[{"data":1,"prerenderedAt":23},["ShallowReactive",2],{"news-en-why-dental-llm":3},{"slug":4,"lang":5,"title":6,"summary":7,"date":8,"category":9,"tags":10,"cover":15,"order":16,"author":17,"authorTitle":18,"authorSlug":19,"people":20,"bodyLang":5,"html":22},"why-dental-llm","en","Benyou Wang: Why we train a dedicated dental large model","General-purpose large models keep getting stronger, yet they still stumble in dentistry. Drawing on the data in our DentalGPT and GlobalDentBench papers, here is why we train a model specifically for dentistry.","2026-11-16","观点",[11,12,13,14],"DentalGPT","GlobalDentBench","dental LLM","perspective","/images/news/why-dental-llm/cover.jpg",0,"Benyou Wang","FreedomAI Chief AI Scientist","wang-benyou",[21],"王本友","\u003Cp>For the past three years I have been involved in research on medical large models: from HuatuoGPT and HuatuoGPT-II, to HuatuoGPT-Vision, which reads medical images, to HuatuoGPT-o1, which handles complex medical reasoning. Since we turned to oral health, the question we hear most is: general-purpose models keep getting stronger, so why train a dedicated dental model?\u003C/p>\n\u003Cp>My answer: in dentistry, &quot;general&quot; is not enough. Rather than argue the point, I will just share the data from our own papers.\u003C/p>\n\u003Ch2>1. Dentistry is almost a blind spot in general data\u003C/h2>\n\u003Cp>A large model&#39;s abilities come from its data, and dentistry makes up a tiny share of general data. In the DentalGPT preprint we counted that in PubMedVision, a very large medical image–text dataset, images involving teeth account for only about \u003Cstrong>0.3%\u003C/strong>.\u003C/p>\n\u003Cp>Dental images also demand fine detail. A panoramic radiograph shows the whole dentition, while fillings, root canal treatments, impacted teeth and periapical lesions often occupy only a small area. Intraoral photos, periapical films, lateral cephalograms and CBCT each have their own way of being read. A model that has rarely seen such images cannot be expected to understand them.\u003C/p>\n\u003Ch2>2. General models can &quot;see&quot; but cannot count or judge correctly\u003C/h2>\n\u003Cp>On MMOral-OPG-Bench, a question-answering benchmark on panoramic radiographs, leading models such as GPT-5, Gemini-2.5-Pro and Claude-Sonnet-4.5 all score below 60% accuracy. We also found that even with &quot;thinking mode&quot; switched on, their gains on oral tasks are clearly smaller than in other medical fields.\u003C/p>\n\u003Cp>The DentalGPT paper includes a telling example: asked to count the teeth with fillings on a panoramic radiograph, where the correct answer is 10, GPT-5 answers about 6, the non-thinking version of Qwen3-VL-235B counts 16, and its thinking version settles on 8 after repeated self-checking. The models find the relevant visual cues but fail to reach the right answer from them. The problem is not only &quot;not seeing&quot;; it is the missing ability to match visual details with dental knowledge.\u003C/p>\n\u003Ch2>3. The closer to real cases, the harder it gets\u003C/h2>\n\u003Cp>To see these weaknesses clearly, we worked with dental experts around the world to build the GlobalDentBench benchmark: 8,978 expert-verified questions drawn from 88 countries and regions and 14 dental specialties. Across 12 frontier models, the results are clear:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>The more a question resembles real clinical practice, the lower the score\u003C/strong>: multiple-choice questions average 81.34%, short-answer questions 64.53%, and case questions taken from real case reports only \u003Cstrong>22.34%\u003C/strong>.