DeLoop uses repeated actions as self-supervision for long-horizon agents

What the research found

DeLoop is a self-supervised preference optimization method (published Oct 14) that cuts repeated tool-call loops in agents; on CodeGym under compression it lowers Qwen3-8B loop share from 57.3% to 6.1% and lifts Qwen3-32B success from 37.7% to 44.8%.

Claims and how well the evidence supports them

1. DeLoop reduces Qwen3-8B tasks containing loops from 57.3% to 6.1% on CodeGym under compression 99%
2. DeLoop raises Qwen3-32B task success from 37.7% to 44.8% 99%
3. DeLoop needs no task rewards, gold answers, or LLM judges 99%

Evidence

• huggingface.co https://huggingface.co/papers/2610.19376

Actionability rated low.