Recent efforts to combine low-rank adaptation (LoRA) with mixture-of-experts (MoE) for adapting large language models (LLMs) to multiple tasks still exhibit prevailing limitations: they either swap entire attention/feed-forward layers for switch experts or bolt on parallel expert branches, diluting parameter efficiency and task fidelity. We propose the LoRA-Mixer, a modular and lightweight MoE framework that integrates LoRA experts. Our core innovation lies in replacing the projection matrices of the attention module's input/output linear layers with dynamically routed, task-specific LoRA experts. This design ensures seamless compatibility with diverse foundation models, including transformers and state space models (SSMs), by leveraging their inherent linear projection structures. The framework supports two operational paradigms: (1) joint optimization of LoRA experts and routing mechanisms via a novel hard-soft routing strategy, or (2) direct deployment of pre-trained, frozen LoRA modules sourced from external repositories. To enable robust router training with limited data while ensuring stable routing decisions and maximizing expert reuse, we introduce an adaptive Specialization Balance Loss (SBL) that jointly optimizes expert balance and task-specific alignment. Extensive experiments on seven benchmark datasets, including MedQA, CoLA, SST-2, GSM8K, ARC-E, ARC-C, and HumanEval, demonstrate the effectiveness of LoRA-Mixer. On datasets such as GSM8K, HumanEval, and MedQA, LoRA-Mixer achieves significant improvements of 7.61%, 4.88%, and 3.08% over the base models, respectively. Compared with state-of-the-art methods, LoRA-Mixer achieves additional improvements of 1.09%, 1.45%, and 1.68%, respectively, using only 48% of the parameters, demonstrating its efficiency and strong performance.
We present Gazal-R1, a 32-billion-parameter language model that achieves state-of-the-art performance in medical reasoning while providing transparent, step-by-step explanations for clinical decision-making. Built upon Qwen3 32B, our model demonstrates that strategic training can enable mid-sized models to outperform significantly larger counterparts in specialized domains. We developed a novel two-stage training pipeline: first, supervised fine-tuning on a carefully curated dataset of 107,033 synthetic medical reasoning examples that teaches structured clinical thinking, enhanced by advanced parameter-efficient techniques including Weight-Decomposed Low-Rank Adaptation (DoRA) and Rank-Stabilized LoRA (rsLoRA); second, reinforcement learning using Group Relative Policy Optimization (GRPO) with a sophisticated multi-component reward system that refines accuracy, format adherence, and reasoning quality. Gazal-R1 achieves exceptional performance across medical benchmarks, scoring 87.1% on MedQA, 81.6% on MMLU Pro (Medical), and 79.6% on PubMedQA, surpassing models up to 12x larger. Beyond its strong empirical results, this work provides detailed insights into the challenges of training reasoning-capable models in specialized domains, including issues with reward hacking, training instability, and the fundamental tension between factual recall and detailed reasoning. Our methodology offers a reproducible framework for developing high-capability, domain-specific language models that balance performance, efficiency, and explainability.
