XNews

ContextPilot: 8B Model with 32K Context Beats 128K Self via Active Memory Management

ContextPilot: 8B Model with 32K Context Beats 128K Self via Active Memory Management

Researchers from Tsinghua University, Tencent YouTu Lab, and Shanghai AI Lab introduced ContextPilot, a framework enabling agents to actively manage context. It adds planning, long-term memory, and soft offloading tools. SFT data was synthesized with Qwen3.5-397B teacher, filtered to ~3K trajectories under 32K. The core RL method uses context-aware partial rollback to allocate exploration budget to critical editing actions and fine-grained credit assignment that averages scores from all continuations of a snapshot. ContextPilot-8B-RL averaged 69.4 on long-document QA, surpassing the 128K-window Qwen3-8B (45.9) and StateLM-8B-RL by 3.6 points. Gains were larger on harder tasks (BrowseComp+ +5.3). Ablations showed long-term memory contributed most, and fine-grained credit assignment was the most effective RL improvement.

Source posts