A Communication-Theoretic Framework for LLM Agents: Cost-Aware Adaptive Reliability

Researchers present a communication-theoretic framework that unifies LLM reliability techniques (retry, majority voting, self-consistency) under classical information theory, introducing a cost-aware router that achieves 56% lower costs than fixed approaches while maintaining quality. The work demonstrates that no single reliability technique dominates across all tasks, supporting dynamic per-task allocation strategies.

This research bridges two previously disconnected domains: LLM engineering practices and Shannon's communication theory. By modeling sampled LLMs as discrete stochastic channels, the authors reframe reliability techniques as special cases of six classical coding operators. This theoretical consolidation matters because LLM applications increasingly demand reliability guarantees while facing cost pressures in production deployments.

The work emerges from a critical observation: the reliability techniques developed for LLMs evolved independently without unifying principles, leading to suboptimal deployment choices. By grounding these methods in communication theory, the researchers establish formal conditions—such as noise-variance thresholds for averaging strategies—that guide when to apply which technique. Their cost-aware semantic-nearest-neighbor router implements a Lagrangian knob allowing practitioners to traverse the quality-cost frontier dynamically without retraining.

For developers and practitioners, these findings have immediate relevance. The empirical results across 69 hard tasks demonstrate that fixed model-technique-budget combinations consistently underperform adaptive routing. On MMLU, GSM8K, and HumanEval subsets, the router achieves 56% cost reduction at matched quality levels and 7% quality improvements at matched costs—significant margins in resource-constrained environments. The contractivity criterion for generator-critic refinement also provides theoretical insight into why larger models behave differently, explaining observed transitions between 3B and 14B parameter scales.

Looking ahead, the consolidation of these techniques into a tunable layer with communication-theoretic foundations could accelerate development of more efficient inference systems. As LLM deployment scales, the economic incentives for cost-optimized reliability grow stronger, making this framework increasingly valuable for infrastructure providers and edge-deployed applications.

• →LLM reliability techniques can be unified under Shannon's communication theory as special cases of six classical coding operators.
• →A cost-aware router achieves 56% lower normalized costs than fixed techniques while maintaining equivalent quality on hard reasoning tasks.
• →No single reliability method dominates across all task types, requiring dynamic per-task allocation for optimal resource utilization.
• →Formal noise-variance thresholds and contractivity criteria provide theoretical guidance for technique selection across model scales.
• →The framework enables traversing the quality-cost frontier without retraining through a single Lagrangian parameter.