Large language models have shown promise in clinical decision making, but current approaches struggle to localize and correct errors at specific steps of the reasoning process. This limitation is critical in medicine, where identifying and addressing reasoning errors is essential for accurate diagnosis and effective patient care. We introduce Med-PRM, a process reward modeling framework that leverages retrieval-augmented generation to verify each reasoning step against established medical knowledge bases. By verifying intermediate reasoning steps with evidence retrieved from clinical guidelines and literature, our model can precisely assess the reasoning quality in a fine-grained manner. Evaluations on five medical QA benchmarks and two open-ended diagnostic tasks demonstrate that Med-PRM achieves state-of-the-art performance, with improving the performance of base models by up to 13.50% using Med-PRM. Moreover, we demonstrate the generality of Med-PRM by integrating it in a plug-and-play fashion with strong policy models such as Meerkat, achieving over 80\% accuracy on MedQA for the first time using small-scale models of 8 billion parameters. Our code and data are available at: https://med-prm.github.io/
We introduce Biomed-Enriched, a biomedical text dataset constructed from PubMed via a two-stage annotation process. In the first stage, a large language model annotates 400K paragraphs from PubMed scientific articles, assigning scores for their type (review, study, clinical case, other), domain (clinical, biomedical, other), and educational quality. The educational quality score (rated 1 to 5) estimates how useful a paragraph is for college-level learning. These annotations are then used to fine-tune a small language model, which propagates the labels across the full PMC-OA corpus. The resulting metadata allows us to extract refined subsets, including 2M clinical case paragraphs with over 450K high-quality ones from articles with commercial-use licenses, and to construct several variants via quality filtering and domain upsampling. Clinical text is typically difficult to access due to privacy constraints, as hospital records cannot be publicly shared. Hence, our dataset provides an alternative large-scale, openly available collection of clinical cases from PubMed, making it a valuable resource for biomedical and clinical NLP. Preliminary continual-pretraining experiments with OLMo2 suggest these curated subsets enable targeted improvements, with clinical upsampling boosting performance by ~5% on MMLU ProfMed and educational quality filtering improving MedQA and MedMCQA by ~1%. Combinations of these techniques led to faster convergence, reaching same performance with a third of training tokens, indicating potential for more efficient and effective biomedical pretraining strategies.
Large reasoning models have recently made significant strides in mathematical and code reasoning, yet their success has not transferred smoothly to the medical domain. While multiple factors contribute to this disparity, a critical issue is the inadequate focus on the quality of intermediate reflection steps, which is particularly crucial in high-stakes medical scenarios. To address this challenge, we propose Med-REFL, a \underline{\textbf{Med}}ical \underline{\textbf{R}}easoning \underline{\textbf{E}}nhancement via self-corrected \underline{\textbf{F}}ine-grained ref\underline{\textbf{L}}ection. Our method leverages a tree-of-thought approach to decompose medical questions into fine-grained reasoning paths, quantitatively evaluating each step and its subsequent reflections. These assessments enable automatic construction of direct preference optimization data, reducing reliance on expensive expert annotations while guiding models to identify and correct reasoning errors. Experimental results on the MedQA-USMLE benchmark demonstrate Med-REFL achieves consistent improvements, with average gains up to 4.11\%. Notably, it further boosts the state-of-the-art performance of 7B/8B models by an additional 4.13\%. Furthermore, Med-REFL exhibits strong generalization capabilities and robustness across several challenging medical question-answering datasets. Our work illustrates that prioritizing reflection quality leads to more accurate and trustworthy reasoning in medical AI applications. Checkpoints, code, and data can be found in https://github.com/TianYin123/Med-REFL.
Large language models (LLMs) have shown great potential in medical question answering (MedQA), yet adapting them to biomedical reasoning remains challenging due to domain-specific complexity and limited supervision. In this work, we study how prompt design and lightweight fine-tuning affect the performance of open-source LLMs on PubMedQA, a benchmark for multiple-choice biomedical questions. We focus on two widely used prompting strategies - standard instruction prompts and Chain-of-Thought (CoT) prompts - and apply QLoRA for parameter-efficient instruction tuning. Across multiple model families and sizes, our experiments show that CoT prompting alone can improve reasoning in zero-shot settings, while instruction tuning significantly boosts accuracy. However, fine-tuning on CoT prompts does not universally enhance performance and may even degrade it for certain larger models. These findings suggest that reasoning-aware prompts are useful, but their benefits are model- and scale-dependent. Our study offers practical insights into combining prompt engineering with efficient finetuning for medical QA applications.
