Large language model (LLM)‑based agents excel on complex tasks, typically reasoning before each action along an interaction trajectory. However, reasoning is not required at every turn because earlier reasoning can often continue to support later actions. The main challenge is to decide when existing reasoning remains sufficient and when a new reasoning step is needed, without resorting to costly generation‑based verification.
We observe that the drop in likelihood of subsequent reference actions after removing additional reasoning closely mirrors whether those actions are still recoverable from earlier reasoning. This provides an effective, lightweight signal for cross‑turn action support estimation. Building on this insight, we introduce RACE (Reasoning Adaptation through Cross‑Turn Estimation), a training approach that enables agents to adapt their reasoning behavior. RACE incorporates a LoGiC (Likelihood‑Guided Progressive Reasoning Cover Detection) procedure that progressively identifies reasoning turns whose removal has limited impact on the current and future reference actions. The removal signals are fed into both supervised fine‑tuning and agentic reinforcement learning, allowing the policy to learn when to reason and when to act directly.
Extensive experiments on four representative agent benchmarks demonstrate that RACE substantially reduces reasoning cost while maintaining or even improving task performance.
Review