Reasoning systems increasingly improve their answers by incorporating additional context into the model's input, such as when agents read each other's responses in multi-agent debate or when a model revisits its own output during self-refinement. A new paper on arXiv introduces a token-budget theory to analyze this practice.

The theory frames the decision as a trade-off between thinking more—generating additional reasoning tokens—and reading more—consuming tokens from shared context. The authors propose a framework to determine the optimal allocation of a fixed token budget across these two activities.

The work is presented as a theoretical contribution, covering scenarios like multi-agent debate and self-refinement. The abstract does not provide specific numerical results, but the framework is intended to guide the design of reasoning systems that must balance computational cost and answer quality.