Adaptive RAN Slicing Control via Reward-Free Self-Finetuning Agents

A team of researchers has introduced a novel framework for creating self-improving AI agents capable of continuous control in complex environments like telecommunications networks. The core innovation is a 'reward-free self-finetuning' process where the agent, built on a generative AI model, learns directly from interaction. It uses a bi-perspective reflection mechanism to autonomously generate linguistic feedback on its own actions, constructing preference datasets from its history. This data is then used to fine-tune the model's own parameters, allowing it to internalize long-term experience rather than relying on a finite context window or external reward signals.

The framework was rigorously tested on a challenging dynamic Radio Access Network (RAN) slicing control task. This real-world problem requires balancing acute trade-offs between spectrum efficiency, service quality, and network stability under volatile conditions. Experimental results showed the self-finetuning agent outperformed standard Reinforcement Learning (RL) baselines and existing Large Language Model (LLM)-based agents. Key advantages included superior sample efficiency, greater operational stability, and better optimization across multiple competing metrics.

This work addresses fundamental limitations of using current LLMs for control tasks, such as finite context and a lack of explicit reward mechanisms. By enabling agents to distill experience into their parameters, it paves the way for more robust, autonomous AI systems that can manage and optimize critical infrastructure like 5G and 6G networks without constant human oversight or reward engineering.