Lihuan Li and colleagues separate an assistant’s progress record from the evidence it needs to act. Their CarryOn system keeps a short account of completed and unfinished work. A trained selector retrieves original observation fragments, preserving identifiers and records that a summary might omit.

Omitted fragments remain in an archive and can return later. The selector ranks evidence within a base budget, then temporarily adds records predicted to matter for the next action. The base budget is therefore not a hard prompt-size limit.

  • A 0.6-billion-parameter selector is trained separately for each environment, then reused across three acting models. It adds local computation without extra text-generation calls.
  • Across AppWorld, OfficeBench and Tau2 Telecom, the authors report 75.7% mean task success, 6.7 percentage points above their strongest overall baseline, SWE-Pruner.
  • Resource scores include failed tasks but exclude local selection and pruning computation. They measure calls and tokens, not all-in hardware cost.
  • On AppWorld with Claude, removing progress state lowers success from 79.8% to 66.1%. Replacing learned selection with recent-evidence packing still achieves 76.2%.

The learned selector cannot explain every gain. In Gemma’s telecom runs, it is never invoked: the full history fits inside the base evidence budget. Across all three telecom models, only 13 of 120 trajectories invoke learned selection. Better progress instructions and context construction matter too.

Caveat: Each setting has one reported run, and transfer to unfamiliar applications remains untested. The paper’s anonymous code page currently cannot load its contents, so an installable release is unverified.

This is a different team and method from the October CarryOn paper. That earlier work organizes context around reusable objects; this paper trains an evidence selector. They should retain separate identities despite sharing a name.

Daily coverage