RHyVE Framework: Building Robust Computer Agents via Capability-Aware Validation and Phase-Aware Training

This article introduces the RHyVE framework, which aims to address the reliability gap of computer agents from demonstration to production. The framework achieves continuous iteration of validation and training through the collaboration of two components: Capability-Aware Validation (CAV) and Phase-Aware Training (PAT). Its core goals are to clarify the agent's capability boundaries, dynamically adjust learning strategies, and provide solutions for building trustworthy computer agents.

Computer agents perform well in controlled environments, but real-world deployment faces the "demonstration-production gap": ambiguous and variable tasks, dynamic environments, and ignored tail risks. Existing methods have three major limitations: lack of systematic capability boundary assessment, one-size-fits-all training strategies, and disconnection between validation and training.

The core of CAV is the concept of capability boundaries (high-confidence reliable regions in the task space). It uses multi-dimensional evaluation methods: functional testing (validation of basic operations), adversarial testing (finding failure inputs), out-of-distribution detection (identifying unfamiliar scenarios), and compositional generalization testing (skill combination ability). It also classifies failure modes (perception, reasoning, execution, context failures) and outputs capability reports for pre-deployment evaluation and runtime monitoring.

PAT divides training into three phases: Exploration (extensive exploration of the task space), Specialization (optimizing effective strategies), and Generalization (enhancing cross-scenario capabilities). It dynamically adjusts hyperparameters: exploration rate (high to low), learning rate (high to low then periodic adjustment), curriculum learning (increasing difficulty), and reward shaping (sparse to dense then adversarial). Phase transitions are detected through performance saturation, strategy stability, and capability coverage.

Synergy between CAV and PAT: CAV feedback guides PAT adjustments (e.g., increasing the proportion of failed tasks), and PAT progress is validated via CAV to form a closed loop. Experiments show: RHyVE has comparable or slightly higher success rates on OSWorld/WebArena benchmarks, with less performance degradation in adversarial/out-of-distribution tests; improved training efficiency (faster convergence, optimized resources, stability); and more compact and reliable capability boundaries.

Deployment is suitable for human-machine collaboration (graded handling of low/medium/high risks) and supports continuous learning (retraining with real-world experience feedback). Limitations: CAV relies on test coverage, phase division can be more granular, and computational overhead is high. Future directions: formal verification, meta-learning, and multi-agent validation.