[{"data":1,"prerenderedAt":24},["ShallowReactive",2],{"news-en-neurips-2026-oral-medical-vision":3},{"slug":4,"lang":5,"title":6,"summary":7,"date":8,"category":9,"tags":10,"cover":16,"order":17,"author":18,"authorTitle":18,"authorSlug":18,"people":19,"bodyLang":5,"html":23},"neurips-2026-oral-medical-vision","en","NeurIPS 2026 Oral | FreedomAI and The Chinese University of Hong Kong, Shenzhen release a medical multimodal large model that uses reinforcement learning to unlock medical vision capabilities","HuatuoGPT-Vision-2, a medical multimodal large model trained with \"pre-training + reinforcement learning\". Its 32B version averages 74.0 across eight medical benchmarks, ahead of GPT-5.4-mini.","2026-11-16","论文",[11,12,13,14,15],"NeurIPS 2026","medical multimodal AI","reinforcement learning","open data","HuatuoGPT-Vision-2","/images/news/neurips-2026-oral-medical-vision/cover.jpg",1,"",[20,21,22],"蔡镇阳","冯相义","王本友","\u003Cp>Getting a multimodal large model to understand medical images is hard for more than one reason. Scarce data is only part of it: even when a model has &quot;seen&quot; huge numbers of medical images during pre-training, that knowledge is not necessarily available when it answers questions. A new study finds that conventional supervised fine-tuning struggles to bring this knowledge out.\u003C/p>\n\u003Cp>The paper &quot;Incentivizing Medical Vision Capabilities from Large-Scale Multimodal Pre-training&quot;, by researchers at The Chinese University of Hong Kong, Shenzhen together with members of the FreedomAI team, has been accepted to NeurIPS 2026 as an \u003Cstrong>Oral\u003C/strong>. It lays out a simple recipe for training medical vision-language models: first run large-scale multimodal pre-training on medical image–text data mined from the PubMed literature, then run reinforcement learning on questions synthesized directly from the pre-training data, &quot;incentivizing&quot; the model to use the medical visual knowledge it has learned. Across eight medical multimodal benchmarks, the 32B version of HuatuoGPT-Vision-2 averages \u003Cstrong>74.0\u003C/strong>, ahead of the closed-source GPT-5.4-mini (71.1) and of open-source general and medical models of the same size, substantially narrowing the gap to the leading closed-source models.\u003C/p>\n\u003Cp>FreedomAI CTO Zhenyang Cai is the paper&#39;s second author, Vice President of Medical Products Xiangyi Feng is one of its authors, and FreedomAI Chief AI Scientist Benyou Wang, an assistant professor at The Chinese University of Hong Kong, Shenzhen, is the corresponding author. The paper will be presented orally at NeurIPS 2026 in Sydney on December 8.\u003C/p>\n\u003Cblockquote>\n\u003Cp>\u003Cstrong>Paper\u003C/strong>: Incentivizing Medical Vision Capabilities from Large-Scale Multimodal Pre-training\u003Cbr>\u003Cstrong>Conference\u003C/strong>: NeurIPS 2026 (Oral)\u003Cbr>\u003Cstrong>Link\u003C/strong>: \u003Ca href=\"https://openreview.net/forum?id=0BXKsjKlFJ\" target=\"_blank\" rel=\"noopener noreferrer\">https://openreview.net/forum?id=0BXKsjKlFJ\u003C/a>\u003Cbr>\u003Cstrong>Code\u003C/strong>: \u003Ca href=\"https://github.com/FreedomIntelligence/HuatuoGPT-Vision-2\" target=\"_blank\" rel=\"noopener noreferrer\">https://github.com/FreedomIntelligence/HuatuoGPT-Vision-2\u003C/a>\u003Cbr>\u003Cstrong>Data\u003C/strong>: \u003Ca href=\"https://huggingface.co/datasets/FreedomIntelligence/PubMedVision-Plus\" target=\"_blank\" rel=\"noopener noreferrer\">https://huggingface.co/datasets/FreedomIntelligence/PubMedVision-Plus\u003C/a>\u003Cbr>\u003Cstrong>Authors\u003C/strong>: Junying Chen, Zhenyang Cai, Yunjin Yang, Kunyuan Cai, Rongsheng Wang, Xidong Wang, Geyang Yu, Xiangyi Feng, Benyou Wang (corresponding author)\u003Cbr>\u003Cstrong>Affiliations\u003C/strong>: The Chinese University of Hong Kong, Shenzhen; Shenzhen Loop Area Institute; Shenzhen Research Institute of Big Data; National Institute of Health Data Science (Shenzhen)\u003C/p>\n\u003C/blockquote>\n\u003Ch2>Background: medical knowledge goes in but doesn&#39;t come out\u003C/h2>\n\u003Cp>Building a useful medical vision-language model (VLM) faces two hurdles: high-quality medical data is scarce, and the medical visual knowledge a model learns during pre-training is hard to bring out.