ACC: Compiling Agent Trajectories for Long-Context Training

Researchers introduced Agent Context Compilation (ACC), which turns multi-turn agent traces into supervised QA data for long-context reasoning.

ACC compiles trajectories from search, coding, and database agents so evidence spread across turns can be learned directly.

The method improves performance on long-range dependency benchmarks and distant-context integration tasks.

It also preserves general model abilities, addressing a key training gap in agent fine-tuning.