CarryOn: Learning What to Keep in Context for Long-Horizon LLM Agents
CarryOn reports 75.7% task success across three benchmarks, beating the strongest baseline by 6.7 percentage points.
The context manager pairs a compact execution state with fine-grained evidence linked to its source. A learned selector and adaptive retention policy preserve relevant details without additional LLM calls for context management. The authors tested it on AppWorld, OfficeBench and Tau2 Telecom using three LLM actors. CarryOn also earned the best average scores on two proposed metrics measuring success alongside resource use and resources spent per successful task.
ArXiv · AI/CL/LG's note
The context manager pairs a compact execution state with fine-grained evidence linked to its source. A learned selector and adaptive retention policy preserve relevant details without additional LLM calls for context management. The authors tested it on AppWorld, OfficeBench and Tau2 Telecom using three LLM actors. CarryOn also earned the best average scores on two proposed metrics measuring success alongside resource use and resources spent per successful task.
ArXiv · AI/CL/LG's note
score 5