often produces rollout trees where many trajectories share early interaction histories. Training each root-to-leaf trajectory independently recomputes those shared prefixes, which is wasteful. HARTS treats the tree as a compact ordering, plans microbatches, data-parallel replica assignments, and microbatch-slot schedules, then executes all branches in a round as one packed call. It supports both full-attention and linear-attention layers, propagates gradients through differentiable state handoffs, and works with activation recomputation.
On an Agentic RL workload derived from SWE-bench tasks, HARTS reports 4.81-4.87x forward, backward, and gradient speedup across multiple parallel configurations. Its numerical differences are comparable to normal self-rerun variation in the baseline, and its reward trend matches the baseline over the first 120 steps of τ^3-Bench training. The design also preserves MoE routing semantics without repeating expert computation.