Skip to content

🩺 Medical LLM

💬 EMNLP2026 · 47 paper notes

📌 Same area in other venues: 🧠 NeurIPS2026 (16) · 📷 CVPR2026 (1) · 🔬 ICLR2026 (20) · 💬 ACL2026 (47) · 🧪 ICML2026 (4) · 🤖 AAAI2026 (12)

🔥 Top topics: Medical Imaging ×33 · Reasoning ×9 · LLM ×8 · Dialogue ×4 · Alignment/RLHF ×3

AP-GRPO: Anchor-Gated Phonetic Alignment with Policy Optimization for Pathological Speech Reconstruction

Addressing acoustic degradation and severe hallucinations when speech language models transcribe dysarthric recordings, AP-GRPO introduces a group relative policy optimization framework that preserves reliable audible anchor words and scores corrupted inter-anchor spans via soft-DTW phonetic alignment, slashing word error rate in severe ALS from 0.75 to 0.29.

auditing evidence use in medical llm diagnosis

Addressing the limitation where diagnostic accuracy hides improper clinical reasoning, this paper presents a behavioral evidence-use audit via controlled subset interventions and diagnostic margin interaction mining, establishing a diagnosis-relative framework with multi-stage stability filtering to isolate true evidence failures.

Can MLLMs Perceive and Describe Medical Image Quality Beyond Scores?

MedQ-Bench establishes the first comprehensive multimodal benchmark for language-based medical image quality assessment across 5 modalities, 40+ degradation types, and 3,308 images. Evaluating 23 frontier MLLMs reveals substantial gaps against human experts, identifying severe performance drops in mild degradations and highlighting a pervasive diagnostic bias where medical-specialized models overlook catastrophic imaging artifacts.

PsyCLIENT: Client Simulation via Conversational Trajectory Modeling for Both Trainee Practice and Model Evaluation in Mental Health Counseling

PsyCLIENT decouples static client profiles from dynamic conversational trajectories, conditioning LLMs on 12 discrete client behavior labels and de-identified reference utterances alongside coherence soft constraints, achieving higher expert-rated authenticity and clinical training utility across 120 detailed Chinese client profiles spanning 60 counseling topics.

ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning

ClinHallu introduces a benchmark of 7,031 validated multimodal medical VQA instances augmented with structured 3-stage traces (Visual Recognition, Knowledge Recall, Reasoning Integration) and stage-replacement counterfactual interventions, pinpointing task-specific cognitive bottlenecks and demonstrating that full trace supervision significantly mitigates stage-wise hallucinations.

Cognitive-Level Prefix-MoE Reward Modeling for Curriculum Alignment of Medical Language Models

To resolve the inherent reward conflict between exam-style multiple-choice questions and complex open-ended clinical reasoning, this paper proposes Prefix-MoE RM—a cognitive-level reward model that softly fuses trainable prefix embeddings across Bloom-inspired levels via a frozen router before scalar scoring—paired with prompt-type dispatch and four-stage curriculum GRPO.

Criterion-Activated Preference Synthesis: Attributed Alignment for Medical Dialogue Agents

This paper introduces Criterion-Activated Preference Synthesis (CAPS), a criterion-first preference synthesis framework that seeds context generation with explicit quality standards, decomposes evaluations into rubric-bound and general blind judges, and applies group-level Empirical Activation Tests (EAT) and pair-level Attribution Ratio (AR) filters to isolate clean, causal preference signals—outperforming a 22k SFT baseline and matching 35k human preference pairs on HealthBench with only 2k synthesized pairs.

CRU-Reason: Knowledge Graph-Grounded and Interpretable Clinical Prediction with Clinical Rationale Units

CRU-Reason integrates machine learning priors and constrained knowledge graph exploration to construct disease-centered Clinical Rationale Units (CRUs), guiding large language models to conduct fine-grained, decoupled clinical reasoning and delivering a 9.4% average improvement over strong baselines in drug recommendation and diagnosis prediction.

