Google's DeLoop cuts agent loops via self-supervised DPO, lifting Qwen3-32B success to 44.8%

google · hf · 2026-10-22

Google and university researchers introduce DeLoop, a self-supervised preference optimization method that reduces repetitive tool-use loops in long-horizon LLM agents.

Problem: Agents re-issue tool calls and get identical observations back; these loops appear natively in Qwen, Llama, and Gemini agents and worsen sharply once history is compressed to fit the context window.
Method: The agent's own rollouts provide supervision — looping actions serve as negative examples and the first later non-looping action in the same rollout as positives, trained via DPO without task rewards, gold answers, or LLM judges.
Results: On CodeGym under compression, DeLoop cuts the share of Qwen3-8B tasks containing loops from 57.3% to 6.1% and uses about one-third as many steps with similar success; on Qwen3-32B it raises task success from 37.7% to 44.8%.

Code and data will be publicly released.

Original post →

Google's DeLoop cuts agent loops via self-supervised DPO, lifting Qwen3-32B success to 44.8%

More from coding & agent