TL;DR
Visual agents often rely on common-sense "guesses" rather than actually reading high-resolution details, a problem standard reinforcement learning (RL) fails to fix. This paper introduces Cross-Input Relative Policy Optimization (CrossRPO), which trains agents on "contrast sets"—multiple versions of an image where only a tiny detail changes. By rewarding the agent only when it solves the entire set correctly, CrossRPO forces models to actually inspect the evidence, allowing a 27B model to outperform proprietary giants like GPT-6-Luna and Claude Haiku 5.5 on ultra-high-resolution tasks.
What Problem Does This Paper Address?
Modern cameras capture images at 20 to 120 megapixels (MP), but most Vision-Language Models (VLMs) are capped at much lower resolutions (often around 1MP) to keep computational costs manageable. To see fine details—like the text on a distant sign or a small object in a pedestrian's hand—visual agents must learn to "zoom in."
However, the authors identify a major hurdle: shortcut learning. Because common sense often provides the right answer (e.g., a person holding a small rectangular device is usually holding a smartphone), agents learn to guess correctly without ever looking at the high-resolution crop. Standard RL methods like Group Relative Policy Optimization (GRPO) can't fix this because a "lucky guess" on a single image receives the same reward as a "diligent read."
Methods
The core innovation is Cross-Input Relative Policy Optimization (CrossRPO). Instead of treating every image as an independent data point, CrossRPO trains on contrast sets.
A contrast set consists of a single question applied to several versions of an image that are identical except for one small region containing the evidence. For example, in one version, a woman holds a smartphone; in another, a calculator; in a third, a TV remote. To get the "set reward," the agent must answer correctly for all versions.
How CrossRPO works: * Joint Reward: It adds a "joint term" to the objective function that rewards the agent for solving the entire set. * Cross-Input Weight: During training, the advantage (reward signal) for a specific rollout is scaled by how well the agent performed on the other versions of that set. If the agent fails version A, it gets less credit for succeeding on version B, because it hasn't yet proven it can distinguish the two. * Unbiased Gradient: The authors mathematically prove that CrossRPO provides an unbiased gradient estimate and, crucially, does not reinforce "blind" policies that rely on priors.
Datasets
The authors introduced three new datasets to support this approach:
- ViRA-Train: 9,261 contrast sets (and some single images) built by editing small objects in 21–120 MP photos.
- ViRA-Bench: 762 sets used for evaluation, where the target object covers a median of only 0.0075% of the total image area.
- ViRA-Bench-Real: To ensure the models weren't just picking up on "editing traces," the authors photographed 78 sets of real-world scenes where objects (like signs or posters) were physically swapped between shots.
Key Insights and Ideas
- The "Set Accuracy" Metric: The paper argues that standard per-image accuracy is a misleading metric for high-res agents. "Set accuracy"—the share of contrast sets where every version is correct—is a much tougher and more meaningful test of whether a model is actually "reading" the pixels.
- Reading vs. Finding: Interestingly, the authors found that CrossRPO doesn't necessarily make the agent better at finding where to zoom; rather, it makes the agent better at interpreting what it sees once it gets there. Without the joint term, agents often zoomed into the right spot but still hallucinated the common-sense answer.
- Efficiency: The ViRA models were trained on only about 3,500 items. This suggests that forcing "hard" reasoning through contrast sets is significantly more data-efficient than massive-scale standard RL.
Key Results
The resulting agents, ViRA-9B and ViRA-27B (based on Qwen), showed dominant performance:
- Outperforming Proprietary Models: ViRA-27B outperformed Claude Haiku 5.5 and GPT-6-Luna (at their highest thinking levels) on most high-resolution benchmarks.
- Set Accuracy: On ViRA-Bench, CrossRPO achieved 63.3% set accuracy, significantly higher than standard GRPO (59.3%).
- Generalization: Despite being trained on specific contrast sets, the models showed improved accuracy on general benchmarks like MME-RealWorld and HR-Bench, proving the "reading" skill transfers to diverse tasks.
Limitations and Future Work
- Zoom Precision: The current reward is based only on final answer accuracy. The authors note that adding rewards for "tighter" zoom boxes could further improve performance at very low pixel budgets.
- Dataset Complexity: Creating contrast sets is more expensive than collecting single images, as it requires careful image editing or multiple physical photos of the same scene.
- Architecture: The study focused on the Qwen family; it remains to be seen if CrossRPO provides similar gains across all VLM architectures.
- Text Scaling: While the authors showed CrossRPO works for text-based contrast sets (e.g., swapping facts in a passage), these experiments were small-scale and leave room for future exploration in pure NLP.
Summary updated