CRU-Reason: Knowledge Graph-Grounded and Interpretable Clinical Prediction with Clinical Rationale Units

CRU-Reason synergizes machine learning sequential statistical priors, medical knowledge graphs, and large language models by decomposing complex electronic health records into disease-centric Clinical Rationale Units (CRUs) via constrained graph exploration, driving fine-grained evidence-grounded clinical reasoning and outperforming top EHR baselines by an average of 9.4% on MIMIC-III and MIMIC-IV.

Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis

To address the challenge of reasoning over complex, non-monotonic disease progression trajectories under sparse delayed rewards in unstructured clinical notes, this paper presents Dementia-R1, a two-stage reinforcement learning framework that cold-starts via GRPO on verifiable intermediate clinical cognitive indices and fine-tunes on binary prognosis, achieving 84.02% AUROC on real-world clinical notes and outperforming models up to 10× larger.

Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis

Dementia-R1 introduces a two-stage reinforcement learning framework that leverages Cold-Start RL with tolerance-aware verifiable clinical rewards (MMSE, GDS, CDR) to master intermediate disease trajectory forecasting before task fine-tuning, achieving 84.02% AUROC on real-world unstructured notes and outperforming models up to 10× larger.

Dense Process Supervision for Search Agents via Fact Utility Estimation

Addressing the credit assignment challenge in RL-trained search agents caused by sparse outcome rewards and implicit intermediate contributions, FactAgent represents accumulated evidence within an explicit fact store, infers the posterior utility of semantic fact clusters from group rollouts via Bayesian estimation, and applies potential-based reward shaping to deliver dense step-level process supervision.

Dental-TriageBench: Benchmarking Multimodal Reasoning for Hierarchical Dental Triage

Dental-TriageBench introduces the first expert-annotated multimodal benchmark for hierarchical dental triage from paired chief complaints and orthopantomograms (OPGs). Benchmarking 21 MLLMs against three junior dentists demonstrates that frontier models substantially trail human clinicians on fine-grained referral planning, exhibiting severe omission-heavy under-coverage on complex multi-specialty cases.

Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards

Fathom-Vaidya introduces a sequential reinforcement learning and rubric-generation framework that pairs rule/error-rubric RL on diagnostic questions with actor-critic synthesis of 5.3k multi-turn clinical dialogues and multi-dimensional continuous rubrics under GSPO, boosting 30B performance on HealthBench-Hard to 50.1% and outperforming GPT-5 Thinking.

FDARxBench: Benchmarking regulatory and clinical reasoning on FDA generic drug assessment

Developed in direct partnership with FDA regulatory assessors, FDARxBench is the first regulatory-grade benchmark (17,223 QA pairs spanning factual, cross-section multi-hop, and negative-control refusal tasks) grounded in structured prescription drug labeling (SPL), exposing severe limitations in frontier LLMs regarding long-context evidence localization, multi-hop synthesis, and conservative abstention.

GPAgentBench-2K: Benchmarking Large Language Model Agents in Complex Clinical Action Space

Addressing the oversimplified action spaces and unconstrained objectives in current clinical agent benchmarks, this paper introduces GPAgentBench-2K, the first Constrained Markov Decision Process (CMDP) benchmark for primary care featuring a 6-action clinical space, action masking priors, and safety-abstention referral outcomes, exposing a critical quality-safety gap across frontier LLMs and establishing a reference baseline via C-GRPO.

Graph2Counsel: Clinically Grounded Synthetic Counseling Dialogue Generation from Client Psychological Graphs

Graph2Counsel generates clinically grounded synthetic counseling dialogues by conditioning LLMs on Client Psychological Graphs (CPGs) and real counselor intervention strategies, producing high-fidelity multi-turn therapeutic sessions that outperform prior synthetic datasets across expert clinical evaluations and downstream benchmarks.

Graph2Counsel: Clinically Grounded Synthetic Counseling Dialogue Generation from Clinical Graphs

Graph2Counsel introduces a clinically grounded synthetic counseling dialogue generation framework anchored in Client Psychological Graphs (CPGs) and extracted therapist strategies, producing realistic multi-turn sessions that outperform existing synthetic benchmarks across expert evaluations and downstream counseling tasks.

How Well Can Modern LLMs Act as Agent Cores in Radiology Environments?

