On-Policy Data Evolution: Visual-Native Search Agents Get Self-Improving Training Through Image Bank Protocol

arXiv·medium signal

Proposes a visual-native agent harness where every tool-returned image becomes an addressable reference reusable by later tools, solving the 'transient image' problem. On-Policy Data Evolution (ODE) runs a closed-loop generator that refines training data from the policy's own rollouts. Improves Qwen3-VL-8B from 24.9% to 39.0% and at 30B from 30.6% to 41.5% across search benchmarks.