Turn-budget allocation in multi-turn agentic reasoning

Determine the optimal allocation of interaction turn budgets between internal reasoning tokens and external tool calls for large language model agents performing multi-turn agentic reasoning.

Background

Agentic reasoning introduces a budget across multiple rounds of interaction, where LLM agents must balance internal inference and external tool invocations. Poor allocation can lead to overthinking or premature tool reliance.

The authors identify unresolved questions around this allocation, motivating empirical analyses of deliberative versus reactive tool-use strategies and their effects on accuracy and efficiency.