| name | step-gui-calibrated-rewards |
| title | Step-GUI: Calibrated Step Rewards and Self-Evolving Training for Autonomous GUI Agents |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2512.15431 |
| keywords | ["GUI-automation","reinforcement-learning","reward-calibration","self-evolving","agent-training"] |
| description | Train GUI agents through calibrated step-level reasoning anchored to trajectory-level validation. Use trajectory verification rather than step annotation achieving 90% accuracy with 10-100× cost reduction. Implement self-evolving training framework with generation-refinement cycles and verifiable reward signals. |
Skill Summary
Step-GUI presents three interconnected innovations for GUI automation. The core contribution is a Calibrated Step Reward System (CSRS) that anchors LLM-generated dense step-level reasoning to trajectory-level evaluation signals through validation rather than step-level annotation, achieving 90% accuracy with 10-100× cost reduction. A self-evolving training framework enables continuous improvement through generation-refinement cycles. Supporting infrastructure includes GUI-MCP (hierarchical protocol for LLM-device interaction) and AndroidDaily benchmark for real-world evaluation.
When To Use
- Training autonomous GUI agents for mobile or web automation
- Projects where annotation cost is prohibitive for step-level supervision
- Scenarios requiring self-improving agent training loops
- Research on efficient reward calibration for RL
When NOT To Use
- Scenarios with abundant step-level annotations making calibration unnecessary
- Real-time applications where training overhead isn't justified
- Domains without reliable trajectory-level evaluation signals (success/failure)
- Simple scripted automation already meeting requirements
Core Technique
Three key innovations enable efficient agent training:
1. Calibrated Step Reward System (CSRS)
Convert expensive annotation problem into scalable data-refinement process. Rather than annotating each step, use trajectory-level validation (automated verifiers or human judges confirming task success/failure), then apply thinking models to generate rich chain-of-thought reasoning for successful trajectories. Achieves ">90% annotation accuracy with 10-100× cost reduction."
2. Self-Evolving Training Framework
Implement three-stage progressive training paradigm:
- Mid-Training: Generate rollouts and refine via rejection sampling
- Cold-Start Fine-Tuning: Initialize from successful trajectories
- Reinforcement Learning: Optimize with verifiable reward signals
Orchestrate two parallel data flows—generation (from model rollouts) and refinement (via rejection sampling)—enabling continuous performance enhancement.
3. Supporting Infrastructure
- GUI-MCP: Hierarchical protocol standardizing LLM-device interaction with privacy protection through local specialist model execution
- AndroidDaily: Evaluation benchmark grounded in real-world usage patterns with 3,146 static actions and 235 end-to-end tasks
Implementation Notes
Establish trajectory-level evaluation: automated verifier or human judgment confirming success/failure. Apply thinking model to generate step-level reasoning chains for successful trajectories. Implement three-stage training: mid-training with rejection sampling, cold-start fine-tuning, and RL with verifiable rewards. Use hierarchical protocol for LLM-device interaction.