ViRA moves the reward from one image to a contrast set: 63.3% set accuracy at maximum resolution, against 59.3% for GRPO
Synopsis
The detail that answers a question about an ultra-high-resolution photo is often only a few dozen pixels wide, and visual agents that can zoom still often answer from common sense instead of reading it. ViRA trains with CrossRPO on contrast sets: one question about several versions of an image that differ only in a small region with the evidence and have different answers, so only reading solves the whole set. On ViRA-Bench, built from 21 to 120 megapixel photos, set accuracy at maximum resolution is 63.3%, against 59.3% for GRPO trained on each version separately and 52.6% for pooled GRPO.
Interpretation
The training signal moves from the final answer on a single image to the answers on a whole contrast set, so a lucky guess no longer earns the same reward as reading. Reinforcement learning on one image gave a guess the same reward as reading the evidence, and the guess was easier, so it became a shortcut the reward could not detect. On ViRA-Bench at maximum resolution, CrossRPO reaches 63.3% set accuracy, against 59.3% for GRPO trained on each version separately and 52.6% for pooled GRPO.
Each version's rollouts are compared only with rollouts of the same version, and the result is weighted by how often the other versions are answered correctly. That weight is what separates CrossRPO from per-version RLOO: the more other versions are correct, the more credit a correct answer on this version receives for solving the set. Removing the weight and falling back to per-version RLOO lowers ViRA-Bench set accuracy by 4.6 to 7.4 points at every budget; drawing the weight from photos of other questions gives results close to per-version RLOO.
ViRA-9B and ViRA-27B start from Qwen3.5-9B and Qwen3.8-27B and each make one pass over about 3,500 training items. The training data mix filtered contrast sets from ViRA-Train with single images, 3,642 items for ViRA-9B and 3,401 for ViRA-27B. ViRA-9B is the most accurate open model of its size on all ten benchmarks and ViRA-27B on nine of ten; ViRA-27B outperforms the proprietary models tested on ViRA-Bench at every resolution.
Perspective
The demonstrated scope is contrast sets built from 21 to 120 megapixel photos plus 78 unedited photographed sets shot with a Nikon Z8; ViRA-9B and ViRA-27B start from Qwen3.5-9B and Qwen3.8-27B and each train for one pass over about 3,500 items. A reader judging whether the method carries to other architectures or other scientific images should watch the next steps the authors propose: a photographed benchmark at larger scale with outdoor scenes and more object kinds, and larger text experiments.
ViRA-Bench and ViRA-Train are made by image editing, so traces inside an edited object could affect answers; the first zoom covers the object equally often on edited and unedited versions, within 2.4 percentage points, but that test does not cover reading. ViRA-Bench-Real has only 78 sets, shot indoors with one camera and showing mostly posters, signs, stickers and labels. Image sets have at most three versions, so larger sets are tested only on text, and the text experiments are small. The reward is accuracy only, so nothing encourages tighter boxes or further zooms.