\u003C/p>\n\u003Cp>In earlier work, HuatuoGPT-Vision (EMNLP 2024), the group built PubMedVision, 1.3 million medical visual question-answering samples drawn from the PubMed literature. The new work scales the data up further along the same path and focuses on a more fundamental question: how can the knowledge learned in pre-training actually be put to use?\u003C/p>\n\u003Cp>The paper&#39;s answer: after large-scale medical pre-training, conventional supervised fine-tuning (SFT) struggles to activate the pre-trained knowledge; switching to reinforcement learning (RL), with training questions taken directly from the pre-training data itself, &quot;incentivizes&quot; the model to draw on that knowledge.\u003C/p>\n\u003Ch2>Method: a two-stage recipe\u003C/h2>\n\u003Cp>The medical vision-language model proposed in the paper is \u003Cstrong>HuatuoGPT-Vision-2\u003C/strong>, the second-generation medical VLM in the HuatuoGPT series. It is built on Qwen3-VL-8B-Instruct and Qwen3-VL-32B-Instruct and trained in two stages.\u003C/p>\n\u003Ch3>Stage 1: scale up medical multimodal pre-training\u003C/h3>\n\u003Cp>The team extracted figures and their captions from \u003Cstrong>15.2 million\u003C/strong> PubMed papers published between 2013 and 2026, yielding \u003Cstrong>60.51 million\u003C/strong> medical image–text pairs, named PubMedVision-Plus. According to the paper, it is the largest medical image–text dataset to date.\u003C/p>\n\u003Cp>The data then went through three rounds of cleaning. First, medical-image detection removed charts, tables, diagrams and other non-medical-image content (18.42 million pairs remained). Next, a large language model filtered by the informativeness and clinical value of the captions (13.7 million remained). Finally, a vision-language model checked image–text consistency and image quality, producing about \u003Cstrong>12 million\u003C/strong> high-quality core pairs, PubMedVision-Plus-Core. Throughout, the original figures and author-written captions were kept without generative rewriting, to avoid hallucinations introduced by synthetic captions.\u003C/p>\n\u003Cp>By imaging modality, the core data covers microscopy/pathology (25.6%), digital photographs (18.7%), MRI (7.8%), CT (6.6%), X-ray (3.4%), ultrasound (3.1%), ophthalmic imaging (1.9%), endoscopy (1.7%) and more. By anatomical region, the cellular level accounts for 16.3%, head, neck and oral cavity 14.8%, abdomen/pelvis 11.7% and chest 10.9%. The team continued pre-training Qwen3-VL on this data to obtain the medical pre-trained models.\u003C/p>\n\u003Cp>\u003Cspan class=\"news-figure\">\u003Cimg src=\"/images/news/neurips-2026-oral-medical-vision/fig1-pretraining.png\" alt=\"Figure 1: Stage 1. About 60 million medical image–text pairs are extracted from 15.2 million PubMed papers and cleaned over several rounds into about 12 million high-quality core pairs used for continued pre-training; below, the data filtering process and the distribution by anatomical region and imaging modality (source: the paper)\" loading=\"lazy\" />\u003Cspan class=\"news-caption\">Figure 1: Stage 1. About 60 million medical image–text pairs are extracted from 15.2 million PubMed papers and cleaned over several rounds into about 12 million high-quality core pairs used for continued pre-training; below, the data filtering process and the distribution by anatomical region and imaging modality (source: the paper)\u003C/span>\u003C/span>\u003C/p>\n\u003Ch3>Stage 2: use reinforcement learning to incentivize medical vision capabilities\u003C/h3>\n\u003Cp>Grounded in the real images and expert-written captions of PubMedVision-Plus, the team synthesized 70,000 reinforcement learning training questions, PubMedVision-RL, in two types:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>Verifiable rewards\u003C/strong>: 35,331 single-choice questions with clearly labeled answers, checked by rule: 1 point for a correct answer, 0 for a wrong one. These rewards are sparse but precise, low in noise and easy to judge.