DeLoop Uses Self-Supervision to Reduce Redundant Actions in Long-Horizon Agents

DeLoop is a self-supervised training method that uses an agent's own repetitive tool calls as negative examples to prevent loops. When applied to Qwen3-8B under context compression, it significantly reduces the frequency of redundant, non-informative actions.