MedVLThinker: Framework for Multimodal Medical Reasoning

• MedVLThinker is a framework for chain-of-thought multimodal medical reasoning that generates explicit reasoning traces before producing final answers.
• It integrates a systematic data curation pipeline from text-only and image–text sources with training paradigms such as supervised fine-tuning and reinforcement learning with verifiable rewards.
• Empirical results show that reinforcement learning on curated text data improves performance, setting new state-of-the-art benchmarks in multimodal medical QA.

Searching arXiv for the primary MedVLThinker paper and closely related work to ground the article in current literature. {"query":"MedVLThinker multimodal medical reasoning arXiv", "max_results": 10} MedVLThinker is a fully open framework for building multimodal medical “Large Reasoning Models” (LRMs)—vision-LLMs that explicitly “think before responding” using chain-of-thought (CoT) reasoning. Instantiated on Qwen2.5‑VL (3B/7B/32B), it combines a systematic data curation pipeline for both text-only and image–text medical QA, filtered by reasoning difficulty, with two training paradigms: Supervised Fine-Tuning (SFT) on distilled CoT traces and Reinforcement Learning with Verifiable Rewards (RLVR) based on final answer correctness. Its stated contribution is an open recipe—data, code, model weights, and evaluation scripts—for reproducible research in multimodal medical reasoning (Huang et al., 4 Aug 2025).

1. Conceptual foundation and problem formulation

MedVLThinker treats multimodal medical reasoning through the lens of LRMs. In this formulation, the model is not limited to direct answer prediction; instead, it is trained to produce an explicit reasoning trace before the answer. The paper formalizes this as

$p(a, r \mid x) = p(r \mid x)\, p(a \mid x, r),$

where $x$ is the input, $r$ is a reasoning trace, and $a$ is the final answer. This places MedVLThinker within the broader “think then answer” paradigm rather than standard single-pass vision-language inference (Huang et al., 4 Aug 2025).

The framework is motivated by several barriers identified for multimodal medical QA and VQA. These include lack of open recipes, scarcity of high-quality medical CoT explanations, multimodal data issues such as synthetic noise in PMC‑VQA, evaluation heterogeneity, and underexplored training choices around SFT versus RLVR and text-only versus multimodal data. MedVLThinker is designed as a controlled baseline suite rather than a specialized architectural modification: it uses standard Qwen2.5‑VL as the backbone, vanilla cross-entropy for SFT, and a straightforward application of GRPO with binary, automatically verifiable rewards. This baseline orientation is central to its role in the literature, because the work is framed as a reproducible foundation for comparison rather than as a task-specific agentic system (Huang et al., 4 Aug 2025).

A common misconception is that multimodal medical reasoning systems must derive their gains primarily from multimodal CoT supervision. MedVLThinker explicitly tests this assumption and reports a different outcome. Another common misconception is that SFT on teacher-generated CoT is inherently beneficial for reasoning-centric medical LMMs. The reported experiments do not support that as a general rule. These points matter because the paper’s main significance lies less in proposing a new architecture than in clarifying which training choices actually improve open multimodal medical reasoning.

2. Data curation and difficulty filtering

The data pipeline is a defining component of MedVLThinker. It draws from two sources: the text-only medical QA dataset m23k and the image–text medical VQA dataset PMC‑VQA. The text-only source aggregates MedQA, MedMCQA, and HeadQA into approximately 23,493 multiple-choice medical questions, each with candidate answers, a correct answer, and chain-of-thought reasoning distilled from DeepSeek‑R1. The multimodal source, PMC‑VQA, contains 176,948 visual QA pairs from PubMed Central figures and captions, with approximately 149k images and synthetic QAs annotated by GPT‑3.5 (Huang et al., 4 Aug 2025).

Rather than training on all available examples, MedVLThinker performs difficulty probing with Qwen2.5‑VL‑Instruct. For each question $q$, the procedure samples 16 answers with nucleus sampling at temperature 1.0 and records a pass count $c \in \{0,\dots,16\}$, the number of correct responses. Although pass-count distributions are examined for 3B, 7B, and 32B models, filtering is ultimately based on the 3B model. Questions are retained only when $1 \le c \le 6$; examples with $c = 0$ are treated as too hard, and those with $c \ge 7$ as too easy. The rationale given is that medium-difficulty questions provide the most useful signal for reasoning training, while trivial questions offer little signal and impossible ones produce noisy RL signals (Huang et al., 4 Aug 2025).

This filtering substantially changes the training corpora. The text-only set is reduced from 23,493 to 16,512 questions, and the image–text set is reduced from 176,948 to 115,456 QA pairs. For SFT, the text-only subset uses existing DeepSeek‑R1 CoTs as targets, whereas the image–text subset uses GPT‑4o to generate detailed CoTs given the image and question. The resulting training units therefore differ not only in modality but also in the provenance and likely noise characteristics of their rationales (Huang et al., 4 Aug 2025).