\u003C/li>\n\u003Cli>\u003Cstrong>Rubric rewards\u003C/strong>: 35,154 open-ended questions, each with a weighted scoring rubric. Positive items reward key findings, such as &quot;mentions pleural effusion +5&quot;; negative items penalize errors, such as &quot;hallucinates pneumothorax −6&quot;. A grader model checks the answer against each item, and the reward is the sum of the points of the items satisfied divided by the total of all positive points. These rewards are denser and give partial credit to partially correct answers.\u003C/li>\n\u003C/ul>\n\u003Cp>To ensure quality, the team used GPT-5.4 to double-check every question and every rubric item, removing questions that were ambiguous or went beyond the information in the caption. They also dropped questions that Qwen3-VL-8B-Instruct could already answer correctly, so that every sample provides a useful training signal.\u003C/p>\n\u003Cp>To preserve general text abilities, 15,000 general-domain rubric questions from RubricHub were mixed into training, for a total of 85,000 reinforcement learning samples. Training used the GRPO (Group Relative Policy Optimization) algorithm with the DAPO recipe (including clip-higher, token-level loss, dynamic sampling and reward shaping for overlong responses). Rewards for the open-ended questions come from HuatuoGPT-Vision-2-Grader, a multimodal grader model trained by the team: based on Qwen3-VL-8B-Instruct and trained on 1 million rubric judgments by GPT-4.1, it sees both the medical image and the answer and judges all rubric items in a single generation. The grader is used only to compute rewards during training; all rubric evaluation results reported in the paper are scored by GPT-4.1.\u003C/p>\n\u003Cp>\u003Cspan class=\"news-figure\">\u003Cimg src=\"/images/news/neurips-2026-oral-medical-vision/fig2-rl.png\" alt=\"Figure 2: Stage 2. Single-choice questions (verifiable rewards) and open-ended questions (rubric rewards) are synthesized from the core data, and reinforcement learning incentivizes the medical pre-trained model to use its visual knowledge (source: the paper)\" loading=\"lazy\" />\u003Cspan class=\"news-caption\">Figure 2: Stage 2. Single-choice questions (verifiable rewards) and open-ended questions (rubric rewards) are synthesized from the core data, and reinforcement learning incentivizes the medical pre-trained model to use its visual knowledge (source: the paper)\u003C/span>\u003C/span>\u003C/p>\n\u003Ch2>Results\u003C/h2>\n\u003Ch3>Eight medical benchmarks: the 32B model averages 74.0, ahead of GPT-5.4-mini\u003C/h3>\n\u003Cp>The paper evaluates on eight medical multimodal benchmarks, GMAI-MMBench, the medical subset of MMMU-Pro, MedXpertQA-MM, OmniMedVQA, PMC-VQA, PathVQA, SLAKE and VQA-RAD, comparing against closed-source models, open-source general models and medical models.\u003C/p>\n\u003Cp>\u003Cstrong>Table 1\u003C/strong>: Average accuracy (%) across eight medical benchmarks, sorted by average (excerpt from Table 2 of the paper)\u003C/p>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Model\u003C/th>\n\u003Cth>Type\u003C/th>\n\u003Cth align=\"right\">Eight-benchmark average\u003C/th>\n\u003C/tr>\n\u003C/thead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>Gemini-3-Pro\u003C/td>\n\u003Ctd>Closed-source\u003C/td>\n\u003Ctd align=\"right\">81.1\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Gemini-3-Flash\u003C/td>\n\u003Ctd>Closed-source\u003C/td>\n\u003Ctd align=\"right\">78.4\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>GPT-5.4 (High)\u003C/td>\n\u003Ctd>Closed-source\u003C/td>\n\u003Ctd align=\"right\">75.0\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>\u003Cstrong>HuatuoGPT-Vision-2-32B\u003C/strong>\u003C/td>\n\u003Ctd>\u003Cstrong>This