TraceR1: Anticipatory Planning via Two-Stage RL Achieves Substantial Gains Across 7 Agent Benchmarks (CVPR 2026)

arXiv 2603.16777·high signal

TraceR1 introduces anticipatory trajectory reasoning — agents forecast short-horizon action sequences before execution rather than acting reactively. A two-stage RL framework first trains trajectory-level consistency, then applies grounded fine-tuning using execution feedback from frozen tool agents. Evaluated across 7 benchmarks covering online/offline computer-use and multimodal tool-use, TraceR1 achieves substantial improvements in planning stability, execution robustness, and generalization; accepted to CVPR 2026 Findings Track.