ACC: Compiling Agent Trajectories for Long-Context Training

The article introduces Agent Context Compilation (ACC), a method for transforming agent-generated trajectories into long-context QA pairs, enhancing long-context reasoning in large language models (LLMs). By integrating responses and observations from multiple turns, ACC allows for effective training without additional annotation, yielding significant improvements in long-range dependency tasks—achieving a score of 68.3 on MRCR and 77.5 on GraphWalks with the Qwen3-30B-A3B model. This approach is scalable and can be combined with existing long-context training methods, making it valuable for practitioners focused on enhancing LLM capabilities in complex reasoning scenarios.