Semantic Context Handoff Keeps Long-Horizon Agent Evidence Addressable
The paper introduces Semantic Context Handoff (SCH), which separates a compact semantic state (the agent's current interpretation) from the underlying action-observation evidence, so that evidence can re-enter context when later execution makes it relevant without separate summarization or retrieval-model calls. In a controlled 150-task SWE-bench study, SCH's evidence recovery improved Pass@1 from 80.0% to 86.9% at essentially unchanged token use and execution time. In the complete-system comparison, SCH reached 86.9% Pass@1 while using 43% fewer tokens and 41% less wall-clock time than the strongest baseline by task performance. Trajectory analysis indicates SCH recovers evidence from substantially earlier in the interaction history rather than only preserving a recent context window.
It offers agent builders a concrete context-management technique with measured accuracy gains and token/time savings on SWE-bench.
Comments & Questions (0)
No comments yet
Be the first to comment!