Automation Instead of Manual Design
In the era of large language models (LLMs), "test-time scaling" (TTS)—or allocating more computational resources during inference—has proven highly effective. However, most current TTS strategies are still designed manually by researchers, relying heavily on intuition and personal experience.
Recently, a research team introduced AutoTTS, an environment-driven framework that completely changes the game. Instead of designing individual inference rules manually, AutoTTS enables the automated discovery of optimal strategies for specific tasks.
Extraordinary Efficiency at Minimal Cost
AutoTTS works by formulating the strategy design problem as controller synthesis. These controllers dynamically decide when to branch, continue, probe, or halt inference based on feedback signals from the model.
Experimental results on mathematical reasoning datasets show an impressive figure: AutoTTS can reduce token consumption by up to 69.5% compared to today's strongest hand-designed methods, while maintaining or even improving accuracy.
Most notably, the cost to "discover" these strategies is remarkably low. According to the research team, the entire search process costs only about $39.9 and takes 160 minutes to complete.
The Future of Efficient AI Inference
The introduction of AutoTTS not only helps save operational costs for enterprises deploying AI but also opens up new directions for optimizing computational reasoning. Furthermore, the strategies discovered by AutoTTS generalize well across different models and scales.
The project's complete codebase and data have been open-sourced on GitHub, promising to drive the AI community toward further improvements in the efficiency of automated systems.