Researchers use AI agents to automatically discover inference strategies that cut token usage by roughly 69.5% compared to running 64 parallel chains while matching accuracy.

A team across UMD, UVA, WUSTL, UNC, Google, and Meta built AutoTTS (automated test-time scaling), where a Claude Code agent writes and iteratively improves controller code that decides when an AI should branch reasoning paths, stop early, or cut diverging traces. The agent tests each version against cached reasoning traces without making new model calls; a full discovery run costs $39.90 in API calls and takes 160 minutes.

The discovered Confidence Momentum Controller watches confidence levels across reasoning traces and adapts in real time: it stops early when confidence rises consistently, opens new branches when confidence stagnates or drops, allocates more compute to branches agreeing with consensus, and cuts persistently diverging branches. This strategy was written by an AI agent, not designed by humans.

At the β = 0.5 operating point (balanced setting between speed and accuracy), AutoTTS cuts token usage by roughly 69.5% compared to running 64 parallel chains while matching accuracy on held-out benchmarks across four different Qwen3 model sizes. The controller generalized to AIME25 and HMMT25 benchmarks it never saw during search.

To use this system today requires collecting offline reasoning traces from a target model, API access to Claude Code for the discovery loop, and engineering setup; no pretrained weights are available for direct deployment. ML researchers and inference engineers can evaluate the discovered CMC controller on their own replay data without running discovery.