Large Language Models (LLMs) have significant impact on the healthcare scenarios but remain prohibitively large for deployment in real-time, resource-constrained environments such as edge devices. In this work, we introduce a novel medical assistant system, optimized through our general-purpose compression framework, which tailors Large Language Models (LLMs) for deployment in specialized domains. By measuring neuron saliency on domain-specific data, our method can aggressively prune irrelevant neurons, reducing model size while preserving performance. Following pruning, we apply post-training quantization to further reduce the memory footprint, and evaluate the compressed model across medical benchmarks including MedMCQA, MedQA, and PubMedQA. We also deploy the 50\% compressed Gemma and the 67\% compressed LLaMA3 models on Jetson Orin Nano (18.7W peak) and Raspberry Pi 5 (6.3W peak), achieving real-time, energy-efficient inference under hardware constraints.
Current Large Language Models (LLMs) exhibit significant limitations, notably in structured, interpretable, and verifiable medical reasoning, alongside practical deployment challenges related to computational resources and data privacy. This report focused on the development of WiNGPT-3.0, the 32-billion parameter LLMs, engineered with the objective of enhancing its capacity for medical reasoning and exploring its potential for effective integration within healthcare IT infrastructures. The broader aim is to advance towards clinically applicable models. The approach involved a multi-stage training pipeline tailored for general, medical, and clinical reasoning. This pipeline incorporated supervised fine-tuning (SFT) and reinforcement learning (RL), leveraging curated Long Chain-of-Thought (CoT) datasets, auxiliary reward models, and an evidence-based diagnostic chain simulation. WiNGPT-3.0 demonstrated strong performance: specific model variants achieved scores of 66.6 on MedCalc and 87.1 on MedQA-USMLE. Furthermore, targeted training improved performance on a clinical reasoning task from a baseline score of 58.1 to 62.5. These findings suggest that reinforcement learning, even when applied with a limited dataset of only a few thousand examples, can enhance medical reasoning accuracy. Crucially, this demonstration of RL's efficacy with limited data and computation paves the way for more trustworthy and practically deployable LLMs within clinical workflows and health information infrastructures.
Despite the growing clinical adoption of large language models (LLMs), current approaches heavily rely on single model architectures. To overcome risks of obsolescence and rigid dependence on single model systems, we present a novel framework, termed the Consensus Mechanism. Mimicking clinical triage and multidisciplinary clinical decision-making, the Consensus Mechanism implements an ensemble of specialized medical expert agents enabling improved clinical decision making while maintaining robust adaptability. This architecture enables the Consensus Mechanism to be optimized for cost, latency, or performance, purely based on its interior model configuration. To rigorously evaluate the Consensus Mechanism, we employed three medical evaluation benchmarks: MedMCQA, MedQA, and MedXpertQA Text, and the differential diagnosis dataset, DDX+. On MedXpertQA, the Consensus Mechanism achieved an accuracy of 61.0% compared to 53.5% and 45.9% for OpenAI's O3 and Google's Gemini 2.5 Pro. Improvement was consistent across benchmarks with an increase in accuracy on MedQA ($\Delta\mathrm{Accuracy}_{\mathrm{consensus\text{-}O3}} = 3.4\%$) and MedMCQA ($\Delta\mathrm{Accuracy}_{\mathrm{consensus\text{-}O3}} = 9.1\%$). These accuracy gains extended to differential diagnosis generation, where our system demonstrated improved recall and precision (F1$_\mathrm{consensus}$ = 0.326 vs. F1$_{\mathrm{O3\text{-}high}}$ = 0.2886) and a higher top-1 accuracy for DDX (Top1$_\mathrm{consensus}$ = 52.0% vs. Top1$_{\mathrm{O3\text{-}high}}$ = 45.2%).
