ZICQ
中 Log in / Sign up
Newsroom LLMs & Foundation Models #Hugging Face #SWR Framework #Terminal Agents #Self-supervised Learning #Scientific Domains

Hugging Face Introduces SWR Framework: Revolutionizing Terminal Agent Training in Scientific Domains

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 3 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:Hugging Face has introduced the Software-in-the-Loop Reconstruction (SWR) framework, a novel self-supervised approach designed to address key challenges in training terminal agents for scientific domains. By extracting reference outputs and verification targets from existing software workflows, SWR leverages editable programs to construct training data, reducing reliance on human annotations. Experiments demonstrate that the SWR method, powered by Qwen3.8-Max, significantly enhances task-solving


Background and Challenges

As terminal agents are increasingly deployed in scientific and other specialized domains, the process of constructing training environments has become more complex. Traditional methods rely on executable behavioral references and domain-specific verifiers, which not only require substantial human engineering but also limit cross-task reusability.

Innovations of the SWR Framework

The SWR framework introduced by Hugging Face revolutionizes the agent training paradigm through the following innovations:

  • Self-supervised Learning: Automatically extracts reference outputs and verification targets from existing software workflows, reducing reliance on human annotations.
  • Editable Program Construction: Based on instructions, input schemas, and public input-output observations, the agent constructs an editable program without accessing the source workflow.
  • Hierarchical Verifier: Combines domain-specific semantic comparison, structural validity, and anti-shortcut checks, while leveraging public feedback to support iterative revision.

Experimental Results

In the experiments, the SWR framework utilized 500 workflows and 46 software families across six domains. Qwen3.8-Max, attempting each task three times, solved 838 tasks and generated 1,422 verified trajectories, which were oversampled to 3,000 reconstruction-only training examples. Supervised fine-tuning of Qwen3.8-27B improved the mean Terminal-Bench 2 performance from 47.94% to 53.56% across three seeds and achieved the highest mean among four matched-token corpus controls on all four reported evaluations.

Industry Impact and Developer Recommendations

  • AI Applications in Science: The SWR framework provides a more efficient and reliable training method for AI applications in scientific fields, helping to accelerate the adoption of AI in scientific research.
  • Developer Tool Support: Developers can leverage the SWR framework to build smarter terminal agents, reducing reliance on human annotations and improving training efficiency.
  • Cross-Task Reusability: The modular design of the framework allows for easy extension to new workflows and configurations, supporting cross-task reusability.

Future Outlook

The release of the SWR framework marks a significant advancement in the AI training paradigm, particularly in scientific domains that require high specialization and complex verification. As more workflows and configurations are supported, SWR is expected to play a broader role in a wider range of applications and scenarios.


Source: Hugging Face Daily Papers (2026-10-02)

— END —

Tags: #Hugging Face #SWR Framework #Terminal Agents #Self-supervised Learning #Scientific Domains

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.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…