Found first: a primary source the press has not covered yet.
A new paper finds that context compression causes 57.3% of Qwen3-8B tasks on the CodeGym benchmark to collapse into loops, where the agent re-issues an identical tool call, receives the same observation, and adds no new information. The authors introduce DeLoop, a self-supervised fix that reduces the loop rate to 6.1% using roughly one-third fewer steps while maintaining comparable task success. The paper is available on arXiv.
What the source says
DeLoop is authored by Tianle Chen, Pushkar Singh, Ketan Todi, Biswajit Paria, and seven colleagues (11 authors total), submitted October 14, 2026; institutional affiliations are not listed on the abstract page. The method works entirely from the agent's own execution traces: each looping action becomes a negative example and the next non-looping action becomes the positive, with training via direct preference optimization. No task rewards, ground-truth answers, or LLM judges are required. On Qwen3-32B, DeLoop raises task success from 37.7% to 44.8% while cutting both loop rate and total steps. The authors also observe looping behavior in Qwen, Llama, and Gemini agents without any compression applied.
Why it matters
Context compression is standard in production long-horizon agents, so a 57.3% loop rate under compression is a realistic failure mode, not a contrived benchmark artifact. The label-free training approach matters for deployment: production systems typically lack gold-labeled rollouts or reward signals, so a method that works from the agent's own behavior requires no additional annotation infrastructure. The gains transfer across model scale: on Qwen3-32B, task success rises from 37.7% to 44.8% alongside the reduction in loops and steps.