AI Agents Discovered a Reasoning Strategy That Cuts LLM Tokens by 70% (firethering.com)

Researchers from a collaboration involving UMD, UVA, WUSTL, UNC, Google, and Meta have developed AutoTTS (automated test-time scaling), a novel approach that enables AI systems to independently discover efficient reasoning strategies, cutting token usage by approximately 70% while maintaining the accuracy of 64 parallel reasoning chains. This breakthrough is significant as it transforms inference from a computationally expensive task into a more efficient process, making AI more accessible and sustainable for large-scale deployment. The innovation revolves around the Confidence Momentum Controller (CMC), an AI-generated policy that dynamically adjusts the reasoning process based on confidence levels across reasoning paths. Unlike traditional methods that rely on brute-force parallel processing, the CMC allows the system to make informed decisions about when to continue exploring or cut branches, optimizing both compute cost and response accuracy. Although the framework and controller code are available for researchers and inference engineers, practical application requires setting up an environment to collect and analyze reasoning traces. This work lays the groundwork for future systems that can autonomously enhance their reasoning strategies, heralding a new frontier in AI/ML research where the design of inference strategies can evolve independently of human input.