work\u003C/strong>\u003C/td>\n\u003Ctd align=\"right\">\u003Cstrong>74.0\u003C/strong>\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>GPT-5.4-mini (High)\u003C/td>\n\u003Ctd>Closed-source\u003C/td>\n\u003Ctd align=\"right\">71.1\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>\u003Cstrong>HuatuoGPT-Vision-2-8B\u003C/strong>\u003C/td>\n\u003Ctd>\u003Cstrong>This work\u003C/strong>\u003C/td>\n\u003Ctd align=\"right\">\u003Cstrong>68.5\u003C/strong>\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Gemma-4-31B-IT\u003C/td>\n\u003Ctd>Open-source general\u003C/td>\n\u003Ctd align=\"right\">67.3\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Lingshu-32B\u003C/td>\n\u003Ctd>Open-source medical\u003C/td>\n\u003Ctd align=\"right\">65.8\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Qwen3-VL-32B-Instruct\u003C/td>\n\u003Ctd>Open-source general (32B base)\u003C/td>\n\u003Ctd align=\"right\">63.9\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Qwen3.5-27B\u003C/td>\n\u003Ctd>Open-source general\u003C/td>\n\u003Ctd align=\"right\">63.3\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>GPT-4.1\u003C/td>\n\u003Ctd>Closed-source\u003C/td>\n\u003Ctd align=\"right\">63.1\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Qwen3-VL-8B-Instruct\u003C/td>\n\u003Ctd>Open-source general (8B base)\u003C/td>\n\u003Ctd align=\"right\">58.6\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>InternVL3.5-8B\u003C/td>\n\u003Ctd>Open-source general\u003C/td>\n\u003Ctd align=\"right\">57.9\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>GPT-5.4-nano (High)\u003C/td>\n\u003Ctd>Closed-source\u003C/td>\n\u003Ctd align=\"right\">57.0\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>HuatuoGPT-Vision-7B\u003C/td>\n\u003Ctd>Open-source medical (previous generation)\u003C/td>\n\u003Ctd align=\"right\">53.9\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>MedGemma-27B-IT\u003C/td>\n\u003Ctd>Open-source medical\u003C/td>\n\u003Ctd align=\"right\">50.7\u003C/td>\n\u003C/tr>\n\u003C/tbody>\u003C/table>\n\u003Cp>Key takeaways:\u003C/p>\n\u003Cul>\n\u003Cli>\u003Cstrong>The 32B version averages 74.0\u003C/strong>, clearly above the same-size open-source general model Gemma-4-31B-IT (67.3) and medical model Lingshu-32B (65.8), and above the closed-source GPT-5.4-mini (71.1). A gap remains to the strongest closed-source models such as GPT-5.4 (75.0) and Gemini-3-Pro (81.1).\u003C/li>\n\u003Cli>\u003Cstrong>The 8B version averages 68.5\u003C/strong>, 9.9 points above its base, Qwen3-VL-8B-Instruct, and ahead of several larger open-source models.\u003C/li>\n\u003Cli>On OmniMedVQA, the 32B version scores \u003Cstrong>88.4\u003C/strong>, the highest of all models compared.\u003C/li>\n\u003Cli>Medical training does not come at the expense of general ability: across six general benchmarks including MMLU-Pro, IFEval and GPQA, the 8B version&#39;s average rises slightly and the 32B version&#39;s stays about the same.\u003C/li>\n\u003C/ul>\n\u003Ch3>Ablation: why reinforcement learning\u003C/h3>\n\u003Cp>With the same 85,000 training prompts, different training recipes perform very differently (the numbers in Table 2 are changes in the eight-benchmark average relative to the base model):\u003C/p>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Training recipe\u003C/th>\n\u003Cth align=\"right\">8B\u003C/th>\n\u003Cth align=\"right\">32B\u003C/th>\n\u003C/tr>\n\u003C/thead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>Reinforcement learning only (no medical pre-training)\u003C/td>\n\u003Ctd align=\"right\">+1.9\u003C/td>\n\u003Ctd align=\"right\">+4.7\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Medical pre-training + supervised fine-tuning\u003C/td>\n\u003Ctd align=\"right\">+1.1\u003C/td>\n\u003Ctd align=\"right\">−1.7\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Medical pre-training + reinforcement learning (this work)\u003C/td>\n\u003Ctd