A team of machine learning researchers has identified and solved a fundamental problem in how visual artificial intelligence agents learn to interpret ultra-high-resolution images. Rather than carefully examining image details to answer questions, these systems often rely on statistical patterns and common sense reasoning, delivering correct answers for the wrong reasons.
According to arXiv, researchers from Huawei and collaborating institutions developed Cross-Input Relative Policy Optimization (CrossRPO), a new reinforcement learning approach that fundamentally changes how agents receive training feedback. The key innovation involves creating "contrast sets," multiple versions of the same image where only a single small object or region changes, forcing the agent to identify and examine the evidence needed to distinguish between them.
How Contrast Sets Eliminate Guessing
Conventional reinforcement learning rewards an agent whenever it produces the correct answer. This creates a critical loophole: if an agent gets lucky with a guess, it receives the same reward as an agent that methodically read the relevant details. CrossRPO solves this by requiring agents to correctly answer entire sets of related questions. Because each image in a set differs only in the region containing the actual evidence, an agent cannot consistently succeed through pattern matching alone.
Researchers constructed three separate datasets to validate their approach:
- ViRA-Train: approximately 3,500 edited image sets for training purposes
- ViRA-Bench: ultra-high-resolution test images reaching 120 megapixels, with edited regions containing evidence
- ViRA-Bench-Real: additional test sets created by photographing real-world changes, eliminating any trace of digital editing that could bias results
The researchers then trained two models using this approach: ViRA-9B and ViRA-27B, referring to their respective parameter counts in billions. Despite using only about 3,500 training examples per model, the results demonstrated substantial improvements over competing methods.
Outperforming Established Baselines
When tested on ViRA-Bench at maximum resolution, ViRA-27B achieved 63.3% accuracy compared to 59.3% for GRPO, an existing reinforcement learning technique. More significantly, the smaller ViRA-9B model matched or exceeded performance from closed-source competitors, including Claude Haiku 5.5 and GPT-6-Luna, both running at their most resource-intensive inference settings.
These results carry implications for how vision-language models handle real-world applications requiring detailed visual analysis. Medical imaging, remote sensing, document processing, and other domains depend on AI systems that can reliably locate and analyze specific regions rather than making broad assumptions based on training data distributions.
"Only reading the evidence solves the whole set," the researchers explained in their abstract, highlighting why contrast sets force meaningful visual reasoning rather than statistical shortcuts.
The work suggests that training methodology may matter as much as model scale when developing visual reasoning capabilities. By restructuring how agents receive feedback during learning, CrossRPO achieves stronger performance on smaller parameter budgets. This efficiency could accelerate development of practical visual agents for resource-constrained environments.
As enterprises explore deploying AI agents for specialized tasks involving image analysis, this research demonstrates how careful experimental design and training methodology can push systems toward more reliable, verifiable behavior. The open availability of ViRA-9B positions it as a reference implementation for researchers investigating visual reasoning and training techniques.



