Skill Graph Auditing: Enabling Verifiable and Transparent Continual Self-Improvement in XSkill Agents

TL;DR: Could XSkill’s learned skills and experiences be organized into an auditable, transparent skill graph, so that improvements can be traced, verified, and trusted—especially for safety-critical applications? By fusing XSkill’s dual-stream knowledge with ASG-SI’s skill-graph and verifier-backed improvement pipeline, we can create agents whose learning is both powerful and trustworthy.

Research Question: Can the explicit organization and auditing of XSkill’s learned skills and experiences as a verifiable skill graph improve the transparency, reproducibility, and safety of continual self-improving multimodal agents?

Hypothesis: Agents whose improvements are compiled into verifiable, auditable skill graphs, with replay-backed validation, will not only match or exceed the adaptability of standard XSkill agents, but will also offer superior traceability and operational governance.

Experiment Plan: - Setup: Extend XSkill with the ASG-SI approach, converting distilled skills and experiences into a growing, explicitly structured skill graph, with verifier-backed promotion and audit logging.

• Benchmarks: Use complex, open-ended domains with high safety or compliance requirements (e.g., simulated industrial or medical tool use).
• Comparisons: Evaluate XSkill, XSkill+ASG-SI, and baseline self-improving agents lacking explicit audit mechanisms.
• Measurements: Track performance, frequency of reward hacking or behavioral drift, auditability (ability to trace skill origin and improvement), and reproducibility.
• Expected Outcome: XSkill+ASG-SI will achieve similar or better continual learning efficiency, while enabling robust verification and traceability of all learned improvements.

References:

• Jiang, G., Su, Z., Qu, X., & Fung, Y. R. (2026). XSkill: Continual Learning from Experience and Skills in Multimodal Agents.
• Huang, K., & Huang, J. (2025). Audited Skill-Graph Self-Improvement for Agentic LLMs via Verifiable Rewards, Experience Synthesis, and Continual Memory. arXiv.org.