MedAD-R1: Consistency-Reinforced Policy Optimization for Interpretable Medical Anomaly Detection

arXiv:2602.01081 · cs.CV · Submitted 2026-02-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "MedAD-R1: Consistency-Reinforced Policy Optimization for Interpretable Medical Anomaly Detection".

Tom: MedAD-R1 introduces a novel two-stage training framework, combining Supervised Fine-Tuning with Consistency Group Relative Policy Optimization (Con-GRPO),

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So, to get into specifics, this paper introduces "MedAD-R1: Consistency-Reinforced Policy Optimization for Interpretable Medical Anomaly Detection," and the authors are Haitao Zhang and his team. This research is about using a two-stage training framework involving Supervised Fine-Tuning followed by Consistency Group Relative Policy Optimization to create transparent reasoning pathways <ref:2602.01081#pg0>.

Jane: That framework is essentially designed to force the model to build a logical chain of thought that actually supports its final medical diagnosis, rather than just spitting out a result based on superficial patterns. It’s about enforcing coherence throughout the entire reasoning process.

Lu: What’s interesting is their approach to overcoming the limitations of models like LLaVAMed and HuatuoGPT-Vision, which they point out are often hindered by that simple Supervised Fine-Tuning paradigm <ref:2602.01081#pg1>. They argue that this method helps models construct a verifiable, step-by-step causal argument instead of just memorizing statistical patterns

Gudibande et al., two thousand twenty-three: <ref:2602.01081#pg1>.

Meng: I’ve read about some of these papers focusing on RLVR and POPO, and it seems like this paper builds on that idea by adding this specific consistency reward mechanism to enforce the logic layer ATOD. It’s interesting how they combine the policy optimization with a reward structure that prioritizes internal coherence.

Lalam: For me, the focus on interpretability through a "verifiable reasoning process" is what makes this really powerful for cultural adoption in healthcare. When we can see the steps, it builds confidence and helps integrate these tools into clinical workflows more smoothly.

The paper's summary: Tom: So, the core idea of MedAD-R1 is that they first build a large multi-modal benchmark called MedAD-38K, which includes diagnostic Chain-of-Thought annotations alongside structured Visual Question Answering pairs <ref:2602.01081#pg0>. They then use Supervised Fine-Tuning to get a baseline policy, and then apply Consistency Group Relative Policy Optimization to fix the disconnect between the thinking process and the final answer <ref:2602.01081#pg1>.

Jane: Essentially, MedAD-38K is their foundation—a massive dataset covering ten imaging modalities across different anatomical regions, where every image has five specific diagnostic probes like anatomy identification and pathology characterization <ref:2602.01081#pg2>. The model learns the structure from this rich data first before the RL phase begins.

Lu: The reward function they introduce is key here: R(Y) = lambda fmt R fmt(Y) + lambda acc R acc(A, A*) + lambda con R con(T, A, Q), where the consistency reward checks if the deduced answer matches what the model itself generated in its thought process T <ref:2602.01081#pg1>. This explicitly prioritizes logical coherence alongside accuracy and format adherence.

Meng: That composite reward structure is clever because it doesn't just punish a wrong final answer; it actively rewards the model for producing a correct internal narrative, which is what they call the Consistency Reward <ref:2602.01081#pg1>. It’s like training the model to write a good story, not just get the right ending.

Lalam: I find that emphasis on internal logic really speaks to how we build better AI systems for complex tasks. If the system is only rewarded for getting the final label right, it can learn shortcuts; rewarding the process keeps it honest about its thinking.

The paper's improvements: Tom: The main improvement they highlight is overcoming that critical weakness where standard SFT leads to reasoning disconnected from the answer <ref:2602.01081#pg1>. They show how their two-stage training framework, moving from SFT to Con-GRPO, creates a model that delivers transparent, accurate, and logically coherent diagnostic reasoning <ref:2602.01081#pg1>.

Jane: The paper suggests that the improvement lies in injecting this cognitive structure early through SFT—the "Cognitive Injection"—which sets up a capable baseline for the subsequent reinforcement learning phase <ref:2602.01081#pg1>. This initial alignment is what prepares the model to handle the complexity of the Con-GRPO phase.

Lu: They specifically improve performance by showing that rewarding a correct process is more effective than just rewarding an outcome, as evidenced by their results in accuracy, where the consistency-focused setting yielded a result of eighty-four point two one percent <ref:2602.01081#pg3>. This shows that enforcing the reasoning path directly leads to better generalization across modalities like MRI and X-ray <ref:2602.01081#pg4>.

Meng: From a practical standpoint, the improvement is that this methodology allows for high accuracy even with relatively small models, achieving an Overall Accuracy of eighty-five point one five percent on the test set using only a lightweight 3B parameter architecture <ref:2602.01081#pg4>. That efficiency is important when deploying these tools in environments where computational resources are limited.

Lalam: I think this efficiency combined with high consistency is what makes it so valuable for future medical applications. It means we can deploy AI systems that provide a traceable diagnostic pathway without needing massive, expensive computational power for every inference.

