DeLoop cuts repeated tool-call loops in long-horizon agents
Original title: DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents
Under CodeGym compression, Qwen3-8B tasks containing loops fall from 57.3% to 6.1%.
The authors define a loop as re-issuing a tool call that returns the same observation. Loops appear in Qwen, Llama, and Gemini agents and become more common when history is compressed. DeLoop pairs a looping action with a later non-looping action and trains with direct preference optimization, without rewards, gold answers, or LLM judges.
Unconfirmed: Loop rates and success changes are reported by the authors.
- Qwen3-8B uses about one third as many steps, with similar success.
- Qwen3-32B task success rises from 37.7% to 44.8%.
- Supervision comes only from the agent's own rollout pairs.
- History compression makes loops far more common.
From the original article
1 min read at huggingface.co
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. We identify this failure as a loop. 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.
By Tianle Chen, Pushkar Singh, Ketan Todi · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.