Researchers let Claude Code discover AI scaling algorithms that humans probably wouldn't have designed

Researchers used Claude Code to auto-discover a new test-time scaling algorithm in an offline simulation called AutoTTS.

Instead of human-written branching and stopping rules, the agent learned to allocate compute based on confidence shifts during reasoning.

The method used about 70% fewer tokens than standard self-consistency while matching or improving accuracy on math and other benchmarks.

The algorithm also transferred to another model and task, suggesting AI agents can design better inference-control strategies than humans.