Conclusion: Tom: So, to wrap up, the authors of "MedAD-R1: Consistency-Reinforced Policy Optimization for Interpretable Medical Anomaly Detection" have shown how a two-stage training approach using Supervised Fine-Tuning and Con-GRPO can significantly improve Large Multimodal Models by enforcing logical coherence in their reasoning <ref:2602.01081#pg0>. They move past simple correlation to build verifiable, step-by-step causal arguments for medical diagnostics.

Jane: Exactly, Tom. The core implication is that for high-stakes tasks like medical anomaly detection, we need training objectives that prioritize the process of thinking as much as the final decision itself <ref:2602.01081#pg1>. This gives us a tool where the reasoning pathway is transparent and auditable for clinicians.

Lu: The impact on research, I think, is showing that explicit reinforcement mechanisms tied to internal consistency are a viable way to guide model behavior in complex multimodal settings <ref:2602.01081#pg3>. It opens up new avenues for how we structure policy optimization beyond just maximizing final output likelihood

Gudibande et al., two thousand twenty-three: <ref:2602.01081#pg1>.

Meng: For engineering, the result is a framework that’s effective even with smaller models, proving that quality of training signals trumps sheer parameter count for achieving strong reasoning in this context <ref:2602.01081#pg4>. That scalability is what makes it ready for real-world testing.

Lalam: I see the future here as AI systems that are inherently trustworthy because their decisions aren't just black boxes; they have a demonstrated, logical path to get there <ref:2602.01081#pg1>. This level of transparency is crucial for building widespread clinical trust in these powerful tools.

Haitao Zhang, Yingying Wang, Jiaxiang Wang, Haote Xu, Hongyang Zhang, Yirong Chen, Yue Huang

School of Informatics, Xiamen University · Institute of Artificial Intelligence, Xiamen University

cs.CV

Submitted: 2026-02-01

Updated: 2026-10-04

Comments: Revised manuscript with an updated title, expanded experiments, and supplementary material

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: MedAD-R1 introduces a novel two-stage training framework, combining Supervised Fine-Tuning with Consistency Group Relative Policy Optimization (Con-GRPO), to enable Large Multimodal Models (LMMs) to

Key concepts

MedAD Benchmark
This is the first large, multi-modal benchmark for medical imaging (MRI, CT, X-ray) with 38K images. It includes structured Visual Question Answering (VQA) pairs and diagnostic Chain-of-Thought (CoT) annotations across ten anatomical regions to test detailed reasoning.
Cognitive Injection
The first training stage uses Supervised Fine-Tuning (SFT) on the benchmark data. This step teaches the model foundational medical knowledge and aligns its thinking process with a structured 'think-then-answer' paradigm, creating a strong starting point for later optimization.
Consistency Reward
This reward is part of Con-GRPO and ensures logical coherence. It gives a high score only if the final deduced answer perfectly matches the reasoning steps generated by the model itself, forcing the thought process to support the conclusion.

Terminology

Summary

MedAD-R1 introduces a novel two-stage training framework, combining Supervised Fine-Tuning with Consistency Group Relative Policy Optimization (Con-GRPO), to enable Large Multimodal Models (LMMs) to generate transparent and logically coherent reasoning pathways for medical anomaly detection. This method addresses the critical weakness of standard SFT in producing reasoning disconnected from the final answer by incorporating a consistency reward, achieving state-of-the-art performance on a large, multi-modal benchmark.

MedAD Benchmark and Foundation

The research begins by constructing MedAD-38K, described as the first large-scale, multi-modal, and multi-center benchmark for MedAD featuring diagnostic Chain-of-Thought (CoT) annotations alongside structured Visual QuestionAnswering (VQA) pairs. This benchmark aggregates data from 10 distinct imaging modalities across 10 anatomical regions, including MRI, CT, X-ray, and Endoscopy. Crucially, every image in MedAD-38K is enriched with a structured VQA task probing five core diagnostic axes: Anatomy Identification, Modality Classification, Anomaly Detection (e.g., Is there an anomaly present in the image?), Lesion Localization (using discrete spatial regions), Modality Classification, and Pathology Characterization. The benchmark's cornerstone is the integration of diagnostic reasoning pathways created through a multi-stage pipeline that involves generating initial CoT using models like MedGemma and Gemini 2.5 Pro, followed by a rigorous manual verification process to ensure annotations demonstrate self-consistent reasoning.

Two-Stage Training Framework

The proposed training paradigm is structured in two distinct stages designed to instill foundational knowledge and then enforce logical coherence. The first stage is the Cognitive Injection, which uses Supervised Fine-Tuning (SFT) on MedAD-38K to instill foundational medical knowledge and align the model with a structured think-then-answer paradigm. This stage minimizes the negative log-likelihood loss, training the model to maximize the likelihood of generating ground-truth outputs. The resulting policy, denoted as πSFT, serves as a capable baseline and a crucial starting point for the subsequent reinforcement learning phase.

