A 4B Chest X-Ray AI Was Trained With Rewards a Clinician Can Audit

A 4-billion-parameter chest X-ray report model called DobicVLM beat Google's Gemini 2.5 Flash on a paper-defined metric set, and the authors credit the result to a training recipe built from plain, hand-written rules rather than a learned neural judge, per their arXiv preprint.

DobicVLM is a fine-tune of MedGemma-4B, a medical-specialized open model from Google. The authors, affiliated with Dobic Health and DiagnoCare AI, treat chest X-ray report drafting as the workflow a radiologist actually follows: look at the film, write a structured report covering the right anatomical structures, use the right terminology, stay within a reasonable length. They encoded that workflow as four rule-based reward signals and optimized the model with Group Relative Policy Optimization (GRPO), a reinforcement learning method that compares groups of candidate outputs against each other rather than against a fixed target.

The four reward components, as described in the paper, are programmatic rather than learned. Structural verification asks whether the generated report follows the expected section layout, including findings and impression. The anatomical checklist asks whether the report mentions the right regions for a chest film. Semantic similarity measures how close the report is to a reference, using embedding overlap. The length constraint keeps the report within a target token range. Each is a short, hand-written rule. A reviewer can read the function and ask whether the rule is the right one. The same move is harder with a 7-billion-parameter reward model trained on radiologist preference labels, because that judge is itself a black box.

The reported results are on a 69-case held-out set, blindly reviewed by a single expert. The authors report 27.2% impression accuracy and 86.5% medical terminology coverage, and say the model outscored both Gemini 2.5 Flash and the MedGemma-4B base model on the majority of their criteria. The 86.5% terminology figure is the more useful number for a non-radiologist, because it means the model is using the right words. The 27.2% impression accuracy is low in absolute terms, and the paper itself flags trade-offs in completeness and referral language.

Three caveats travel with the result. First, the training set is 1,000 image-report pairs from a private, de-identified clinical dataset, enough to demonstrate a recipe but not a system. Generalization to other sites, scanners, and patient populations isn't established. Second, the comparator is Gemini 2.5 Flash in a generic multimodal configuration, not a radiology-specialized product, so the paper is comparing its recipe against a general-purpose model run with default prompting. Third, the paper is an arXiv preprint that isn't peer-reviewed. The authors disclose affiliation with Dobic Health and DiagnoCare AI, and the "transparent alignment in resource-limited settings" framing is author-stated positioning rather than third-party-validated outcome.

Most medical AI alignment work uses a learned reward model, a separate neural network trained on radiologist preferences that scores candidate reports. That judge is itself a black box. A regulator or hospital ethics board can audit the medical model's output, but they can't easily audit the thing that shaped its training. Replacing that with four short, hand-written rule sets trades expressiveness for readability: less powerful than a neural judge, but contestable and adjustable per site.

The dataset is too small and the evaluation is too narrow for any clinical deployment claim. The paper delivers a documented recipe that a clinical-AI team can take apart: here are the four rewards, here is how they were combined via GRPO, here is what the model did on 69 cases against a general-purpose baseline. A team building a chest X-ray report drafter in a low-resource setting now has a paper-defined starting point they can read end to end.

The next check is whether the recipe transfers. The authors haven't released the training data, the weights, or the reward code publicly, so independent reproduction has to wait. The wider question for the field is whether rule-based programmatic rewards stay legible as report generation gets harder, or whether the expressiveness gap to a neural judge pulls the same opacity back in.