Claude Code agent discovers efficient test-time scaling algorithms
anthropic's claude code agent designed test-time scaling algorithms that outperform human methods, reducing compute by 70% on math benchmarks.
a research team introduced autotts, a framework where a coding agent searches for optimal test-time scaling control algorithms within a simulated environment. instead of human-designed rules for when to spawn or terminate solution paths, claude code evolved a logic that dynamically allocates compute based on confidence shifts across rounds. this method achieves better accuracy per compute unit than standard approaches like self-consistency, cutting token usage by roughly 70%. the $40 discovery process produced an algorithm that generalizes to different models and benchmarks, relying on detailed logging to avoid shortcuts that degrade performance.