Research map / Vision-language and multimodal / Multimodal LLMs

OralGPT-Omni: A Versatile Dental Multimodal Large Language Model

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models · 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.

Builds OralGPT-Omni, a dental-specialized MLLM fine-tuned from Qwen2.5-VL-7B with a four-stage pipeline (including GRPO with a TRACE-based reasoning reward), and introduces the MMOral-Uni benchmark of 2,809 open-ended dental VQA pairs.

Problem

Multimodal LLMs perform poorly on dental imaging due to scarce domain data, lack of expert annotations, modality heterogeneity, and unreliable, unexplained outputs.

Approach

A clinically structured five-step chain-of-thought (TRACE-CoT) generated from sparse annotations and guideline knowledge, used as SFT supervision and as a judge-based reward in GRPO, combined with a four-stage training recipe for a dental MLLM.

Key results

• OralGPT-Omni reaches 51.84 overall on MMOral-Uni, versus 36.42 for GPT-5 in the paper's table (the text cites GPT-5 as 15.42, an inconsistency).
• On MMOral-OPG, OralGPT-Omni scores 45.31 overall versus 42.42 for GPT-5 and 39.38 for GPT-4V.
• Ablation: SFT raises the MMOral-Uni overall score from 24.00 to 48.67, and RLT adds 3.17 points to 51.84; TRACE-CoT in SFT adds 4.36 points (44.31 to 48.67).

Surprising findings

• OralGPT-Omni underperforms proprietary models on the treatment planning task, attributed to treatment planning being only 0.006% of training data.
• Report generation on MMOral-OPG lags GPT-4V (28 vs 43.70 range per table), and existing medical MLLMs show no clear advantage over general MLLMs on dental tasks.

Limitations

• Underperforms proprietary models on treatment planning due to very little training data for that task.
• Report generation lags GPT-4V on MMOral-OPG.
• LLM-judge scores diverge from dentist scores by about 2 points overall and vary per subcategory.

Experimental setup

Datasets: 31 public dental imaging datasets dental hospital dataset (Hong Kong) dental textbooks MMOral-OPG

Benchmarks: MMOral-Uni MMOral-OPG

Metrics: overall score (0-1 per sample, scaled) per-modality score

Baselines: GPT-5 o3 Grok-4 Claude-Sonnet-4.5 Gemini-2.5-flash Qwen3-VL-235B-A22B Lingshu-7B HealthGPT GPT-4V MedDr

Builds on: Qwen2.5-VL GRPO LLaMA-Factory ms-swift VLMEvalKit

Closest papers by meaning

• Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis (NeurIPS 2025 · similarity 0.92)
• OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis (CVPR 2026 · similarity 0.89)
• OralXrays-9: Towards Hospital-Scale Panoramic X-ray Anomaly Detection via Personalized Multi-Object Query-Aware Mining (CVPR 2025 · similarity 0.87)
• MedMO: Grounding and Understanding Multimodal Large Language Model for Medical Images (CVPR 2026 · similarity 0.86)
• Teeth-SEG: An Efficient Instance Segmentation Framework for Orthodontic Treatment based on Multi-Scale Aggregation and Anthropic Prior Knowledge (CVPR 2024 · similarity 0.86)
• OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.86)