Research · Oct 22, 2026, 00:00 UTC

DeLoop uses an agent's own repeated actions as self-supervision to cut loops

DeLoop trains agents with direct preference optimization on loop and non-loop action pairs from their own rollouts, without task rewards or judges. On CodeGym under compression, Qwen3-8B tasks with loops fall from 57.3% to 6.1%, and Qwen3-32B task success rises from 37.7% to 44.8%.

• LabAlibaba
• LabMeta
• LabGoogle
• ModelQwen3-8B
• ModelQwen3-32B
• ModelLlama
• ModelGemini
• Topicresearch
• Topicagents