Better Eyes, Better Thoughts: Why Vision Chain-of-Thought Fails in Medicine

Large vision language models (VLMs) report strong accuracy on medical question-answering, yet it remains unclear whether they reason from visual evidence or exploit textual shortcuts. We introduce a counterfactual evaluation framework that decouples visual and textual contributions by substituting input images with controlled surrogates blank, pixel-shuffled, image-absent, and CLIP-retrieved hard negatives and derive a suite of grounding metrics including the Visual Reliance Score (VRS) and Visual Hallucination Rate (VHR). We further introduce CORAL (COntrastive Retrieval-Augmented Learning), a 7B-parameter LoRA fine-tune of Qwen2.5-VL-7B trained with a Contrastive Grounding Objective (CGO) that penalises answer invariance under hard-negative image swaps. On a paired controlled evaluation across four closed-form medical VQA benchmarks (PathVQA, PMC-VQA, SLAKE, VQA-RAD; n=400 total), CORAL improves macro accuracy by +6.7 pp (P(Delta>0)=0.988) and reduces VHR by 8.0 pp (P<0.001) over the matched Qwen2.5-VL-7B base; neither MedVLThinker RL variant achieves a significant gain on either metric. Cross-domain diagnostics further reveal that image substitution costs only <=6.5 pp on medical benchmarks versus 48-61 pp on general-domain tasks, situating the grounding gap that CGO targets. We discuss evaluation limitations openly including train/eval benchmark overlap and underpowered secondary metrics and release our framework, training code, and model weights to support reproducible grounding audits of medical VLMs.

Medical Vision-Language Models have shown promising potential in clinical decision support, yet they remain prone to factual hallucinations due to insufficient grounding in localized pathological evidence. Existing medical alignment methods primarily operate at the response level through preference optimization, improving output correctness but leaving intermediate reasoning weakly connected to visual regions. Although chain-of-thought (CoT) enhances multimodal reasoning, it remains largely text-centric, limiting effective integration of clinical visual cues. To address this gap, we propose ClinCoT, a clinical-aware visual chain-of-thought framework that transforms preference optimization from response-level correction to visual-driven reasoning. We introduce an automatic data generation pipeline that constructs clinically grounded preference pairs through reasoning with hypotheses-driven region proposals. Multiple Med-LLMs evaluators rank and assign scores to each response, and these rankings serve as supervision to train the target model. We further introduce a scoring-based margin-aware optimization strategy that incorporates both preference ranking and score difference to refine region-level reasoning trajectories. To maintain alignment as the model's policy evolves during training, we adopt an iterative learning scheme that dynamically regenerates preference data. Extensive experiments on three medical VQA and report generation benchmarks demonstrate that ClinCoT consistently improves factual grounding and achieves superior performance compared with existing preference-based alignment methods.

Vision-language models (VLMs) are increasingly adapted through domain-specific fine-tuning, yet it remains unclear whether this improves reasoning beyond superficial visual cues, particularly in high-stakes domains like medicine. We evaluate four paired open-source VLMs (LLaVA vs. LLaVA-Med; Gemma vs. MedGemma) across four medical imaging tasks of increasing difficulty: brain tumor, pneumonia, skin cancer, and histopathology classification. We find that performance degrades toward near-random levels as task difficulty increases, indicating limited clinical reasoning. Medical fine-tuning provides no consistent advantage, and models are highly sensitive to prompt formulation, with minor changes causing large swings in accuracy and refusal rates. To test whether closed-form VQA suppresses latent knowledge, we introduce a description-based pipeline where models generate image descriptions that a text-only model (GPT-5.1) uses for diagnosis. This recovers a limited additional signal but remains bounded by task difficulty. Analysis of vision encoder embeddings further shows that failures stem from both weak visual representations and downstream reasoning. Overall, medical VLM performance is fragile, prompt-dependent, and not reliably improved by domain-specific fine-tuning.