DeLoop cuts Qwen3-8B agent loop rate from 57.3% to 6.1%

Language-model agents under context compression spend a lot of their time redoing work they have already done. On CodeGym under compression, 57.3% of Qwen3-8B agent tasks contained at least one "loop" (a tool call that re-issues an earlier action and gets back exactly the same observation), according to a new paper on arxiv from a team led by Tianle Chen.

The authors state the failure flatly: "Language-model agents that run long tasks often redo work they have already done: they re-issue a tool call they made earlier and get back exactly the same observation, a step that adds no new information." They write that loops "appear natively in Qwen, Llama, and Gemini agents and become far more common once the agent's history is compressed to fit its context window."

Their proposed fix, DeLoop, trains on the agent's own rollouts. A looping action is used as a negative example and the first later non-looping action in the same rollout as the positive. Direct preference optimization does the rest. No task rewards, no gold answers, no LLM judge.

The reported numbers: on CodeGym under compression, the share of Qwen3-8B tasks containing loops drops from 57.3% to 6.1%, with "about one third as many steps per task, with similar task success." On Qwen3-32B, task success rises from 37.7% to 44.8% while loops and steps both fall.

The abstract reports no per-model numbers for Llama or Gemini.