LiteMedCoT-VL: Parameter-Efficient Adaptation for Medical Visual Question Answering

Researchers introduce LiteMedCoT-VL, a technique that transfers chain-of-thought reasoning from large language models to compact 2B parameter models for medical visual question answering, achieving 64.9% accuracy on the PMC-VQA benchmark without relying on image captions. The breakthrough demonstrates that smaller models enhanced with reasoning distillation can match or exceed the performance of larger models, enabling deployment of sophisticated medical AI on resource-constrained clinical devices.

LiteMedCoT-VL addresses a critical challenge in medical AI deployment: the efficiency gap between powerful large models and the portable systems required in clinical environments. While 235B parameter models excel at complex reasoning, their computational demands make real-time clinical deployment impractical. The research demonstrates that knowledge transfer focused specifically on reasoning chains—not just final answers—enables compact models to punch above their weight class.

The medical imaging domain provides an ideal testing ground for this approach because it demands structured reasoning that combines visual evidence with clinical knowledge. Traditional knowledge distillation methods transfer answers without exposing the reasoning process, leaving student models unable to explain their conclusions. By training on explanation-enriched data using LoRA-based fine-tuning, LiteMedCoT-VL bridges this gap, enabling 2B models to achieve reasoning capabilities previously thought to require 4B+ parameters.

The 11 percentage point improvement over baseline performance has significant implications for healthcare technology deployment. Portable devices in hospitals, clinics, and remote settings could run interpretable AI diagnostic support without cloud connectivity or expensive infrastructure. Visual grounding analysis confirms the model relies on image content rather than exploiting textual shortcuts, suggesting genuine clinical utility rather than pattern gaming.

The work signals a broader trend toward efficiency in AI model design. As parameter counts stabilize or decrease across the industry, performance gains increasingly come from training methodology rather than scale. For healthcare specifically, this enables clinically relevant AI at the point of care rather than centralized processing, improving accessibility and reducing latency-sensitive diagnostic delays.

• →Reasoning distillation enables 2B parameter models to match 4B+ model performance on medical visual question answering tasks
• →LiteMedCoT-VL achieves 64.9% accuracy on PMC-VQA, outperforming all published baselines without image captions
• →The approach transfers explainable chain-of-thought reasoning from teachers to students, not just answers
• →Visual grounding analysis confirms models rely on image content rather than textual shortcuts, supporting clinical validity
• →Parameter-efficient fine-tuning enables deployment of interpretable medical AI on resource-constrained clinical devices