This paper introduces RadA-BenchPlat, a two-layer benchmark decoupling idealized reasoning from real-world execution across 24.2k synthetic QA pairs and 165 real-world multimodal cases, revealing that modern LLM agent cores face a 15.5% performance gap and a drop to 0.327 on complex workflows due to execution fidelity drift.

Import What You Need: Learning When and How to Augment Domain Knowledge in Clinical LLMs

Addressing the noise dilution and computational bloat caused by indiscriminate external knowledge graph expansion in longitudinal EHR prediction, this paper introduces ReTA—a reinforcement learning-based routing framework that dynamically selects between semantic Soft Import, topological Hard Import, or adaptive Skip over an offline-distilled template pool, harmonizing high predictive accuracy with low inference latency via decoupled dual-channel encoding.

Investigating Linear Probe Robustness to Linguistic Register, Medical Specialty, and Corpus Shifts in Medical QA

The paper isolates linguistic register, medical specialty, and corpus shifts in medical QA linear probing, demonstrating that the truth direction is robust to surface style (\(\Delta_{\text{register}} \approx -0.095\) AUROC) and specialty (\(\Delta_{\text{specialty}} \approx -0.031\)) but degrades sharply under corpus transfer (\(\Delta_{\text{dataset}} \approx -0.21\)).

KnowSim: Evaluating Information Calibration in LLM Assistants with User Simulators that Learn

Introduces KNOWSIM, an evaluation framework powered by a user simulator that models dynamic knowledge state transitions over an Information Unit (IU) prerequisite DAG with cognitive load constraints, providing explainable multi-turn information calibration benchmarks.

Learning Personalized Prompts for Healthcare Guidance

Proposes PPL (Personalized Prompt Learning), an end-to-end framework that constructs coarse prompts from a patient's historical medical checkups and peer-informed similar cases, and refines them into discrete patient-tailored hard prompts via reinforcement learning with physician recommendations as reward signals, delivering superior personalized healthcare advice with proprietary LLMs.

Learning Personalized Prompts for Healthcare Guidance

This paper introduces Personalized Prompt Learning (PPL), a framework that integrates longitudinal self-informed medical history with peer-informed collaborative clinical cases into tailored prompts, subsequently refining them via reinforcement learning against real physician recommendations to steer proprietary black-box LLMs toward high-quality personalized healthcare guidance without model fine-tuning.

Lung-R1: A Knowledge Graph-Guided LLM for Pulmonary Diagnostic Reasoning

To bridge the Pulmonary Knowledge-to-Diagnosis Gap, this paper introduces LungKG (59,038 nodes, 164,308 directed edges) and develops Lung-R1 via KG-constrained Chain-of-Thought fine-tuning and triple-consistency reinforcement learning, scoring 4.3583 on authentic EMR diagnosis and surpassing Claude-Sonnet-4.5 by 0.1476 points.

Making Clinical Language Models Auditable: Concept-Guided Fine-Tuning for Robust Prediction

Addressing the critical flaw of clinical language models exploiting documentation artifacts and note templates as predictive shortcuts, this paper introduces CAST (Concept-guided Artifact Suppression Tuning), which leverages Sparse Autoencoders to expose internal features, rigorously filters non-clinical artifacts via LLM consensus with ICD-10 retrieval safeguards, suppresses artifact latents via residual subtraction during fine-tuning, and delivers a transparent first-order gradient attribution trail for patient-level auditability.

MedF^2Route: Evidence-Routed Graph Verification for Hallucination Detection in Clinical LVLMs

Addressing the intertwined nature of faithfulness and factuality hallucinations and evidence misalignment in clinical LVLMs, MedF²Route structures clinical contexts into graphs, dynamically routes response triplets to heterogeneous patient and knowledge evidence sources for dual-path verification, and achieves substantial gains in detection precision and overall Macro-F1 across diverse backbones.

MedF2Route: Evidence-Routed Graph Verification for Hallucination Detection in Clinical LVLMs

MedF2Route casts black-box clinical LVLM hallucination detection as an evidence routing problem, converting free-text case queries and generated responses into structured clinical graphs, dynamically dispatching response triplets across patient history, medical imaging, and external knowledge graphs, and achieving substantial precision and F1 gains while eliminating false alarms.