align=\"right\">\u003Cstrong>+9.9\u003C/strong>\u003C/td>\n\u003Ctd align=\"right\">\u003Cstrong>+10.1\u003C/strong>\u003C/td>\n\u003C/tr>\n\u003C/tbody>\u003C/table>\n\u003Cp>\u003Cstrong>Table 2\u003C/strong>: Effect of different training recipes (Table 2 of the paper; answers for supervised fine-tuning generated by Qwen3-VL-32B-Instruct)\u003C/p>\n\u003Cp>Even when the &quot;teacher&quot; for supervised fine-tuning is switched to GPT-5.4 or Gemini-3-Pro, the 8B model averages only 60.8 and 60.0, far below the 68.5 of &quot;pre-training + reinforcement learning&quot;. The paper concludes that, with the same prompts and amount of data, reinforcement learning is better at activating the medical visual knowledge learned in pre-training, and that medical pre-training, which injects domain knowledge, is a prerequisite for reinforcement learning to work.\u003C/p>\n\u003Ch3>Expert evaluation: rubric scoring and blind review by physicians\u003C/h3>\n\u003Cp>To evaluate open-ended answers, the team built the MedMM-ExpertRubric evaluation set: \u003Cstrong>500\u003C/strong> questions, 538 images and \u003Cstrong>11,088\u003C/strong> rubric items, about 22 weighted items per question on average. Every item was screened by licensed physicians, keeping only the core image-reading points. The questions cover radiology, pathology, ultrasound, endoscopy and other fields; 32 involve multiple images, and none were used in training. On this set, the 8B version improves from 22.1 for its base to \u003Cstrong>35.7\u003C/strong>, and the 32B version from 30.9 to \u003Cstrong>40.8\u003C/strong>.\u003C/p>\n\u003Cp>The team also asked two attending physicians to compare answers to 100 open-ended questions pairwise and blind: the two models&#39; answers were anonymized and randomly ordered, and the physicians picked the better one or called a tie. Results for HuatuoGPT-Vision-2-32B:\u003C/p>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Compared with\u003C/th>\n\u003Cth>Win / tie / loss (%)\u003C/th>\n\u003Cth>Verdict\u003C/th>\n\u003C/tr>\n\u003C/thead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>GPT-5.4 (High)\u003C/td>\n\u003Ctd>33 / 19 / 48\u003C/td>\n\u003Ctd>Slightly behind\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>GPT-5.4-mini (High)\u003C/td>\n\u003Ctd>41 / 21 / 38\u003C/td>\n\u003Ctd>On par\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Qwen3-VL-32B-Instruct\u003C/td>\n\u003Ctd>55 / 18 / 27\u003C/td>\n\u003Ctd>Clearly ahead\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>Lingshu-32B\u003C/td>\n\u003Ctd>65 / 14 / 21\u003C/td>\n\u003Ctd>Clearly ahead\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>GPT-5.4-nano (High)\u003C/td>\n\u003Ctd>68 / 16 / 16\u003C/td>\n\u003Ctd>Clearly ahead\u003C/td>\n\u003C/tr>\n\u003C/tbody>\u003C/table>\n\u003Cp>\u003Cstrong>Table 3\u003C/strong>: Pairwise blind evaluation by physicians (Table 13 of the paper)\u003C/p>\n\u003Cp>The paper notes that the physicians&#39; preferences agree with the MedMM-ExpertRubric rankings, suggesting the evaluation set can serve as a low-cost substitute for expert evaluation.\u003C/p>\n\u003Ch3>Is the grader reliable?\u003C/h3>\n\u003Cp>In reinforcement learning, the rewards for open-ended questions come from the grader model, so the reliability of its judgments directly affects training. On held-out data, the 8B grader agrees with GPT-4.1&#39;s judgments \u003Cstrong>91.5%\u003C/strong> of the time. The team also sampled 300 rubric judgments and compared them with physicians&#39; annotations: agreement was 88.7% for the 8B grader and 90.3% for GPT-4.1.