AutoTTS: Claude Code Autonomously Discovers a 70% Token Efficiency Algorithm

A 160-minute, $40 autonomous Claude Code session produced AutoTTS — a test-time scaling method that cuts token usage by 70% compared to standard self-consistency approaches with no accuracy loss on math reasoning benchmarks.

The Agentic Discovery

Research teams from UMD, UVA, WUSTL, UNC, Google, and Meta made a specific design choice: they gave Claude Code a search environment and let it find novel test-time scaling algorithms without specifying what those algorithms should look like. The researchers “set up the search environment instead of inventing the rules themselves” — letting the model discover methods that human designers would be unlikely to construct through conventional research.

AutoTTS (Automated Test-Time Scaling) is the result. It identifies and applies optimal scaling strategies at inference time without requiring changes to model weights or training pipelines. The discovery run cost $40 and ran for 160 minutes — well within the range of a CI job, not a multi-week research effort.

What AutoTTS Actually Does

The primary benchmarks were AIME and HMMT math competitions — tasks where test-time compute is already known to move accuracy. AutoTTS achieved 70% token reduction versus standard self-consistency sampling while maintaining accuracy. At typical API prices, that compression directly translates to cost reduction at scale.

The more significant result is transferability: AutoTTS generalizes across different models and non-math domains. That is the difference between a benchmark curiosity and a deployment-relevant finding. Teams already running agentic pipelines — multi-step reasoning chains, RAG-augmented workflows, any task where a single user query triggers multiple inference calls — see token cost as the primary scaling constraint. The Alibaba Qwen 3.7-Max run that demonstrated a 10x speedup on chip optimization operated for 35 hours; at that duration, a 70% token reduction changes the economics of whether long-horizon autonomous runs are viable.

AutoTTS is research output — it requires engineering integration before it slots into production stacks. But the implication is clear: the optimization frontier for test-time scaling has not been exhausted, and the algorithms that will matter most may not come from human researchers following established paradigms.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.