SujiKim/learnweak-evocua-8b-lora-r32-libreoffice-writer

LearnWeak: Automated Domain Specialization for Small Computer-Use Agents

This repository contains a domain-specialized LoRA adapter for EvoCUA-8B, developed as part of the LearnWeak framework.

LearnWeak is an annotation-free specialization framework for small computer-use agents (CUAs). It uses a stronger reference agent to identify a student model's weaknesses in a target domain, synthesizes targeted tasks, and constructs supervision automatically. This model focuses on specializing the agent for desktop software environments.

• —Project Page: https://learnweak.github.io/
• —Repository: https://github.com/sujiikim/LearnWeak
• —Paper: Learn from Weaknesses: Automated Domain Specialization for Small Computer-Use Agents

Model Description

Small open computer-use agents are practical specialization targets but often exhibit domain-specific failures. LearnWeak introduces an error-aware specialization objective that disentangles planning and execution errors, enabling more behaviorally precise updates. On OSWorld, LearnWeak achieves significant performance gains across various domains such as GIMP, LibreOffice, and VS Code.

How to Get Started with the Model

Serve with vLLM

You can serve the base model with this LoRA adapter enabled using vLLM. Replace the LoRA module name and path as appropriate for the specific domain:

vllm serve meituan/EvoCUA-8B-20260105 \
  --enable-lora \
  --max-lora-rank 32 \
  --lora-modules learnweak-adapter={MODEL_ID}

Use the LoRA module name (e.g., learnweak-adapter) when calling the served model's API.

Training Details

• —Base Model: meituan/EvoCUA-8B-20260105
• —Framework: PEFT (LoRA)
• —Rank: 32
• —Alpha: 64
• —Target Modules: qproj, downproj, kproj, upproj, oproj, vproj, gate_proj

Citation

If you find this work useful, please consider citing:

@article{kim2026learnweaknessesautomateddomain,
  title   = {Learn from Weaknesses: Automated Domain Specialization for Small Computer-Use Agents},
  author  = {Kim, Suji and Kim, Kangsa and Hwang, Sung Ju},
  journal = {arXiv preprint arXiv:2605.28775},
  year    = {2026}
}

Acknowledgments

This project builds on OSWorld, LlamaFactory, and EvoCUA.