Interesting paper from Tencent.

Tencent trains an agent to manage its own working context, and assigns credit at the level of individual context edits.

Long-horizon tasks force a model to retrieve, integrate and maintain scattered information across many turns, and keeping every interaction history makes the working context grow without bound.

Recent proactive methods let a model edit its own context with tools, but the toolset stops at search, deletion and summarization.

ContextPilot adds global planning, long-term memory and adaptive soft compression, so the agent can offload information rather than only discard it.

The training side is where it gets interesting. Standard RL hands the final trajectory reward to every intermediate edit equally. ContextPilot uses context and entropy variation to find which editing decisions actually mattered, samples branches at those points, and estimates action-level advantages from all branched trajectories passing through that edit.

On long-context QA and deep search it beats existing baselines across several base models while holding a more compact working context. Code is available.

Paper: https://t.co/buM6galu4x

Chat with Paper: https://t.co/VN5J5PO4Rt