Researchers Let Claude Code Discover AI Scaling Algorithms that Humans Probably Wouldn't Have Designed

Why It Matters

AutoTTS shows that AI can design its own efficiency strategies, slashing compute costs and accelerating the deployment of more capable language models. This shifts the research bottleneck from manual rule‑crafting to environment design, opening faster innovation cycles for the industry.

Key Takeaways

• AutoTTS discovered a control algorithm that cuts token use by ~70%
• The agent‑written strategy outperforms self‑consistency across four model sizes
• Search cost was only $40 and runtime 160 minutes
• Algorithm adapts width and depth based on model confidence shifts
• Researchers shift from hand‑coding rules to designing search environments

Pulse Analysis

Test‑time scaling (TTS) has become a cornerstone for squeezing extra performance out of large language models without expanding model size. Traditional approaches rely on human‑engineered heuristics that decide when to spawn additional answer paths, prune them, or halt computation. While effective, these hand‑crafted rules often represent narrow slices of a much larger control space, limiting the ability to balance accuracy against compute in a nuanced way.

AutoTTS flips the paradigm by providing an offline replay environment where Claude Code iteratively writes and refines a single high‑level controller. The agent evaluates thousands of candidate policies against pre‑generated solution paths, using feedback from scaling curves to guide its search. The resulting algorithm dynamically adjusts width and depth based on confidence trends, achieving up to a 70% reduction in token consumption while preserving or improving accuracy on math benchmarks like AIME and HMMT, as well as the GPQA‑Diamond reasoning test. The entire discovery process cost roughly $40 and ran for just 160 minutes, demonstrating a cost‑effective pathway to superior TTS strategies.

The broader implication is a shift in research focus: instead of labor‑intensive rule design, engineers now construct expressive search spaces and let powerful coding agents explore them. This could accelerate the development of adaptive inference techniques across domains, from finance to healthcare, where compute budgets are tight but performance demands are high. Future work may expand AutoTTS beyond width‑depth trade‑offs to more complex tree‑search structures, further democratizing AI efficiency gains for enterprises.