StructVRM: Aligning Multimodal Reasoning with Structured and Verifiable Reward Models
This addresses the need for better multimodal reasoning in AI, particularly for complex tasks like STEM, though it is an incremental improvement over existing reward mechanisms.
The paper tackles the problem of Vision-Language Models struggling with complex, multi-question reasoning tasks by introducing StructVRM, which uses structured and verifiable reward models for fine-grained feedback, achieving state-of-the-art performance on six out of twelve public multimodal benchmarks and a new STEM-Bench.
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.