A specialist AI model for veterinary X-rays matched or exceeded the reported score of a much larger fine-tuned model in some runs, according to an arXiv preprint. V-REX variants recorded micro-F1 scores of 0.34 to 0.4, compared with 0.34 to 0.38 for PaliGemma with full fine-tuning; PaliGemma adapted with LoRA scored 0.11 to 0.13, while o3 scored 0.2.
Micro-F1 is the paper’s main report-comparison score, summarising how closely findings in generated reports match those in veterinarian-written reports. V-REX was trained from scratch on 15 million proprietary veterinary text-image pairs, making this an offline model benchmark rather than a study of animal or patient outcomes.
A model built around veterinary text
V-REX uses a decoder-only transformer that receives image tokens and predicts report text. Its RAPTOR image tokens are described as a DINO-CLIP hybrid trained on the same data. The paper also uses image generative pre-training, presented as a way to improve data use and potentially make use of unlabeled images.
The domain tokenizer lowercases and typo-cleans the text, splits it first at word level and then with BPE, a way of breaking text into reusable pieces, and uses a 32,768-token vocabulary. For text averaging about 520 characters, it produced about 30.8 tokens on average, versus 106 for the comparison tokenizer, making the comparison sequences roughly six times longer.
Reported accuracy varied by dataset size
Across the reported dataset sizes, top-1 accuracy—the rate at which the model’s first choice was correct—was 0.210 for 0.1 million samples, 0.327 for 0.5 million, 0.351 for 3 million and 0.394 for 15 million, after 30 million samples had been seen.
On a 3 million-sample subset after 15 epochs, the 512-centre generative-pretraining setting had the best listed top-1 accuracy, 0.3898, and top-5 accuracy, 0.6857, while also posting the lowest listed perplexity, 13.85. Perplexity is a measure of how well a model predicts text.
The authors added image-focused decoding
At inference, when the model generates a report, the paper’s image-free guidance procedure runs it twice for each next token—once with the image and once without it—then scales the difference in their token probabilities. The authors report that amplifying the image’s influence was associated with a noticeable increase in correct reports, particularly on smaller datasets, although the supplied results give no numeric F1 values for those settings.
For examples with multiple images, a rule that chose the most likely next token across the images was reported to add up to 0.03 to top-1 accuracy without changing the model.
The result remains tied to a private benchmark
The figures are descriptive: no confidence intervals or significance tests are reported, and the models differ in parameter count, training duration and adaptation method. The o3 entries for training time and trainable parameters are unknown, so the comparison does not isolate which factor drove the differences.
The dataset is proprietary, and the paper does not report exact numbers of animals, images, reports, unique subjects or train/test splits. No data-availability statement is reported in the supplied text.
The document is an arXiv version 1 preprint dated 20 Aug 2026. Its evidence is limited to offline experiments on proprietary veterinary X-ray and report data; it does not establish performance on outside datasets or readiness for clinical use.
Paper data and sources
Original title: V-REX: Efficient Specialist VLM Training for Veterinary X-Rays
Authors: Tim Elsner, Nicole McNally, Andre Dourson, Michael Fitzke
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text