\u003C/p>\n\u003Ch2>Open resources\u003C/h2>\n\u003Cp>As of November 16, 2026, the project has released the following on GitHub and Hugging Face:\u003C/p>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Resource\u003C/th>\n\u003Cth>Contents\u003C/th>\n\u003Cth>Status\u003C/th>\n\u003C/tr>\n\u003C/thead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>\u003Ca href=\"https://huggingface.co/datasets/FreedomIntelligence/PubMedVision-Plus\" target=\"_blank\" rel=\"noopener noreferrer\">PubMedVision-Plus\u003C/a>\u003C/td>\n\u003Ctd>About 12 million high-quality medical image–text pairs (Core)\u003C/td>\n\u003Ctd>Released\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>\u003Ca href=\"https://huggingface.co/datasets/FreedomIntelligence/PubMedVision-RL\" target=\"_blank\" rel=\"noopener noreferrer\">PubMedVision-RL\u003C/a>\u003C/td>\n\u003Ctd>70,000 reinforcement learning training questions (single-choice + open-ended)\u003C/td>\n\u003Ctd>Released\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>\u003Ca href=\"https://huggingface.co/datasets/FreedomIntelligence/MedMM-ExpertRubric\" target=\"_blank\" rel=\"noopener noreferrer\">MedMM-ExpertRubric\u003C/a>\u003C/td>\n\u003Ctd>500 open-ended medical visual question-answering evaluation questions with physician-verified rubrics\u003C/td>\n\u003Ctd>Released\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>\u003Ca href=\"https://huggingface.co/FreedomIntelligence/HuatuoGPT-Vision-2-Grader\" target=\"_blank\" rel=\"noopener noreferrer\">HuatuoGPT-Vision-2-Grader\u003C/a>\u003C/td>\n\u003Ctd>Multimodal grader model\u003C/td>\n\u003Ctd>Released\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>\u003Ca href=\"https://github.com/FreedomIntelligence/HuatuoGPT-Vision-2\" target=\"_blank\" rel=\"noopener noreferrer\">Training code\u003C/a>\u003C/td>\n\u003Ctd>Code for both pre-training and reinforcement learning stages, supporting the Qwen3-VL and Qwen3.8 series\u003C/td>\n\u003Ctd>Released\u003C/td>\n\u003C/tr>\n\u003Ctr>\n\u003Ctd>HuatuoGPT-Vision-2-8B / 32B\u003C/td>\n\u003Ctd>Medical vision-language models\u003C/td>\n\u003Ctd>Coming soon\u003C/td>\n\u003C/tr>\n\u003C/tbody>\u003C/table>\n\u003Cp>Note that, according to the project, HuatuoGPT-Vision-2 is intended for academic research only and must not be used for medical decisions such as diagnosis, treatment or triage.\u003C/p>\n\u003Ch2>Significance and outlook\u003C/h2>\n\u003Cp>This work offers a reusable recipe for medical multimodal models: scale up pre-training with public medical literature, synthesize verifiable questions and rubric questions from the same data, and use reinforcement learning to &quot;incentivize&quot; the pre-trained knowledge. The authors hope these data, models and findings will accelerate open research on medical vision-language models. Next, the weights of the 8B and 32B HuatuoGPT-Vision-2 models will be released.\u003C/p>\n\u003Cp>For FreedomAI, the expertise built up through the HuatuoGPT series in data construction, model training and professional evaluation is the methodological foundation for DentalGPT and oral-health applications. Notably, images of the head, neck and oral cavity make up 14.8% of the PubMedVision-Plus core data.\u003C/p>\n\u003Ch2>The team\u003C/h2>\n\u003Cp>The paper&#39;s first author is Junying Chen. FreedomAI team members Zhenyang Cai, Xiangyi Feng and Benyou Wang took part in the research.\u003C/p>\n\u003Ch3>Zhenyang Cai | FreedomAI CTO, second author\u003C/h3>\n\u003Cp>A PhD student at The Chinese University of Hong Kong, Shenzhen, he leads multimodal medical model development at FreedomAI. He contributed to models including HuatuoGPT-o1 and DentalGPT and has published first-author papers at NeurIPS and ACL.\u003C/p>\n\u003Ch3>Xiangyi Feng | FreedomAI Vice President of Medical Products, author\u003C/h3>\n\u003Cp>An attending surgeon with experience of more than 4,000 operations, he is responsible for clinical requirements and bringing products into practice at FreedomAI. He works on data construction and application optimization for medical models, combining AI product design and project management experience with frontline clinical experience.\u003C/p>\n\u003Ch3>Benyou Wang | FreedomAI Chief AI Scientist, corresponding author\u003C/h3>\n\u003Cp>An assistant professor and doctoral supervisor at the School of Data Science, The Chinese University of Hong Kong, Shenzhen.\u003C/p>\n",1794841746734]