Reinforcement learning (RL) is a powerful way to adapt foundation models to specialized tasks, but its reliance on large-scale human-labeled data limits broad adoption. We introduce Synthetic Data RL, a simple and general framework that reinforcement fine-tunes models using only synthetic data generated from a task definition. Our method first generates question and answer pairs from the task definition and retrieved documents, then adapts the difficulty of the question based on model solvability, and selects questions using the average pass rate of the model across samples for RL training. On Qwen-2.5-7B, our method achieves a 29.2% absolute improvement over the base model on GSM8K (+2.9 pp vs. instruction-tuned, +6.6 pp vs. Self-Instruct), 8.7% on MATH, 13.1% on GPQA (+7.0 pp vs. SynthLLM), 8.9% on MedQA, 17.7% on CQA (law) and 13.7% on CFA (finance). It surpasses supervised fine-tuning under the same data budget and nearly matches RL with full human data across datasets (e.g., +17.2 pp on GSM8K). Adding 100 human demonstrations improves the performance of GSM8K only by 0.4 pp, showing a limited added value. By reducing human data annotation, Synthetic Data RL enables scalable and efficient RL-based model adaptation. Code and demos are available at https://github.com/gydpku/Data_Synthesis_RL/.
Recent advances such as Chain-of-Thought prompting have significantly improved large language models (LLMs) in zero-shot medical reasoning. However, prompting-based methods often remain shallow and unstable, while fine-tuned medical LLMs suffer from poor generalization under distribution shifts and limited adaptability to unseen clinical scenarios. To address these limitations, we present TAGS, a test-time framework that combines a broadly capable generalist with a domain-specific specialist to offer complementary perspectives without any model fine-tuning or parameter updates. To support this generalist-specialist reasoning process, we introduce two auxiliary modules: a hierarchical retrieval mechanism that provides multi-scale exemplars by selecting examples based on both semantic and rationale-level similarity, and a reliability scorer that evaluates reasoning consistency to guide final answer aggregation. TAGS achieves strong performance across nine MedQA benchmarks, boosting GPT-4o accuracy by 13.8%, DeepSeek-R1 by 16.8%, and improving a vanilla 7B model from 14.1% to 23.9%. These results surpass several fine-tuned medical LLMs, without any parameter updates. The code will be available at https://github.com/JianghaoWu/TAGS.
Doctors and patients alike increasingly use Large Language Models (LLMs) to diagnose clinical cases. However, unlike domains such as math or coding, where correctness can be objectively defined by the final answer, medical diagnosis requires both the outcome and the reasoning process to be accurate. Currently, widely used medical benchmarks like MedQA and MMLU assess only accuracy in the final answer, overlooking the quality and faithfulness of the clinical reasoning process. To address this limitation, we introduce MedCaseReasoning, the first open-access dataset for evaluating LLMs on their ability to align with clinician-authored diagnostic reasoning. The dataset includes 14,489 diagnostic question-and-answer cases, each paired with detailed reasoning statements derived from open-access medical case reports. We evaluate state-of-the-art reasoning LLMs on MedCaseReasoning and find significant shortcomings in their diagnoses and reasoning: for instance, the top-performing open-source model, DeepSeek-R1, achieves only 48% 10-shot diagnostic accuracy and mentions only 64% of the clinician reasoning statements (recall). However, we demonstrate that fine-tuning LLMs on the reasoning traces derived from MedCaseReasoning significantly improves diagnostic accuracy and clinical reasoning recall by an average relative gain of 29% and 41%, respectively. The open-source dataset, code, and models are available at https://github.com/kevinwu23/Stanford-MedCaseReasoning.
Medical reasoning in large language models (LLMs) aims to emulate clinicians' diagnostic thinking, but current benchmarks such as MedQA-USMLE, MedMCQA, and PubMedQA often mix reasoning with factual recall. We address this by separating 11 biomedical QA benchmarks into reasoning- and knowledge-focused subsets using a PubMedBERT classifier that reaches 81 percent accuracy, comparable to human performance. Our analysis shows that only 32.8 percent of questions require complex reasoning. We evaluate biomedical models (HuatuoGPT-o1, MedReason, m1) and general-domain models (DeepSeek-R1, o4-mini, Qwen3), finding consistent gaps between knowledge and reasoning performance. For example, m1 scores 60.5 on knowledge but only 47.1 on reasoning. In adversarial tests where models are misled with incorrect initial reasoning, biomedical models degrade sharply, while larger or RL-trained general models show more robustness. To address this, we train BioMed-R1 using fine-tuning and reinforcement learning on reasoning-heavy examples. It achieves the strongest performance among similarly sized models. Further gains may come from incorporating clinical case reports and training with adversarial and backtracking scenarios.
