Claude Code Discovers AI Scaling Algorithms Humans Would Never Have Designed
Researchers have taken a bold step back and let an AI do what they usually do themselves, design AI scaling algorithms. A team from the University of Maryland, UVA, WUSTL, UNC, Google, and Meta built a system called AutoTTS that hands the algorithm-writing job to a coding agent, and the result is turning heads.
Test-time scaling is a technique that makes large language models perform better by letting them spend more compute on a response, whether by running several solution paths in parallel or extending chains of thought. Until now, human-written rules have always governed when to start a new path, push deeper into a promising one, or abandon a dead end. AutoTTS changes that entirely.
Instead of manually crafting the rules, the researchers designed an offline environment where a coding agent, specifically Claude Code, hunts for better AI scaling algorithms on its own. The environment pre-stores solution paths so thousands of algorithm variants can be tested without running the full language model each time. Claude Code reviews what previous rounds produced, spots where compute was wasted, and writes a new control algorithm directly in code.
The AI scaling algorithms it discovered are not just theoretically interesting, they work. On math benchmarks like AIME and HMMT, the agent-written algorithm cuts token usage by roughly 70 percent compared to standard self-consistency, which typically generates 64 parallel answers and picks the majority winner, while holding accuracy steady. The algorithm also transferred successfully to DeepSeek-R1-Distill-Llama-8B and a non-math benchmark, GPQA-Diamond. The entire discovery run cost about $40 and took just 160 minutes.
What makes the result particularly striking is how the discovered algorithm actually thinks. It monitors how the model’s confidence shifts across multiple rounds rather than stopping as soon as a majority answer emerges. When confidence rises quickly, it skips opening new solution paths. When it stalls, it opens more. Paths that align with the current majority get extra compute, while diverging ones are only cut after consistently pointing in the wrong direction over several rounds. The authors describe this kind of coordination as something that would have been nearly impossible for a human to design deliberately.
AutoTTS joins a growing line of work, including Google DeepMind’s FunSearch and AlphaEvolve, that uses language models as program searchers rather than just answer generators. The bigger shift AutoTTS represents is where human effort now goes: not into writing the strategy, but into designing the search space where the strategy can emerge.