Medical Bilingualism: Cross-Ontology Graph Retrieval for Incommensurable Knowledge Bases

Addressing semantic collapse caused by forcing incommensurable medical paradigms (Traditional Chinese Medicine and Western Medicine) into a shared embedding space, XONTO-RAG constructs a dual-ontology graph with verified typed bridging relations and applies a query-aware RGAT retriever to assemble provenance-grounded multi-hop evidence subgraphs.

MedInsightBench: Evaluating Medical Analytics Agents on Multi-Modal Clinical Records

This paper introduces MedInsightBench, the first benchmark evaluating medical analytics agents on goal-driven, multi-step insight discovery from paired whole-slide pathology images and clinical reports, and proposes MedInsightAgent, an agentic framework that synergizes visual root finding, specialized pathology tool inspection, and iterative question composition to substantially surpass baseline multi-modal models and generic agents.

MedPriv-Bench: Benchmarking the Privacy-Utility Trade-off of Large Language Models in Healthcare

MedPriv-Bench is the first benchmark designed to jointly evaluate privacy preservation and clinical utility in medical open-ended question answering, establishing a multi-agent synthetic pipeline and a fine-tuned RoBERTa-NLI judge to reveal pervasive contextual leakage risks across nine major LLMs and demonstrating the superiority of supervised fine-tuning and brevity constraints.

MedRoundsQA: A Persona and Difficulty Aware Evaluation for Multi-Turn Medical Consultations

This paper introduces MedRoundsQA, a multi-turn diagnostic consultation benchmark derived from 1,387 board-exam vignettes normalized into 24-slot clinical records, demonstrating that interactive diagnosis causes a severe 13–39 point drop in accuracy across 15 LLMs, dialogue benefits plateau after 6–12 turns, and health-literacy disparities shift diagnosis accuracy by 7–8 points under identical underlying clinical facts.

MedSNIP: Building and Benchmarking Snippet-Level Granularity for Medical Fact Verification

Addressing the issue where traditional atomic claim decomposition shatters clinical context and causal reasoning, this paper reformulates medical fact-checking around snippet-level granularity, introducing the expert-annotated MEDSNIP-BENCH and the automated MEDSNIP pipeline to significantly improve false-class F1 while reducing verifier calls.

Meerkat2: Bridging Global Reasoning and Local Knowledge for Efficient Medical AI

Meerkat2 tackles the dual challenges of missing local legal-institutional knowledge and excessive generation verbosity in clinical LLMs by coupling source-grounded statute QA synthesis with Verbosity-Decoupled Distillation (VDD), allowing a 30B-A3B MoE model to match 235B frontier accuracy on the Korean Medical Licensing Examination with only 3B active parameters while slashing token inference cost by 54%–65%.

MMedAgentBench: A Multi-Stage Benchmark for Evaluating LLM-based Medical Agents

MMedAgentBench introduces a realistic multi-stage benchmark comprising 167 de-identified tertiary hospital patient journeys across 527 sequential clinical stages, evaluating LLM agents on diagnostic accuracy, therapeutic appropriateness, and resource cost control under progressive information disclosure and expert key-point gating.

omniv med a scalable training recipe for unified medical vision language underst

OmniV-Med addresses the fragmentation of heterogeneous medical visual inputs by presenting a scalable training recipe that unifies 2D images, 3D volumes, and surgical videos using a single 2D ViT with 2D RoPE, Redundancy-Guided Token Pruning (RGTP) cutting 60% of visual tokens, and a three-stage progressive curriculum, outperforming domain-specific baselines across nine clinical benchmarks.

OntologyBench: Can Dense Retrieval Satisfy Structured Biomedical Constraints?

Constructs OntologyBench, a 597k-pair tiered benchmark across concept grounding, relational retrieval, and compositional phenotype constraints, revealing that dense retrievers suffer severe aggregation failures on joint constraints (scoring only 0.313 nDCG@10 vs. 0.739 for an ontology-aware baseline), which downstream late-interaction or LLM rerankers fail to resolve.

