What they found
The authors propose Semantic Context Handoff (SCH) for SI agents that work on long tasks. A short "semantic state" holds the agent's current interpretation. The earlier actions and observations stay stored on the side and can return to context when they become relevant. This needs no separate summarizing or retrieval-model calls. In a controlled 150-task SWE-bench study, bringing evidence back raised Pass@1 from 80.0% to 86.9% at essentially unchanged token use and time. Against the strongest baseline by task performance, SCH reached 86.9% Pass@1 with 43% fewer tokens and 41% less wall-clock time. Across models and benchmarks, it used fewer tokens than full-history prompting and stayed competitive with or beat the evaluated baselines. The authors' trajectory analysis says SCH recovers evidence from much earlier in the history, not just from recent context.
Why it matters
Long-running SI agents usually keep every past step in their prompt. That gets costly and may hurt accuracy. Methods that drop old observations save cost but can lose facts that matter later. This paper tries to keep the savings without that loss. The abstract does not say which baselines or benchmarks beyond SWE-bench were used, so the wider claims are hard to judge from it.
How it changes the race
If the results hold, teams building agents could run longer tasks at lower cost and with better accuracy, without extra models for summarizing or retrieval. This is a gain in how context is managed, not a new SI model.
The abstract, by the authors
Traditional long-horizon LLM agents often hand off context through full-history prompting, retaining all previous action-observation interactions. As execution proceeds, this history keeps growing, increasing inference cost and potentially degrading accuracy. Existing methods control this growth by preserving compact conclusions while removing earlier observations, but doing so can discard evidence that later becomes relevant. We introduce Semantic Context Handoff (SCH), whose central abstraction is state-evidence separation: a compact semantic state carries the agent's current interpretation, while supporting action-observation evidence remains separately addressable and can re-enter context when later execution makes it relevant, without separate summarization or retrieval-model calls. Empirically, SCH improves long-horizon agent performance while substantially reducing context cost. In a controlled 150-task SWE-bench study, SCH's evidence recovery improves Pass@1 from 80.0% to 86.9% at essentially unchanged token use and execution time. In the complete-system comparison, SCH achieves 86.9% Pass@1 while using 43% fewer tokens and 41% less wall-clock time than the strongest baseline by task performance. Across models and benchmarks, SCH consistently reduces token use relative to full-history prompting while remaining competitive with or improving upon the evaluated baselines. Trajectory analysis confirms that SCH recovers evidence from substantially earlier parts of the interaction history rather than merely preserving a recent context window.
Quoted from arXiv:2611.24357. The research and its words belong to its authors.