In this paper, we introduce S-MedQA, an English medical question-answering (QA) dataset for benchmarking large language models in fine-grained clinical specialties. We use S-MedQA to check the applicability of a popular hypothesis related to knowledge injection in the knowledge-intense scenario of medical QA, and show that: 1) training on data from a speciality does not necessarily lead to best performance on that specialty and 2) regardless of the specialty fine-tuned on, token probabilities of clinically relevant terms for all specialties increase consistently. Thus, we believe improvement gains come mostly from domain shifting (e.g., general to medical) rather than knowledge injection and suggest rethinking the role of fine-tuning data in the medical domain. We release S-MedQA and all code needed to reproduce all our experiments to the research community.
Sanitizing sensitive text data typically involves removing personally identifiable information (PII) or generating synthetic data under the assumption that these methods adequately protect privacy; however, their effectiveness is often only assessed by measuring the leakage of explicit identifiers but ignoring nuanced textual markers that can lead to re-identification. We challenge the above illusion of privacy by proposing a new framework that evaluates re-identification attacks to quantify individual privacy risks upon data release. Our approach shows that seemingly innocuous auxiliary information -- such as routine social activities -- can be used to infer sensitive attributes like age or substance use history from sanitized data. For instance, we demonstrate that Azure's commercial PII removal tool fails to protect 74\% of information in the MedQA dataset. Although differential privacy mitigates these risks to some extent, it significantly reduces the utility of the sanitized text for downstream tasks. Our findings indicate that current sanitization techniques offer a \textit{false sense of privacy}, highlighting the need for more robust methods that protect against semantic-level information leakage.
Small Language Models (SLMs) have potential to be used for automatically labelling and identifying aspects of text data for medicine/health-related purposes from documents and the web. As their resource requirements are significantly lower than Large Language Models (LLMs), these can be deployed potentially on more types of devices. SLMs often are benchmarked on health/medicine-related tasks, such as MedQA, although performance on these can vary especially depending on the size of the model in terms of number of parameters. Furthermore, these test results may not necessarily reflect real-world performance regarding the automatic labelling or identification of texts in documents and the web. As a result, we compared topic-relatedness scores from Microsofts phi-3-mini-4k-instruct SLM to the topic-relatedness scores from 7 human evaluators on 1144 samples of medical/health-related texts and 1117 samples of sports injury-related texts. These texts were from a larger dataset of about 9 million news headlines, each of which were processed and assigned scores by phi-3-mini-4k-instruct. Our sample was selected (filtered) based on 1 (low filtering) or more (high filtering) Boolean conditions on the phi-3 SLM scores. We found low-moderate significant correlations between the scores from the SLM and human evaluators for sports injury texts with low filtering (\r{ho} = 0.3413, p < 0.001) and medicine/health texts with high filtering (\r{ho} = 0.3854, p < 0.001), and low significant correlation for medicine/health texts with low filtering (\r{ho} = 0.2255, p < 0.001). There was negligible, insignificant correlation for sports injury-related texts with high filtering (\r{ho} = 0.0318, p = 0.4466).