\u003C/li>\n\u003Cli>\u003Cstrong>The more individualized the judgment, the harder it is\u003C/strong>: average scores fall from 74.01% for &quot;knowledge recall&quot; to 55.64% for &quot;routine reasoning&quot; and 35.71% for &quot;individualized reasoning&quot;; every model scores below 50% on individualized reasoning.\u003C/li>\n\u003C/ul>\n\u003Cp>\u003Cspan class=\"news-figure\">\u003Cimg src=\"/images/globaldentbench/fig-results.png\" alt=\"Figure 1: Results of 12 frontier models on GlobalDentBench: question types, reasoning levels, overall ranking and reasoning cost (source: GlobalDentBench paper)\" loading=\"lazy\" />\u003Cspan class=\"news-caption\">Figure 1: Results of 12 frontier models on GlobalDentBench: question types, reasoning levels, overall ranking and reasoning cost (source: GlobalDentBench paper)\u003C/span>\u003C/span>\u003C/p>\n\u003Cp>In other words, today&#39;s large models are more like a student who has memorized the textbook than a doctor who can make decisions for a specific patient.\u003C/p>\n\u003Ch2>4. In dentistry, some mistakes cannot be undone\u003C/h2>\n\u003Cp>What I care about most is safety. GlobalDentBench assessed the risk of 19,080 answers to 1,590 case questions: \u003Cstrong>31.01%\u003C/strong> were judged unsafe, and \u003Cstrong>4.51%\u003C/strong> could cause irreversible harm. Orthodontics had the highest share of unsafe answers, at 44.30%.\u003C/p>\n\u003Cp>Teeth do not grow back. Our mission is to &quot;let everyone keep their natural teeth and live a healthy life to 120&quot;. An unnecessary extraction or an overly aggressive treatment recommendation comes at a permanent cost. That is why we believe today&#39;s large models are suited to helping retrieve and organize knowledge under expert supervision, and are not yet able to make dental clinical decisions on their own.\u003C/p>\n\u003Cp>\u003Cspan class=\"news-figure\">\u003Cimg src=\"/images/globaldentbench/fig-risk.png\" alt=\"Figure 2: Safety risk analysis of answers to case questions. S1 means possible reversible harm and S2 possible irreversible harm (source: GlobalDentBench paper)\" loading=\"lazy\" />\u003Cspan class=\"news-caption\">Figure 2: Safety risk analysis of answers to case questions. S1 means possible reversible harm and S2 possible irreversible harm (source: GlobalDentBench paper)\u003C/span>\u003C/span>\u003C/p>\n\u003Ch2>5. A specialist model can be small and strong\u003C/h2>\n\u003Cp>The good news is that dedicated training works. DentalGPT has only \u003Cstrong>7B\u003C/strong> parameters. We first used more than 120,000 dental images for multimodal alignment, teaching the model to &quot;see&quot;, and then used reinforcement learning to elicit complex reasoning, teaching it to &quot;think&quot;. Across five oral benchmarks its average accuracy reaches \u003Cstrong>67.1%\u003C/strong>, above GPT-5&#39;s 59.2%; on panoramic radiograph classification it reaches 84.0%, against 50.0% for its base model and 63.5% for GPT-5. Back to the filling-counting example: after training, DentalGPT checked its work repeatedly and arrived at the correct count of 10.\u003C/p>\n\u003Cp>Ablation experiments also show that the order matters. Multimodal alignment alone raises average accuracy from 46.7% to 63.2%; adding reinforcement learning lifts it to 67.1%. For a model without dental alignment, reinforcement learning brings only limited gains, and the more alignment data there is, the higher the ceiling reinforcement learning can reach. Specialist knowledge is the foundation of reasoning, not an optional extra.\u003C/p>\n\u003Cp>In a blind evaluation by 4 practicing dentists, DentalGPT mostly tied with GPT-5 and Gemini-2.5-Pro, winning more often than it lost; against Claude-Sonnet-4.5 and Qwen3-VL-235B its win rates reached 63.9% and 87.5% respectively.