Hugging Face Releases DeLoop: A New Method to Reduce Redundant Actions in LLM Agents
Summary:Hugging Face introduces DeLoop, a novel self-supervised training method designed to reduce redundant actions in long-horizon language-model agents. DeLoop leverages the agent's own rollout as supervision, treating looping actions as negative examples and the first subsequent non-looping action as a positive example. Learning occurs through direct preference optimization without requiring task rewards, gold answers, or LLM judges. In experiments on CodeGym under compression, DeLoop reduced the sh
Technical Mechanism Analysis
The core mechanism of DeLoop is based on self-supervised learning, utilizing the agent's own execution trajectory to identify and reduce redundant actions. The main steps are as follows:
- Redundant Action Identification: The agent records its action trajectory during task execution. When a repeated action sequence is detected, it is marked as a negative example.
- Positive Example Selection: The first non-repeated action following a repeated action is marked as a positive example.
- Direct Preference Optimization: By contrasting positive and negative examples, DeLoop uses direct preference optimization to train the agent, enabling it to learn a strategy to avoid redundant actions.
The advantage of this approach is that it does not rely on external reward signals, gold answers, or LLM judges, but rather depends solely on the agent's own execution data, thereby reducing training costs and complexity.
Engineering Trade-offs and Performance
In the CodeGym benchmark test, DeLoop demonstrated excellent performance on both Qwen3-8B and Qwen3-32B models:
- Reduction in Redundant Actions: For Qwen3-8B, the proportion of tasks with loops decreased from 57.3% to 6.1%.
- Reduction in Steps: The number of steps per task was reduced to about one-third of the original.
- Increase in Task Success Rate: For Qwen3-32B, the task success rate increased from 37.7% to 44.8%.
These results indicate that DeLoop can significantly reduce redundant actions while maintaining or even improving task success rates. This is particularly important for tasks that require long execution times, as it can significantly reduce the consumption of computing resources.
Developer Implementation and Deployment Recommendations
For developers, DeLoop provides a simple yet effective method to improve the execution efficiency of agents. Here are some deployment recommendations:
- Integration into Existing Systems: DeLoop can be integrated as an independent module into existing agent systems without the need for major modifications to the existing architecture.
- Custom Optimization: Developers can adjust the training parameters of DeLoop according to specific task requirements to achieve the best performance.
- Combination with Other Technologies: DeLoop can be combined with other optimization technologies (such as pruning, quantization, etc.) to further improve the efficiency of agents.
Conclusion
DeLoop demonstrates the great potential of self-supervised learning in reducing redundant actions in agents. It not only improves task execution efficiency but also reduces the consumption of computing resources, providing new ideas for the practical application of large language model agents.
— END —Source: Hugging Face Daily Papers (2026-10-14)
Tags: #Hugging Face #LLMs & Foundation Models #Self-Supervised Learning #Agent Optimization
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments