xue-26/SAWM

VISIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

SAWM (Safety-Augmented World Model) is a vision-language model developed by Xue Yu, Bo Yuan, Pengshuai Yang, Kailin Zhao, Hong Hu, and Junlan Feng, fine-tuned from Qwen3-VL-8B-Instruct. It functions as a plug-and-play, pre-execution safety framework for mobile GUI agents, predicting action consequences and assessing risk. The model specializes in instruction-level screening and action-level risk assessment to prevent unsafe operations on mobile devices.

Loading preview...

SAWM: A Safety-Augmented World Model for Mobile GUI Agents

SAWM is the core component of SeerGuard, a novel pre-execution safety framework designed to protect mobile GUI agents from performing risky operations on real devices. Developed by Xue Yu et al., this model is fine-tuned from Qwen3-VL-8B-Instruct and acts as an external guard, integrating seamlessly without requiring modifications or retraining of the protected agent.

Key Capabilities

  • Instruction-Level Screening: Blocks malicious or unauthorized user requests before task execution begins, ensuring initial safety alignment.
  • Semantic Next-State Prediction: Given a mobile GUI screenshot, user instruction, and a candidate action, SAWM predicts the likely functional consequence in natural language (e.g., "a payment confirmation will be submitted").
  • Action-Level Risk Assessment: Evaluates the predicted consequence for safety, providing a rationale and labeling the action as safe or unsafe. This proactive approach prevents device state changes from unsafe actions.
  • Agent-Agnostic Protection: Can be paired with various open- or closed-source GUI-agent backbones, offering flexible deployment.

Performance Highlights

SAWM demonstrates strong performance across safety benchmarks:

  • On MobileSafetyBench, it significantly improves the safety-utility trade-off for GUI-agent backbones like Qwen3-VL, GPT-5.1, and Gemini-3.1.
  • Achieves an F1 score of 0.922 on the Prompt Injection benchmark for instruction-level screening.
  • Attains an F1 score of 0.723 and a Step Score of 0.361 on MobileRisk for action-level assessment, outperforming compared methods.
  • Reaches 0.762 accuracy on Next-State-QA for semantic next-state prediction.

Intended Use Cases

SAWM is ideal for research and development in:

  • Proactive safety monitoring for mobile GUI agents.
  • Filtering malicious instructions and assessing pre-execution action risks.
  • Evaluating safety across different GUI-agent backbones and analyzing safety-utility trade-offs.