Wuying-Browser-Agent: Real-World Centric Fundamental Long-Horizon Browser Agents

Details

Browser agents that perform well on lab benchmarks often fail in real deployment because real websites require sustained decision-making, error recovery, and navigation of complex UIs across dozens of steps. Wuying-Browser-Agent addresses this with alignment at every pipeline level rather than relying on scale alone. The framework combines four components: a structured browser harness with stable execution primitives and decision-oriented context management; Reflection and UI-specialized Curriculum SFT (RUIC-SFT) that explicitly trains on recovery trajectories and complex-UI interactions; Divergence-Aware Online GRPO (DAO-GRPO) for long-horizon credit assignment via potential-based reward shaping; and BrowserBench, a new bilingual real-web evaluation suite of 350 tasks averaging 37.9 steps — far longer than existing benchmarks. Wuying-Browser-Agent-27B achieves 80.6% on WebVoyager, 66.7% on Online-Mind2Web, and 65.1% on BrowserBench, establishing new open-source state-of-the-art results. The same pipeline transfers to general agentic tasks, reaching 73.8 average on Tau2-Bench, Claw-Eval, and BFCL-v4. The combination of a new long-horizon benchmark and strong open-source performance makes this a notable release for teams building web automation agents. The 40-author AIMAE Team from what appears to be an Alibaba Cloud research group suggests substantial industrial investment.