WiSI: Wireless-in-the-Loop Self-Improvement
eess.SY, cs.SY
Submitted: 2026-11-03
Updated: 2026-11-03
Comments: The code is available at \url{https://github.com/NICE-HKU/WiSI}
Code: https://github.com/NICE-HKU/WiSI
License: http://creativecommons.org/licenses/by/4.0/
The gist: The growing autonomy of AI agents has enabled self-improvement, where agents iteratively refine their programs through execution feedback and performance evaluation.
Terminology
Abstract
The growing autonomy of AI agents has enabled self-improvement, where agents iteratively refine their programs through execution feedback and performance evaluation. As these agents increasingly operate in physical environments and rely on wireless connectivity, communication conditions become an integral part of their improvement process. However, existing approaches largely assume ideal communication, overlooking how wireless conditions shape both the feedback that guides program modifications and the evaluation outcomes that determine which modifications survive. This challenge is further complicated by the coupling between agent behavior and wireless connectivity, as agents influence their own communication conditions through mobility and transmission decisions. To address this problem, we propose WiSI, a wireless-in-the-loop self-improvement framework that integrates spatiotemporal wireless channels, edge-assisted perception, and Large Language Model-driven program improvement. By incorporating wireless execution into the improvement loop, WiSI reveals how wireless execution experience shapes the generation and retention of program structures, and investigates the wireless conditions needed for robust improvement. Experiments show that agents improved under ideal links achieve only 40.0% task success under congestion, compared with 77.2% when improved under realistic wireless conditions. More importantly, wireless-aware feedback and selection jointly enable agents to discover and retain communication-efficient program structures that outperform parameter tuning and remain robust even with inaccurate run-time channel observations. We further find that moderate-load or mixed-network evaluation maintains strong performance across deployment conditions, while simplified models preserving link-rate statistics achieve comparable results without detailed spatial channel information.
Sources
- Dream-RSI: Recursive Self-Improvement through Evolving Worlds
- AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents
- RoboFoundry: System-as-Policy Evolution for Self-Learning Embodied Agents
- Self-Evolving Embodied Agents via Skill-Harness Evolution
- When Non-Adversarial Noise Hurts Learning: Benchmarking Skill Self-Evolution in Agents
- AI2-THOR: An Interactive 3D Environment for Visual AI
Related papers
- Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems
- Exact and Approximate Convex Reformulation of Linear Stochastic Optimal Control with Chance Constraints
- Certified Set Convergence for Piecewise Affine Systems
- Symbolic Control for Stochastic Systems via Finite Parity Games
- Multiobjective optimization-based design and dispatch of islanded, hybrid microgrids for remote, off-grid communities in sub-Saharan Africa
- Physiology-informed layered sensing for intelligent human-exoskeleton interaction