TrACE: Training-Free Adaptive Compute for LLM Agents via Inter-Rollout Action Agreement

arXiv·medium signal

Introduces TrACE, a training-free controller that allocates LLM inference budget adaptively across agent decision steps by measuring how much candidate actions agree across parallel rollouts. Easy decisions (high agreement) get fewer samples; hard decisions get more. This avoids the waste of applying uniform compute budgets to every agent step and works as a drop-in wrapper around existing agents without retraining.