StructVRM: Aligning Multimodal Reasoning with Structured and Verifiable Reward Models

TL;DR

StructVRM introduces a structured, verifiable reward mechanism for multimodal reasoning models, enabling fine-grained feedback and partial credit, leading to state-of-the-art results on complex benchmarks.

Contribution

We propose StructVRM, a novel approach that aligns multimodal reasoning with structured, verifiable rewards, improving model guidance for complex, multi-part questions.

Findings

Achieved state-of-the-art performance on 6 out of 12 benchmarks.

Demonstrated effectiveness on high-difficulty STEM-Bench.

Validated the approach's superiority over traditional reward methods.

Abstract

Existing Vision-Language Models often struggle with complex, multi-question reasoning tasks where partial correctness is crucial for effective learning. Traditional reward mechanisms, which provide a single binary score for an entire response, are too coarse to guide models through intricate problems with multiple sub-parts. To address this, we introduce StructVRM, a method that aligns multimodal reasoning with Structured and Verifiable Reward Models. At its core is a model-based verifier trained to provide fine-grained, sub-question-level feedback, assessing semantic and mathematical equivalence rather than relying on rigid string matching. This allows for nuanced, partial credit scoring in previously intractable problem formats. Extensive experiments demonstrate the effectiveness of StructVRM. Our trained model, Seed-StructVRM, achieves state-of-the-art performance on six out of twelve public multimodal benchmarks and our newly curated, high-difficulty STEM-Bench. The success of StructVRM validates that training with structured, verifiable rewards is a highly effective approach for advancing the capabilities of multimodal models in complex, real-world reasoning domains.