Large language models (LLMs) hold great promise for assisting clinical interviews due to their fluent interactive capabilities and extensive medical knowledge. However, the lack of high-quality interview dialogue data and widely accepted evaluation methods has significantly impeded this process. So we propose CliniChat, a framework that integrates multi-source knowledge to enable LLMs to simulate real-world clinical interviews. It consists of two modules: Clini-Recon and Clini-Eval, each responsible for reconstructing and evaluating interview dialogues, respectively. By incorporating three sources of knowledge, Clini-Recon transforms clinical notes into systematic, professional, and empathetic interview dialogues. Clini-Eval combines a comprehensive evaluation metric system with a two-phase automatic evaluation approach, enabling LLMs to assess interview performance like experts. We contribute MedQA-Dialog, a high-quality synthetic interview dialogue dataset, and CliniChatGLM, a model specialized for clinical interviews. Experimental results demonstrate that CliniChatGLM's interview capabilities undergo a comprehensive upgrade, particularly in history-taking, achieving state-of-the-art performance.
Large language models (LLMs) are increasingly used in healthcare, but their reliability is heavily influenced by user-driven factors such as question phrasing and the completeness of clinical information. In this study, we examined how misinformation framing, source authority, model persona, and omission of key clinical details affect the diagnostic accuracy and reliability of LLM outputs. We conducted two experiments: one introducing misleading external opinions with varying assertiveness (perturbation test), and another removing specific categories of patient information (ablation test). Using public datasets (MedQA and Medbullets), we evaluated proprietary models (GPT-4o, Claude 3.5 Sonnet, Claude 3.5 Haiku, Gemini 1.5 Pro, Gemini 1.5 Flash) and open-source models (LLaMA 3 8B, LLaMA 3 Med42 8B, DeepSeek R1 8B). All models were vulnerable to user-driven misinformation, with proprietary models especially affected by definitive and authoritative language. Assertive tone had the greatest negative impact on accuracy. In the ablation test, omitting physical exam findings and lab results caused the most significant performance drop. Although proprietary models had higher baseline accuracy, their performance declined sharply under misinformation. These results highlight the need for well-structured prompts and complete clinical context. Users should avoid authoritative framing of misinformation and provide full clinical details, especially for complex cases.
Medical Question Answering systems based on Retrieval Augmented Generation is promising for clinical decision support because they can integrate external knowledge, thus reducing inaccuracies inherent in standalone large language models (LLMs). However, these systems may unintentionally propagate or amplify biases associated with sensitive demographic attributes like race, gender, and socioeconomic factors. This study systematically evaluates demographic biases within medical RAG pipelines across multiple QA benchmarks, including MedQA, MedMCQA, MMLU, and EquityMedQA. We quantify disparities in retrieval consistency and answer correctness by generating and analyzing queries sensitive to demographic variations. We further implement and compare several bias mitigation strategies to address identified biases, including Chain of Thought reasoning, Counterfactual filtering, Adversarial prompt refinement, and Majority Vote aggregation. Experimental results reveal significant demographic disparities, highlighting that Majority Vote aggregation notably improves accuracy and fairness metrics. Our findings underscore the critical need for explicitly fairness-aware retrieval methods and prompt engineering strategies to develop truly equitable medical QA systems.
Large Language Models (LLMs) have made significant progress in various fields. However, challenges remain in Multi-Disciplinary Team (MDT) medical consultations. Current research enhances reasoning through role assignment, task decomposition, and accumulation of medical experience. Multi-role collaboration in MDT consultations often results in excessively long dialogue histories. This increases the model's cognitive burden and degrades both efficiency and accuracy. Some methods only store treatment histories. They do not extract effective experience or reflect on errors. This limits knowledge generalization and system evolution. We propose a multi-agent MDT medical consultation framework based on LLMs to address these issues. Our framework uses consensus aggregation and a residual discussion structure for multi-round consultations. It also employs a Correct Answer Knowledge Base (CorrectKB) and a Chain-of-Thought Knowledge Base (ChainKB) to accumulate consultation experience. These mechanisms enable the framework to evolve and continually improve diagnosis rationality and accuracy. Experimental results on the MedQA and PubMedQA datasets demonstrate that our framework achieves accuracies of 90.1% and 83.9%, respectively, and that the constructed knowledge bases generalize effectively across test sets from both datasets.