. Moreover, AdaptMMBench facilitates multi-dimensional process evaluation across key step coverage, tool effectiveness, and computational efficiency. Our evaluation reveals that while adaptive mode selection scales with model capacity, it notably decouples from final accuracy. Conversely, key step coverage aligns with performance, though tool effectiveness remains highly inconsistent across model architectures.