Abstract: External tools substantially expand what multimodal large language models (MLLMs) can retrieve and compute on reasoning-intensive visual questions, and aggregate accuracy typically improves when tools are available. Our empirical study across six VQA benchmarks shows that aggregate accuracy gains under tool access can coexist with answer reversals: questions answered correctly without tools but incorrectly with tools. Trajectory analysis reveals substantially shorter reasoning turns under tool access, suggesting a shift from intrinsic visual reasoning toward reliance on external outputs. We propose ReVision, a training-free test-time method that registers intrinsic reconsideration as an optional action alongside search and code execution. When invoked, a separate tool-free multimodal call re-examines the original image, integrates accumulated evidence, and returns a revised analysis. The MLLM chooses if and when to use it, without fixed reflection stages or paired rollouts. ReVision improves over standard tool use on all six benchmarks (1.28–7.51\%) and outperforms Dual-Rollout Combine on five, with reasoning-token use comparable to standard tool use.
Read the original article:
ReVision: Selective Intrinsic Reasoning for Tool-Augmented, Reasoning-Intensive VQA