The paper’s analysis of data quality is consequential. It characterizes m23k as high-quality, built from human-authored exam questions with strong DeepSeek reasoning, while PMC‑VQA is described as noisy, synthetic, and often shallow or flawed. This directly informs one of the paper’s central findings: the modality of the training data is less important than the quality and reasoning value of the supervision.

3. Training paradigms: SFT, RLVR, and the text-only result

MedVLThinker compares two training paradigms. In SFT, the model is trained to imitate teacher reasoning traces and answers using standard token-level cross-entropy: $$\mathcal{L}_{\text{SFT}} = - \sum_{t} \log p_\theta(y_t \mid y_{<t}, x).$$ In RLVR, the model is trained using only final answer correctness as supervision. The state is the multimodal input $x$0, the action sequence is the generated token sequence $x$1, and the reward is binary: if the output contains the required > ... <answer> ... </answer> format and the extracted final answer matches ground truth, $x$2; otherwise, $x$3 (Huang et al., 4 Aug 2025).

RLVR is implemented with Group Relative Policy Optimization (GRPO). For each training question, MedVLThinker samples $x$4 rollouts, computes rewards $x$5, whitens them into group-based advantages

$x$6

and optimizes a PPO-like clipped objective with an additional KL regularization term. No explicit value network or critic is required. Training uses VERL + GRPO, learning rate $x$7, KL coefficient 0.01, 8 samples per prompt, and sequence length up to approximately 2048 tokens (Huang et al., 4 Aug 2025).

The empirical comparison between SFT and RLVR is the paper’s most discussed result. For 3B, the base Qwen2.5‑VL‑3B‑Instruct averages 49.14% across six benchmarks. SFT on text-only m23k reduces this to 32.80%, whereas RL on text-only m23k raises it to 53.19%, the best 3B result. For 7B, the base model averages 53.50%; SFT on text-only drops it to 43.83%, while RL on text-only raises it to 54.88%, again the best 7B result. RL on image–text PMC‑VQA helps less, and combined strategies such as SFT on text-only followed by RL on image–text or RL on text-only followed by RL on image–text are consistently worse than direct RL on text-only m23k (Huang et al., 4 Aug 2025).

This produces two counter-intuitive conclusions. First, RLVR dominates SFT in these settings. Second, under RLVR, curated text-only reasoning data provides a larger boost than multimodal image–text data, even though the final model is a vision-LLM. The paper attributes this to the relative quality of the two corpora. This suggests that, for multimodal medical reasoning, reasoning supervision quality can matter more than immediate multimodal alignment when the optimization signal is final answer correctness (Huang et al., 4 Aug 2025).

4. Model instantiation, prompting, and reproducibility

MedVLThinker does not modify the Qwen2.5‑VL architecture. All variants are based on Qwen2.5‑VL‑Instruct at 3B, 7B, and 32B scales, with a transformer-based language backbone, a ViT-like vision encoder, and a connector that maps visual tokens into the language token space. Multimodal fusion occurs through the transformer’s self-attention over text and visual tokens. The paper states that it relies on Qwen2.5‑VL’s default image resolution and tokenizer, with a practical context window of approximately 2k tokens in the reported experiments (Huang et al., 4 Aug 2025).

A consistent prompt format is used for both SFT and RLVR:

$x$9

This template is not a minor formatting choice. During RLVR, adherence to the <think> and <answer> structure is part of the reward function, so formatting becomes part of the policy being optimized. The paper reports that, over training, the model learns to separate reasoning from answer and develops more structured, multi-step reasoning traces (Huang et al., 4 Aug 2025).

The training stack is also unusually explicit for a medical multimodal reasoning paper. SFT uses FSDP and gradient checkpointing. RLVR uses VERL, VLLM for efficient rollouts, mixed precision bf16, and 8×A100 GPUs. Reported hyperparameters include SFT learning rate $x$8, batch size 32, and 3 epochs in the main text, while the appendix notes 5 epochs in implementation details. RL uses 5 epochs on text-only data, 1 epoch on image-text data, batch size 128 for text-only RL, 256 for single-stage image-text RL, and 64 when image-text RL is applied as a second stage (Huang et al., 4 Aug 2025).

Reproducibility is treated as a first-class output. The paper releases curated datasets, model weights, code for data curation, SFT and RLVR training, and evaluation scripts for six benchmarks. This includes pass-count probing, difficulty filtering, teacher CoT generation with DeepSeek‑R1 and GPT‑4o, rollout generation, reward implementation, and standardized evaluation protocols. In the context of medical multimodal reasoning, where open end-to-end recipes have often been missing, this release structure is part of the framework’s substantive contribution rather than merely an implementation detail (Huang et al., 4 Aug 2025).

5. Empirical results and benchmark position

MedVLThinker is evaluated on six multimodal medical QA datasets: PMC‑VQA, MMMU‑Health, MedXpert‑MM, PathVQA, SLAKE, and VQA‑Rad. The metric is accuracy with greedy decoding, and each evaluation is run three times with standard deviation below 0.1. The work also compares against open medical LMMs such as LLaVA‑Med v1.5 Mistral‑7B, HuatuoGPT‑Vision‑7B, and MedGemma, as well as proprietary GPT‑4o and GPT‑4o-mini at 32B scale (Huang et al., 4 Aug 2025).

