Research map / Vision-language and multimodal / Multimodal LLMs

LLaDA-MedV: Exploring Large Language Diffusion Models for Biomedical Image Understanding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces LLaDA-MedV, a masked-diffusion large language model adapted to biomedical image understanding via multi-stage visual instruction tuning, compared against autoregressive biomedical VLMs.

Problem

Biomedical vision-language models are dominated by autoregressive models, and the potential of masked diffusion language models for biomedical image understanding is unexplored; ARMs also offer limited control over response length.

Approach

Applies LLaDA-8B-Instruct as a masked-diffusion language backbone with a SigLIP2 vision encoder and MLP projector, trained with a masked-response prediction objective conditioned on image and prompt; response length is controlled by the fixed generation length and the number of sampling steps.

Key results

• Overall open-ended conversation score of 52.605 versus 44.750 for LLaVA-Med (reported as 7.855% relative gain).
• Closed-form accuracy of 84.93% on VQA-RAD, 92.31% on SLAKE, and 95.15% on PathVQA, reported as state of the art on these closed-form subsets.
• Average response length of about 166 words per answer versus about 36 for LLaVA-Med, with inference time per word rising from 0.036s to 0.230s.

Surprising findings

• Initializing from LLaDA-V, a general image/video model, gave worse biomedical performance and more repetitive output than the chosen initialization.
• Open-form VQA performance is weak (45.60 on VQA-RAD, 31.96 on PathVQA), which the authors attribute to limited post-training of the LLaDA backbone.

Limitations

• Token repetition during generation, worse with long responses and few sampling steps.
• Inference cost is much higher than autoregressive LLaVA-Med per word.
• Open-form VQA lags baselines because the backbone lacks post-training and does not follow answer-format constraints.

Experimental setup

Datasets: 600K biomedical alignment image-text pairs 60K multi-turn biomedical dialogue dataset VQA-RAD SLAKE PathVQA Biomedical Visual Chatbot benchmark (193 questions, 50 images)

Benchmarks: Biomedical Visual Chatbot (open-ended conversation, 193 questions) VQA-RAD SLAKE PathVQA

Metrics: relative score vs GPT-4 reference (1-10 judge scale) closed-form accuracy open-form token-level recall response length (words/question) time per word

Baselines: LLaVA-Med LLaVA LLaMA Med-Flamingo MedVLM-R1 Qwen-RL-Med Qwen-VL RetinalGPT LLaDA-V BiomedCLIP

Builds on: LLaDA LLaDA-V LLaVA LLaVA-Med

Closest papers by meaning

• LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning (CVPR 2026 · similarity 0.90)
• Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.89)
• BiomedCoOp: Learning to Prompt for Biomedical Vision-Language Models (CVPR 2025 · similarity 0.89)
• ExGra-Med: Extended Context Graph Alignment for Medical Vision-Language Models (NeurIPS 2025 · similarity 0.89)
• OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.89)
• GMAI-VL & GMAI-VL-5.5M: A Large Vision-Language Model and a Comprehensive Multimodal Dataset Towards General Medical AI (AAAI 2026 · similarity 0.88)