aiqbench analysis
VITAL addresses critical limitations in medical multimodal large language models by introducing visual-semantic dual supervision to enhance latent reasoning without inference overhead. By reconstructing reasoning chains and regressing region-of-interest features, the framework achieves state-of-the-art performance on seven benchmarks while providing essential interpretability for clinical deployment. This approach effectively bridges the gap between efficient continuous-state reasoning and the transparency required in high-stakes medical applications.
VITAL: Visual-Semantic Dual Supervision for Enhanced and Interpretable Latent Reasoning in Medical MLLMs
arXiv:2605.28422v2 Announce Type: replace-cross Abstract: Latent reasoning enables reasoning over continuous hidden states rather than explicit tokens, avoiding the language bottleneck and inference overhead of chain-of-thought for medical VQA. However, existing methods suffer from modality collapse, insufficient visual supervision, and train-inference mismatch. Moreover, their opaque latent states offer no interpretability, which is critical in clinical applications. We propose VITAL, a latent-space reasoning framework for medical MLLMs with visual-semantic dual supervision: an auxiliary text decoder reconstructs reasoning chains from latent states, while a visual projector regresses ROI features from a frozen, independent medical vision encoder. Both modules are discarded at inference with zero overhead, yet can be re-attached post-hoc for dual interpretability, providing textual and visual explanations of the reasoning process without sacrificing efficiency. We construct a 61K dataset spanning 9 imaging modalities, exceeding prior medical visual latent reasoning datasets by an order of magnitude. Experiments on 7 benchmarks show that VITAL consistently and substantially outperforms the backbone, all latent reasoning baselines, and medical MLLMs trained on far larger data, achieving state-of-the-art results competitive with trillion-parameter proprietary models.
Abstract:Latent reasoning enables reasoning over continuous hidden states rather than explicit tokens, avoiding the language bottleneck and inference overhead of chain-of-thought for medical VQA. However, existing methods suffer from modality collapse, insufficient visual supervision, and train-inference mismatch. Moreover, their opaque latent states offer no interpretability, which is critical in clinical applications. We propose VITAL, a latent-space reasoning framework for medical MLLMs with visual-semantic dual supervision: an auxiliary text decoder reconstructs reasoning chains from latent states, while a visual projector regresses ROI features from a frozen, independent medical vision encoder. Both modules are discarded at inference with zero overhead, yet can be re-attached post-hoc for dual interpretability, providing textual and visual explanations of the reasoning process without sacrificing efficiency. We construct a 61K dataset spanning 9 imaging modalities, exceeding prior medical visual latent reasoning datasets by an order of magnitude. Experiments on 7 benchmarks show that VITAL consistently and substantially outperforms the backbone, all latent reasoning baselines, and medical MLLMs trained on far larger data, achieving state-of-the-art results competitive with trillion-parameter proprietary models.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2605.28422 [cs.CV]
(or arXiv:2605.28422v2 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2605.28422 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 270 words
Related coverage
Robust-U1: Can MLLMs Self-Recover Corrupted Visual Content for Robust Understanding?
arXiv · Jun 9, 2026
Decoupling Intention from Trajectory: A Representational Deduction Framework for World Action Models
arXiv · Sep 28, 2026
Anatomy-Anchored Self-Supervision: Distilling Vision Foundation Models for Invariant Ultrasound Representation
arXiv · Jun 4, 2026
Anatomy-Anchored Self-Supervision: Distilling Vision Foundation Models for Invariant Ultrasound Representation
arXiv · Jun 3, 2026