At 7B, the best MedVLThinker variant is RL on text-only m23k, with an average of 54.88%. This exceeds HuatuoGPT‑Vision‑7B at 54.69%, LLaVA‑Med v1.5 Mistral‑7B at 44.05%, and MedGemma 4B/27B at 50.86% / 49.49%. The paper therefore characterizes MedVLThinker‑7B RL(○) as a new state of the art among open models on this benchmark suite. At 32B, the base Qwen2.5‑VL‑32B‑Instruct averages 60.20%, while MedVLThinker‑32B RL(○ m23k) reaches 63.12%, compared with GPT‑4o-mini at 58.24% and GPT‑4o at 63.74%. The 32B result is thus reported as essentially on par with GPT‑4o (Huang et al., 4 Aug 2025).

The benchmark-level trends are also informative. On the challenging MedXpert‑MM subset, 3B RL(○) reaches 22.90% and 7B RL(○) reaches 24.43%, indicating that scaling helps but does not eliminate the difficulty of complex multimodal medical reasoning. The paper further notes that RLVR increases pass counts empirically, meaning that correct answers become more likely across multiple samples, although it does not provide explicit process-level correctness or hallucination metrics (Huang et al., 4 Aug 2025).

Two findings are especially important for interpretation. First, strong performance is obtained without architectural changes to the underlying VLM. Second, the gain is not uniform across all tasks; it is largest where reasoning supervision and answer verifiability align cleanly with the task format. This suggests that MedVLThinker is especially well-matched to single-turn QA with deterministic or multiple-choice answers, and that its apparent “reasoning” gains are tightly linked to the verifiability regime under which RLVR is defined.

6. Relation to adjacent work, limitations, and future directions

MedVLThinker belongs to a broader shift from static medical VLMs toward reasoning-centric and agentic systems. In relation to general vision-language reasoning, “OpenVLThinker” investigates whether reasoning capabilities analogous to DeepSeek‑R1 can be integrated into large vision-LLMs through an iterative SFT+RL loop, with caption-based distillation from text-only R1 models and iterative GRPO training (Deng et al., 21 Mar 2025). MedVLThinker differs in emphasis: it does not propose an iterative self-improvement loop, and its headline result is that a simple RLVR recipe on curated text-only medical data yields the strongest open medical LMM baselines (Huang et al., 4 Aug 2025).

Within medical multimodal reasoning specifically, MedThink is an earlier rationale-driven MedVQA framework that augments VQA‑RAD and SLAKE with “Medical Decision-Making Rationale” annotations and trains a lightweight generative model to produce answers and rationales (Gai et al., 2024). MedVLThinker moves beyond rationale-supervised generation to a reinforcement learning setting with verifiable rewards. A plausible implication is that the two lines of work emphasize different aspects of “thinking”: MedThink foregrounds interpretability through explicit rationales, whereas MedVLThinker foregrounds reproducible scaling behavior under binary correctness rewards.

More agentic medical extensions go further than MedVLThinker’s baseline recipe. “Med‑VRAgent” adds ROI extraction, Visual Token Edit, Teacher–Student–Assessor loops, Monte Carlo Tree Search, Retrieval‑Augmented Reflection, and PPO fine-tuning, specifically to address hallucinations, vague descriptions, inconsistent logic, and poor localization in medical VLMs (Guo et al., 21 Oct 2025). “MEDVISTAGYM” formulates tool-integrated medical visual reasoning as a POMDP and trains MedVistaGym‑R1 to interleave tool use with agentic reasoning through trajectory sampling and end-to-end reinforcement learning, arguing that structured agentic training—not tool access alone—unlocks effective tool-integrated reasoning for medical image analysis (Lu et al., 12 Jan 2026). These systems are not replacements for MedVLThinker’s contribution; rather, they indicate downstream directions in which a reasoning-centric medical baseline can be extended toward visual grounding, interactive tools, and long-horizon agentic behavior.

The limitations of MedVLThinker are explicit. PMC‑VQA is synthetic and noisy; filtering thresholds are static and derived from the 3B model’s pass counts; RLVR is applied to single-turn QA with multiple-choice or short deterministic answers; and even the 32B model does not capture the full complexity and variability of real-world clinical practice. The paper also states that the models are not ready for unsupervised clinical use, citing clinical safety, dataset bias, misuse risk, and the limitation that RLVR optimizes final answer correctness rather than process-level faithfulness (Huang et al., 4 Aug 2025).

The future directions identified are correspondingly pragmatic: better multimodal data, adaptive curricula, multi-turn clinical dialogues, vision–language grounding, temporal reasoning across image sequences, and more nuanced rewards such as partial credit or process-level rewards. This suggests that MedVLThinker’s enduring significance may lie less in any single model checkpoint than in establishing a controlled, open baseline from which more grounded, tool-integrated, and clinically constrained medical reasoning systems can be developed.