ViRA paper: contrast-set RL stops visual agents guessing, 27B beats Claude Haiku 5.5
Huawei Lin · hf · 2026-11-18
A paper from RIT, Purdue, University of Utah and NVIDIA introduces ViRA and a new training method, CrossRPO, for ultra-high-resolution visual agents:
- Problem: visual agents often answer from common sense instead of reading the image evidence; RL on final answers can't fix this since guessing earns the same reward as reading;
- Method (CrossRPO): trains on contrast sets—the same question asked over several versions of an image differing only in a small evidence region with different answers, so only reading the evidence solves the whole set; rewarding full-set success provably doesn't reinforce common-sense answering;
- Data: contrast sets built by editing one small object in 21–120 megapixel images (ViRA-Train/ViRA-Bench), plus really-photographed variants (ViRA-Bench-Real) to exclude editing artifacts;
- Results: ViRA-9B and ViRA-27B trained on only 3,500 items each; CrossRPO reaches 63.3% set accuracy at max resolution vs 59.3% for GRPO; ViRA-9B is the most accurate open model at its size, and ViRA-27B beats Claude Haiku 5.5 and GPT-6-Luna (both at max thinking) on most benchmarks.
More from Research
- Google/Penn State paper: Self-Improvement Memory lifts harness optimization gains by 2.36 points across models — rohanpaul_ai · 2026-11-18
- Google paper: Self-Improvement Memory lifts cross-model harness transfer from 68.41 to 84.75 — rohanpaul_ai · 2026-11-18
- Deep Bundle Adjustment: 590x smaller than VGGT, initialization-free 3D reconstruction beats COLMAP beyond 15 views — Singularitarian · 2026-11-18
- Argon open-sources Alloy, a multi-material physics engine for robot learning on one GPU — ycombinator · 2026-11-18
- Mohit Bansal's distinguished lectures at Columbia and Northwestern on trustworthy agents — mohitban47 · 2026-11-18
- Arc Institute's antiGen and Germinal Multi forecast virus mutations for proactive vaccine design — BrianHie · 2026-11-18