Search papers, labs, and topics across Lattice.
SCAD addresses the challenge of training long-horizon agents by structuring interactions into planning and bounded subtask execution phases, enabling more effective credit assignment and distillation. It distills execution within local contexts while refining planning credit using cross-rollout subtask prefix trees, ensuring planning receives full terminal credit and execution benefits from both terminal and teacher guidance. This approach yields significant improvements, increasing macro-average accuracy by 4.48 pp on text tasks and 4.19 pp on multimodal tasks compared to strong baselines.
Long-horizon agents achieve state-of-the-art performance by decoupling planning credit from execution guidance via structured subtask trees.
Training long-horizon agents to solve complex tasks requires effective supervision over extended interaction sequences. However, sparse terminal rewards obscure intermediate contributions, while on-policy distillation can lose informative teacher guidance as student-generated histories grow. To address this problem, we introduce SCAD, which organizes interactions into planning and bounded subtask execution, distills execution in local contexts, and refines planning credit through cross-rollout subtask prefix trees, with planning receiving full terminal credit and execution receiving positive terminal credit and teacher guidance. Across all evaluated benchmarks, SCAD improves macro-average accuracy over the strongest training baseline by 4.48 percentage points for text tasks and 4.19 points for multimodal tasks. SCAD effectively combines outcome-based credit assignment with teacher-guided distillation to improve planning and execution in long-horizon agents.
