Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie
Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie is an 8 billion parameter SFT fine-tuned reward model based on Qwen3-VL-8B-Instruct, specifically designed for evaluating GUI agent actions on mobile UIs. It classifies agent actions as correct or wrong by generating a discrete preference token, making it ideal for step-level correctness classification and filtering training data. This model excels as a process reward model (PRM) for mobile GUI agent trajectory evaluation.
Loading preview...
Model Overview
Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie is an 8 billion parameter vision-language model, fine-tuned from the Qwen3-VL-8B-Instruct base model. Its primary function is to act as a process reward model (PRM) for evaluating the correctness of GUI agent actions on mobile user interfaces.
Key Capabilities
- Discrete Action Evaluation: Classifies GUI agent actions by generating a specific token:
<|+|>for correct actions (score = 1.0) and<|-|>for wrong actions (score = 0.0). - Contextual Judging: Appends a judge prompt after the agent's tool-call response, allowing the model to assess actions based on the current screen state and task progress.
- SFT Fine-tuned: Trained using Supervised Fine-Tuning (SFT) with LoRA on the UI-Genie-RM-517k dataset, which includes 64k training pairs.
Intended Use Cases
This model is particularly well-suited for:
- Step-level correctness classification during GUI agent rollouts.
- Best-of-N action selection in agent decision-making processes.
- Filtering training data to improve the quality of agent training sets.
Related Models
- Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie-scoring: A related model that builds upon this one by adding a Bradley-Terry scalar reward head for pairwise evaluation.