VITAL: Visual-Semantic Dual Supervision for Enhanced and Interpretable Latent Reasoning in Medical MLLMs

Researchers introduce VITAL, a latent-space reasoning framework for medical AI models that uses dual visual-semantic supervision to improve medical visual question answering while maintaining interpretability. The method addresses modality collapse and inference efficiency issues in existing approaches, achieving state-of-the-art results on 7 benchmarks using a newly constructed 61K medical imaging dataset.

VITAL represents a significant advancement in making medical AI systems both more efficient and more transparent—two critical requirements for clinical adoption. The framework tackles a fundamental challenge in medical MLLMs: reasoning over continuous hidden states rather than generating explicit reasoning chains, which reduces computational overhead while avoiding the 'language bottleneck.' This efficiency gain matters substantially for real-world deployment in resource-constrained healthcare environments.

The dual supervision approach is architecturally elegant. An auxiliary text decoder helps the model learn to represent reasoning chains in latent space, while a visual projector grounds these representations in concrete medical imagery through ROI feature regression. Crucially, both components can be discarded at inference with zero computational cost, then reattached post-hoc for interpretability—a design choice that prioritizes practical deployment without sacrificing explainability.

The interpretability dimension addresses a critical gap in medical AI. Clinical applications demand transparency in decision-making; black-box latent reasoning, however efficient, creates liability and trust issues. VITAL's ability to provide both textual and visual explanations of reasoning bridges this gap. The new 61K dataset spanning 9 imaging modalities—an order of magnitude larger than prior medical vision-language datasets—provides stronger empirical grounding than previous work.

Results showing competitive performance with trillion-parameter proprietary models using significantly smaller-scale training data suggest the architectural innovations drive gains beyond mere scale. For healthcare AI development, this demonstrates that specialized design for medical reasoning can outperform general-purpose scaling approaches.

• →VITAL achieves state-of-the-art medical VQA results through visual-semantic dual supervision with zero inference overhead
• →Interpretability mechanisms provide textual and visual explanations without sacrificing computational efficiency
• →New 61K medical imaging dataset across 9 modalities substantially expands training resources for medical MLLMs
• →Framework outperforms larger proprietary models, suggesting specialized architecture matters more than scale alone
• →Design enables post-hoc reattachment of explanation modules for flexible deployment scenarios