Researchers at the University of Exeter and British Telecom have developed a language model agent for autonomous network control that learns from written critiques of its decisions. In a custom simulator, the system achieved the highest composite utility among five tested methods after one 300-step training trajectory. The reported reinforcement learning schedule used 1,600 trajectories, suggesting a substantial reduction in costly environment interactions.

The preprint applies the method to radio access network slicing, where a controller must divide spectrum among competing services in real time. “Reward-free” has a specific scope here: language judgments and preference optimization drive policy updates, while scalar metrics and a composite utility score remain part of the evaluation.

One Score for Conflicting Goals

Radio Access Network slicing allocates Physical Resource Blocks, the time-frequency units assigned by a cellular scheduler, across virtual network slices. The experiment includes guaranteed bit rate traffic, which has explicit service commitments, and elastic non-GBR traffic.

A controller must balance three objectives:

  • Resource utilization: keep available spectrum productive.
  • Service quality: limit violations across traffic classes.
  • Stability: avoid frequent allocation changes and their operational overhead.

A conventional reinforcement learning reward compresses those objectives into one weighted scalar. Weight selection can produce policies that optimize one metric at the expense of another, and the useful balance may shift as traffic changes. Prompt-based agents can reason over written objectives, but long histories consume context and models may underuse information buried in earlier traces. Updating the model’s weights gives the agent a persistent form of adaptation.

A Critic That Writes

The framework separates online action selection from retrospective credit assignment. A Qwen3-4B Actor chooses resource allocations, while a DeepSeek-R1 Reflector reviews the completed trajectory. The loop proceeds through five stages: