Large language model agents often perform a reasoning step before each action in a trajectory. The new paper argues this uniform approach is wasteful: not every step requires the same level of deliberation.
The authors propose cross-turn estimation, a technique that looks across the interaction history to predict whether reasoning will change the outcome. When the model is confident, it can act directly; when not, it reasons.
This shifts agents from a fixed reasoning budget to an adaptive one. The work addresses a practical bottleneck in deploying LLM agents, where excessive thinking adds latency and cost.