← Back to feed

Test-time scaling — spending more compute at inference to improve accuracy — has been one of the most consequential techniques in frontier AI. Until now, every algorithm that governs how that compute gets spent was written by humans.

A team from the University of Maryland, UVA, WUSTL, UNC, Google, and Meta changed that. Their paper, “LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling” (arXiv:2605.08083), built an environment where Claude Code searched for better control algorithms on its own. The agent did it in 160 minutes for $39.90.

The resulting algorithm outperforms human-designed baselines in the majority of test scenarios.

How AutoTTS Works

The approach, called AutoTTS, sidesteps a practical bottleneck: evaluating thousands of inference strategy candidates normally requires thousands of LLM calls. The team solved this by pre-computing 128 solution paths per problem offline and storing them. A candidate controller algorithm then operates on that stored data — no live LLM calls during evaluation, making the search essentially free at the margin.

Claude Code acts as the explorer. Each round, the agent reviews the results from prior controller proposals, identifies where compute was wasted or accuracy dropped, and writes a new controller in code. A constraint limits each proposal to a single high-level parameter that derives all internal thresholds automatically — preventing the search from overfitting to the 25 training problems.

What It Found

The discovered algorithm, which the team named the Confidence Momentum Controller (CMC), tracks the model’s confidence trend using an exponential moving average rather than instantaneous confidence readings. That distinction matters: instantaneous confidence spikes are noisy. The EMA-based signal makes better branch, probe, prune, and stop decisions.

No human on the team designed this. It emerged from Claude Code’s iterative search.

Results across Qwen-3 models (0.6B to 8B) on AIME24/25 and HMMT25:

• Efficiency mode (β = 0): 70% fewer tokens than standard self-consistency (64 parallel samples, majority vote), accuracy maintained
• Accuracy mode (β = 1.0): Outperforms all human-designed baselines in 5 of 8 test scenarios
• Transferability: The discovered policy carries over to DeepSeek-R1-Distill-Llama-8B and to GPQA-Diamond, a non-math benchmark — ruling out overfitting to the training distribution

Why It Matters

Self-consistency is the dominant test-time scaling method in production today. It is also expensive: generating 64 samples and running a majority vote multiplies inference cost by 64x. A controller that matches that performance at 30% of the token budget changes the economics of deployed reasoning models substantially.

The deeper implication is architectural. AutoTTS sits in a line with FunSearch and AlphaEvolve — systems where researchers build environments for AI to search in, rather than writing the solution directly. AlphaEvolve applied this to mathematical and engineering problems. AutoTTS applies it to the inference stack itself. AI is now optimising how AI reasons.

The discovery cost is the starkest number. Frontier labs spending billions on training and inference infrastructure now have evidence that a $40 Claude Code run can produce techniques that improve the efficiency of every model they deploy. The gap between what’s discoverable and what’s been discovered by hand may be larger than anyone assumed.

Key Numbers

• Discovery cost: $39.90
• Discovery time: 160 minutes
• Token reduction: 70% vs self-consistency (efficiency mode)
• Baselines beaten: 5 of 8 test scenarios (accuracy mode)
• Paper: arXiv:2605.08083, code: github.com/zhengkid/AutoTTS