AutoTTS uses an AI-designed controller to cut reasoning-token use 69.5%

AutoTTS belongs in the research lane because it attacks one of the biggest production AI problems: reasoning cost. VentureBeat reports that researchers from Meta, Google, and several universities built a framework that lets an explorer LLM discover controllers for test-time scaling. Instead of hand-designing when a model should branch, deepen, prune, or stop reasoning, AutoTTS searches over strategies using an offline replay environment of pre-collected reasoning trajectories. On Qwen3 models and a distilled DeepSeek-R1 model, the balanced controller cut token consumption by about 69.5% compared with self-consistency using 64 paths while maintaining average accuracy. On GPQA-Diamond, token cost fell from 510K to 151K tokens. The discovery process reportedly cost $39.90 and took 160 minutes. Watch whether these controllers become standard middleware for expensive reasoning workflows.

Key details: AutoTTS, Meta, Google, May 28, 2026, test-time scaling, 69.5% token reduction, Qwen3, DeepSeek-R1 distilled 8B.

Continue swiping for more AI Brief stories.