Research map / Vision-language and multimodal / Multimodal LLMs

CURE: Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

CURE fine-tunes MedGemma-4B-IT with LoRA on public chest X-ray grounding datasets using an error-aware curriculum that reweights sampling by IoU and CXRFEScore errors, improving grounding and reducing hallucinations versus MAIRA-2.

Problem

Medical vision-language models generating chest X-ray reports often lack visual grounding and produce hallucinated findings, and standard multi-task training handles dataset and class imbalance poorly.

Approach

An error-aware curriculum that periodically evaluates per-dataset and per-class performance (IoU and CXRFEScore) and reweights sampling toward high-error sources, combined with fine-grained instruction reformulation of Chest ImaGenome into Locate, Describe, and Locate-and-describe anatomy-grounded tasks.

Key results

• CURE reaches 0.601 mean IoU on Chest ImaGenome AGRG versus 0.249 for MAIRA-2 (+0.35 IoU).
• On PadChest-GR phrase grounding, CURE reaches 0.453 micro IoU versus 0.280 for MAIRA-2.
• In the hallucination analysis over six anatomies, mean abnormal-finding hallucination falls from 26.50% (MAIRA-2) to 8.78% for CURE (about 18.6 points).

Surprising findings

• Standard phrase grounding training appears biased toward abnormal findings, which the authors link to MAIRA-2's hallucinations; AGRG training exposing normal and abnormal descriptions reduced this.
• CURE does not beat MAIRA-2 on most text-based GRG metrics on PadChest-GR, which the authors attribute to MAIRA-2's proprietary USMix pretraining.

Limitations

• Text-based GRG metrics trail MAIRA-2, which benefits from proprietary USMix pretraining data.
• Hallucination evaluation relies on an LLM (Gemini 2.5 Flash Lite) NLI judge and a 1000-sample subset of the Chest ImaGenome test set.
• Curriculum re-weighting is performed only once after warm-up in the final model, and the best reweighting interval was chosen on the ablation.

Experimental setup

Datasets: MIMIC-CXR Chest ImaGenome PadChest-GR MS-CXR VinDr-CXR

Benchmarks: MS-CXR PadChest-GR VinDr-CXR Chest ImaGenome MIMIC-CXR

Metrics: IoU CheXbert F1 CheXbert cosine similarity CXRFEScore RaTEScore RadGraph F1 NLI entailment/contradiction rate

Baselines: MAIRA-2 MedGemma-4B-IT CXRMate-RRG24

Builds on: MAIRA-2 MedGemma self-paced curriculum learning LoRA

Closest papers by meaning

• S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation (AAAI 2026 · similarity 0.87)
• LLM-RG4: Flexible and Factual Radiology Report Generation Across Diverse Input Contexts (AAAI 2025 · similarity 0.87)
• AHIVE: Anatomy-aware Hierarchical Vision Encoding for Interactive Radiology Report Retrieval (CVPR 2024 · similarity 0.86)
• Bootstrapping Large Language Models for Radiology Report Generation (AAAI 2024 · similarity 0.86)
• ReFINE: A Reward-Based Framework for Interpretable and Nuanced Evaluation of Radiology Report Generation (AAAI 2026 · similarity 0.86)
• DART: Disease-aware Image-Text Alignment and Self-correcting Re-alignment for Trustworthy Radiology Report Generation (CVPR 2025 · similarity 0.85)