ReTerm: A Self-Improving Terminal Agent Recovery Training Framework Released
Summary:ReTerm is a novel terminal agent designed to improve robustness by learning to recover from its own failed tasks. It leverages a customized data pipeline to create ReTerm-10K, a dataset for Supervised Fine-Tuning (SFT) cold start, and introduces ReTerm-GRPO, an online Reinforcement Learning (RL) algorithm that generates multiple self-recovery attempts from failure states. Experiments demonstrate that ReTerm-9B achieves a Pass@1 score of 34.8% on Terminal-Bench 2.0, outperforming the base model b
Technical Mechanism Analysis
The core innovation of ReTerm lies in its novel training mechanism that enables the agent to learn and improve from its own failures. The key components include:
-
ReTerm-10K Dataset Construction: To support the Supervised Fine-Tuning (SFT) cold start of the agent, the team constructed a high-quality dataset, ReTerm-10K, which contains numerous failure cases and corresponding recovery solutions for terminal tasks. This dataset provides essential learning material for the agent.
-
ReTerm-GRPO Algorithm: ReTerm-GRPO is an online Reinforcement Learning (RL) algorithm that generates multiple independent recovery attempts from the agent's own failure states and learns from whether these attempts succeed. This approach allows the agent to continuously improve its recovery strategies and enhance its overall robustness.
-
Qwen3.5-9B Base Model: ReTerm is built on the Qwen3.5-9B model, leveraging its strong language understanding capabilities to parse terminal tasks and generate recovery strategies.
Engineering Trade-offs and Performance
ReTerm's design balances the improvement of recovery capabilities with training efficiency and resource consumption:
-
Training Efficiency: By using an online RL algorithm, ReTerm can dynamically generate recovery attempts during training, avoiding the limitations of traditional methods that require large amounts of pre-annotated data. However, this also introduces additional computational overhead during training.
-
Recovery Rate Improvement: Experimental results on Terminal-Bench 2.0 show that ReTerm-9B achieves a Pass@1 score of 34.8%, outperforming the base model by 14.6 percentage points. This indicates that ReTerm has a higher recovery rate when dealing with diverse error categories.
-
Resource Consumption: Although ReTerm requires more computational resources during training, its improved recovery capabilities may reduce the need for repeated computations due to task failures in practical applications, ultimately achieving resource savings.
Developer Implementation and Deployment Recommendations
For developers, the release of ReTerm provides the following insights:
-
Data-Driven Agent Training: The success of ReTerm demonstrates the importance of building high-quality datasets for agent training. Developers can learn from the construction method of ReTerm-10K and build corresponding datasets for specific application scenarios.
-
Reinforcement Learning and Self-Improvement Mechanisms: The ReTerm-GRPO algorithm showcases the potential of reinforcement learning in agent self-improvement. Developers can explore the application of similar mechanisms to other types of agents to enhance their adaptability and robustness.
-
Base Model Selection and Fine-Tuning: ReTerm is built on the Qwen3.5-9B model, demonstrating the importance of base model selection. Developers should choose the appropriate base model based on the requirements of specific tasks and perform targeted fine-tuning to achieve optimal performance.
Conclusion
The release of ReTerm marks a significant advancement in the robustness and self-improvement capabilities of terminal agents. Its innovative training mechanism and significant experimental results provide a new direction for agent research and offer strong support for the deployment of agents in practical applications.
— END —Source: Hugging Face Daily Papers (2026-10-13)
Tags: #ReTerm #Terminal Agent #Reinforcement Learning #Self-Improvement #Qwen
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