Patient-Zero: Scaling Synthetic Patient Agents to Real-World Distributions without Real Patient Data

Patient-Zero presents an ab initio patient simulation framework that generates clinically realistic and diverse patient agents entirely from medical knowledge without touching real patient data, resolving the privacy-diversity bottleneck and overcoming the Stability-Plasticity Dilemma in dynamic medical dialogue through a Dual-Track Cognitive Memory System with real-time NLI verification.

Patient-Zero: Scaling Synthetic Patient Agents to Real-World Distributions without Real Patient Data

Patient-Zero establishes an ab initio synthetic patient simulation framework that generates clinically grounded, multi-specialty patient records without using any real clinical data, resolving the Stability-Plasticity Dilemma in extended diagnostic consultations via a Dual-Track Cognitive Memory System with NLI verification, achieving over 98% clinical validity and boosting downstream clinical reasoning models by up to +24.0% on MedQA.

RASPER: Reward-Aligned Summarization of Clinical Notes for EHR Outcome Prediction

This paper proposes RASPER, a reward-aligned clinical note summarization framework for EHR outcome prediction that integrates longitudinal structured codes into soft prompts and trains a tunable LLM summarizer via reinforcement learning from downstream predictor feedback (RLPF), preserving patient-specific textual evidence that effectively complements structured codes.

right diagnoses decorative reasoning a perturbation audit of medical chain of th

Addressing the clinical risk where Chain-of-Thought (CoT) rationales are scrutinized as reasoning evidence, this paper proposes a 30-operator clinical perturbation audit paired with a chain-update \(\times\) answer-flip joint analysis, revealing that medical CoT is predominantly decorative (panel-wide Chain-Decoupling Rate reaches 72.9%).

SafeRx-Agent: A Knowledge-Grounded Multi-Agent Framework for Safe and Explainable Medication Recommendation

SafeRx-Agent introduces the first fine-grained ATC-L4 medication recommendation formulation and combines sparse specialty routing, indication evidence grounding, global critique reconciliation, and a patient-conditioned safety verifier to outperform traditional deep learning and LLM baselines while drastically suppressing drug interactions and contraindications.

SAGE: Semantic Anchor-Guided Evolution for Grounded Medical QA Data Synthesis

SAGE leverages the lightweight, authoritative MeSH taxonomy as semantic anchors, executing a dual-track parallel iterative evolution interleaving atomic single-concept grounding and associative multi-concept reasoning to enable small local models to synthesize high-quality medical QA data without large corpora or cloud APIs, achieving a 68.12% average accuracy across five clinical benchmarks.

SARL-ICD: Structure-aware Contrastive Representation Learning with LLM-augmented Label Descriptions for Automatic ICD Coding

SARL-ICD mitigates structural shortcut learning in semi-structured clinical discharge summaries and enriches medical code semantics through Jaccard-guided narrative-truncated MoCo pretraining, a lightweight LLM-augmented label-text interaction module, and an adaptive MoE fusion network, establishing new state-of-the-art performance across MIMIC-III-full and MIMIC-IV.

SARL-ICD: Structure-aware Contrastive Representation Learning with LLM-augmented Label Descriptions for Automatic ICD Coding

To tackle representation shortcut learning caused by templated semi-structured discharge summaries and the underutilization of rich label semantics, SARL-ICD introduces narrative-truncated Jaccard contrastive pretraining (JMoCo), lightweight LLM-augmented label-text interaction, and adaptive MoE fusion across complementary coding backbones, establishing new state-of-the-art results across MIMIC-III and MIMIC-IV.

SoftGene: Protein Language Model-Enhanced Soft Prompt Tuning for Genomic Sequence Modeling

SoftGene introduces a hybrid prompting framework that hierarchically encodes protein amino acid sequences and alternative splice isoforms via ESM-2 into continuous soft prompt tokens, conditioning a local LLM alongside auxiliary biological context to generate accurate and interpretable gene set annotations.

Solve the Missing First Step: Can VLMs Standardize Uncalibrated Clinical Images?

Presents MDS-Bench, the first comprehensive benchmark encompassing 100 diverse medical datasets and 1,939 verified target samples to assess agentic VLMs in discovering, converting, and standardizing raw heterogeneous clinical files into VLM-ready image-text atomic units, showing that even top frontier models like Gemini 3 Flash achieve an end-to-end strict success rate of only 48.6%.