Consistency Group Relative Policy Optimization (Con-GRPO)

The second stage introduces Consistency Group Relative Policy Optimization (Con-GRPO) to overcome the reasoning-answer disconnect. Con-GRPO is built upon Group Relative Policy Optimization (GRPO), a memory-efficient variant of PPO that uses the average reward of a group of sampled outputs as a dynamic baseline, thereby avoiding the need for an additional value function. The core innovation is the composite reward function: R(Y) = λfmtRfmt(Y)+λaccRacc(A, A∗)+λconRcon(T, A, Q), where the weights are set equally to enforce equal priority on structural correctness (Format Reward), final prediction accuracy (Accuracy Reward), and internal logical coherence (Consistency Reward). The Consistency Reward is defined as: 1 if and only if this deduced answer matches the model’s own generated answer, ensuring that the thought process T logically supports the final diagnosis A.

Performance and Contributions

The resulting model, MedAD-R1, achieves state-of-the-art (SOTA) performance on the MedAD-38K benchmark, outperforming strong baselines by more than 10%. Experimental results show that while RL without SFT performs poorly (73.22%), the SFT stage provides a necessary cognitive injection, leading to a strong baseline of 75.41%. Furthermore, the Consistency-focused setting yields the second-best result (84.21%) in accuracy, demonstrating that rewarding a correct process is more effective than merely rewarding an outcome. The final MedAD-R1 model achieves an Overall Accuracy of 85.15% on the test set using a lightweight 3B parameter architecture, validating the framework's efficiency and its ability to produce transparent and internally consistent reasoning pathways that are a prerequisite for clinical trust.

Key Contributions

The paper summarizes its contributions into three main points:

  1. Introduction of MedAD-38K, the first benchmark specifically designed to train and evaluate diagnostic reasoning, enriched with structured VQA pairs and CoT annotations.

  2. Proposal of a novel two-stage training framework featuring Con-GRPO that enables LMMs to cultivate deep, consistent reasoning for medical diagnostics.

  3. The development of MedAD-R1, which significantly outperforms strong baselines on MedAD-38K while generating verifiable reasoning pathways that enhance clinical trustworthiness.

Improvements for AI systems

Based on the scientific paper MedAD-R1: Eliciting Consistent Reasoning in Interpretible Medical Anomaly Detection via Consistency-Reinforced Policy Optimization, here are specific improvements that can be made to existing AI systems, and what those improved systems will be capable of:

) Improved AI Systems and Capabilities:

The core improvement is the development of a robust, two-stage training framework that moves beyond superficial pattern matching in Large Multimodal Models (LMMs) towards generating verifiable, logically coherent diagnostic reasoning.

  1. Improvement Specific Capability of Improved System

2.:---:---

  1. Move from Correlation to Causal Reasoning via Con-GRPO Training The system will no longer just generate fluent text; it will be trained to ensure that every intermediate thought step logically supports the final diagnosis, drastically reducing the risk of hallucinated or contradictory reasoning pathways.

  2. Achieve High Clinical Trust through Consistency Reward The improved system will incorporate a novel reward function that explicitly penalizes inconsistencies between the generated thought process and the final answer. This ensures that if a model claims to diagnose Lymphoma, its internal reasoning steps must be grounded in features characteristic of Lymphoma, making the AI's decision pathway auditable and trustworthy for clinicians.

  3. Enable Robust Multimodal Generalization across Diverse Modalities By constructing the MedAD-38K benchmark (aggregating 10 modalities like MRI, CT, X-ray), the improved system will be capable of performing accurate medical anomaly detection and diagnosis across a vast array of imaging types (e.g., distinguishing pathology in a brain MRI from one in an ultrasound).

  4. Provide Interpretable Diagnostic Outputs The system will output structured data consisting of two distinct parts: a detailed, step-by-step diagnostic trace (blocks) and the final diagnosis (block). This transparency is crucial for clinical decision support, allowing physicians to verify the AI's logic before acting on its recommendation.

  5. Achieve State-of-the-Art Performance with Efficient Architectures The framework (MedAD-R1) demonstrates that a lightweight 3B parameter model can achieve SOTA accuracy (85.15%) by leveraging the structured training signal provided by the benchmark and Con-GRPO, making high-accuracy reasoning accessible for deployment in resource-constrained clinical environments.

  6. Reduce Model Scale Dependency for Reasoning Quality The improvement shows that explicit reasoning reinforcement (Con-GRPO) is more effective than simply scaling up a model. The system can achieve significant performance gains even with smaller backbone models (like the 3B Qwen2.5-VL-3B), suggesting that high-quality reasoning training is a more powerful lever than sheer parameter count for this specific task.

  7. Facilitate Targeted Anomaly Localization The benchmark includes fine-grained spatial tasks (e.g., defining lesion location within a 5-region grid). The improved system will be capable of precisely localizing anomalies on medical images, moving beyond simple presence/absence detection to providing actionable spatial coordinates for clinicians.

Sources

Related papers