ContextPilot framework enables 8B model with 32K context to outperform 128K baseline

Researchers from Tsinghua University, Tencent YouTu Lab, and Shanghai AI Lab introduced ContextPilot, a framework that enables agents to actively manage their own context. It addresses three deficiencies in prior work: narrow tool sets, uniform exploration in RL training, and coarse credit assignment. The framework adds planning, long-term memory, and soft offloading tools; synthesizes SFT data with a teacher model scaffold; and introduces context-aware partial rollback with fine-grained credit assignment in RL. An 8B model with ContextPilot and 32K context window achieved 69.4 average on long-document QA, outperforming the 128K-window baseline Qwen3-8B at 45.9. Token efficiency improved from ~30K to 8-10K per turn in deep search tasks.