Researchers let AI Agents Optimize LLM Reasoning and Cut Tokens by 70%

Researchers from several universities and companies built a system called AutoTTS, where an AI agent automatically writes and refines code that controls how AI models reason — cutting token usage by 70% while matching the accuracy of running 64 parallel reasoning chains, making AI inference significantly cheaper and more efficient.