STITCH: Less-Is-More Training for Software Engineering Agents via Sliding-Memory Trajectory Chunking
arXiv·medium signal
Proposes STITCH (Sliding-memory Trajectory Inference and Task Chunking Heuristic), an end-to-end framework achieving superior agentic capabilities with fewer but higher-quality training trajectories. Extends the 'Less-Is-More' hypothesis from math reasoning to agentic scenarios. Reduces data construction costs for training SWE agents while improving performance — practical for teams building custom coding agents.