CHILL-Harness: Counterfactual Harness Learning for Efficient Reasoning in Long-Horizon Agents
CHILL-Harness intervenes at the orchestration layer to enable advantage-guided workflow adaptation, thereby improving reasoning and execution efficiency while preserving task performance and incorporating a success-preserving objective and advantage-margin authorization constraints into CHILL-Harness to promote reliable adaptation.