\u003C/p>\n\u003Cp>\u003Cspan class=\"news-figure\">\u003Cimg src=\"/images/news/dentalgpt/fig2.png\" alt=\"Figure 3: Pairwise evaluation by practicing dentists: share of DentalGPT wins, ties and losses (%) (source: DentalGPT paper)\" loading=\"lazy\" />\u003Cspan class=\"news-caption\">Figure 3: Pairwise evaluation by practicing dentists: share of DentalGPT wins, ties and losses (%) (source: DentalGPT paper)\u003C/span>\u003C/span>\u003C/p>\n\u003Ch2>6. A specialist model is not built behind closed doors\u003C/h2>\n\u003Cp>The hard part of training a dental model has never been just the algorithm:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>Experts must be deeply involved.\u003C/strong> Six licensed senior dentists oversaw the construction of GlobalDentBench from start to finish, investing 297 person-hours. In DentalGPT&#39;s expert-annotated data, any annotation with a cross-validation agreement below 85% is filtered out.\u003C/li>\n\u003Cli>\u003Cstrong>Evaluation must span countries and standards.\u003C/strong> Clinical guidelines and licensing exams differ from country to country, so we evaluated across countries from the start. OmniDentBench, now under construction, selects challenging questions from real case reports, with every question reviewed by dental experts in China and abroad, to test expert-level clinical decision-making.\u003C/li>\n\u003Cli>\u003Cstrong>Models must use tools and show their evidence.\u003C/strong> Our DentalAgent lets the model call professional tools for tooth numbering, lesion classification and measurement, and retrieve textbooks, guidelines and cases as evidence, so that its judgments can be verified.\u003C/li>\n\u003Cli>\u003Cstrong>Dentists always make the final decision.\u003C/strong> In our collaborations with Shenzhen Stomatology Hospital (Pingshan) of Southern Medical University and Foshan Stomatological Hospital, the AI makes suggestions, lists its evidence and flags what is missing, while diagnosis and treatment decisions always stay with the dentist.\u003C/li>\n\u003C/ul>\n\u003Ch2>Closing thoughts\u003C/h2>\n\u003Cp>We are not building DentalGPT to &quot;replace dentists&quot;. We want professional oral-health judgment to arrive earlier and reach more people: to catch problems while they are still small, and to save teeth while they can still be saved. General-purpose models will keep improving, but in a highly specialized field where mistakes cannot be undone, dedicated data, dedicated training and dedicated evaluation, working together with dentists, are all indispensable.\u003C/p>\n\u003Ch2>References\u003C/h2>\n\u003Cul>\n\u003Cli>DentalGPT: Incentivizing Multimodal Reasoning in Dentistry, Findings of ACL 2026: \u003Ca href=\"https://aclanthology.org/2026.findings-acl.135/\" target=\"_blank\" rel=\"noopener noreferrer\">https://aclanthology.org/2026.findings-acl.135/\u003C/a>\u003C/li>\n\u003Cli>GlobalDentBench: A Multinational Benchmark for Evaluating LLM Clinical Reasoning in Dentistry with Expert Calibration, npj Digital Medicine (accepted); preprint: \u003Ca href=\"https://arxiv.org/abs/2605.24636\" target=\"_blank\" rel=\"noopener noreferrer\">https://arxiv.org/abs/2605.24636\u003C/a>\u003C/li>\n\u003Cli>Towards Medical Complex Reasoning with LLMs through Medical Verifiable Problems (HuatuoGPT-o1), Findings of ACL 2025: \u003Ca href=\"https://aclanthology.org/2025.findings-acl.751/\" target=\"_blank\" rel=\"noopener noreferrer\">https://aclanthology.org/2025.findings-acl.751/\u003C/a>\u003C/li>\n\u003Cli>Learn more: \u003Ca href=\"/en/dentalgpt\">DentalGPT and core technology\u003C/a>, \u003Ca href=\"/en/globaldentbench\">Dental benchmarks: GlobalDentBench and OmniDentBench\u003C/a>\u003C/li>\n\u003C/ul>\n",1794841747109]