CAVE: A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning

TL;DR

CAVE introduces a structured credit assignment method for visual reasoning that enhances the integration of fragmented visual evidence in vision-language models, leading to improved performance and robustness.

Contribution

The paper proposes CAVE, a novel process-reward approach based on GRPO, and introduces TRACER-Bench for evaluating nonlocal visual reasoning, advancing the state of multimodal reasoning.

Findings

CAVE significantly improves performance on fragmented visual evidence tasks.

CAVE enhances robustness in longer-range and deeper cross-region dependencies.

Experiments show CAVE outperforms existing methods on TRACER-Bench and public benchmarks.

Abstract

Vision-Language Models (VLMs) have achieved strong performance on general multimodal reasoning, yet remain challenged in integrating nonlocal visual information to support semantically underdetermined visual reasoning. We describe this challenge as Fragmented Visual Reasoning. To this end, we propose Credit Assignment for Visual Evidence (CAVE), a structured process-reward method based on GRPO for interleaved visual reasoning. Specifically, CAVE evaluates the contribution of intermediate steps at the action level via three complementary reasoning process signals: belief update, evidence acquisition, and adaptive focus control, thereby guiding the model to optimize each reasoning action and learn more reliable visual reasoning strategies. Meanwhile, we construct TRACER-Bench, which covers four nonlocal and semantically confusable reasoning dimensions and provides key intermediate evidence to supervise reasoning paths. Experiments demonstrate that CAVE substantially improves performance on tasks requiring fragmented visual evidence integration, covering both public benchmarks and our newly introduced TRACER-Bench, while retaining competitive performance on general multimodal evaluations. Further analyses reveal that CAVE effectively improves the visual reasoning capacity and exhibits stronger robustness under longer-range and deeper cross-region dependencies.