StepGuard: Guarding Web Navigation via Single-Step Calibration

In three linesStepGuard improves web navigation for AI agents via Dynamic Dual-Policy Optimization (DDPO) to handle reward conflicts and Confidence-Guided Adaptive Navigation Reflection (CANR) to calibrate per-step errors. The framework achieves state-of-the-art performance on standard web navigation benchmarks.

Summary generated by Claude — human-verified