Skip to main content
Back

Hill Sampling Beats AlphaEvolve on Math With 8 GPUs and Frozen Weights

A new paper introduces Hill Sampling, a minimalist test-time search that beats AlphaEvolve and evolutionary weight training on hard math problems.

Paper, Full Access
  • New paper introduces Hill Sampling: repeatedly edit the best program found so far with a frozen LLM.
  • Sets state of the art on circle packing among published methods, beats AlphaEvolve on Erdos' minimum-overlap.
  • Results achieved in hours on 8 NVIDIA H100 GPUs across three open-weight models.
  • Largest study of evolution strategies on LLM weights at test time to date.
  • Training weights at test time was worse than a zero learning rate baseline.
  • Simple rule: exhaust repeated edit sampling before adding archives, diversity, or weight updates.

Researchers Jacob Beck, Philip V. Ogren, and Ari Kobren report that a frozen large language model can outperform more elaborate search systems by repeatedly editing its best verified program. Their method, Hill Sampling, produced a new published result for circle packing, improved on an AlphaEvolve reference for an open Erdős problem, and ranked among the leading methods for a finite-set problem.

The experiments used eight NVIDIA H100 GPUs for a few hours and tested three open-weight models. For developers building verifiable discovery systems, the results support a lean baseline: keep one incumbent program, generate edits, score them automatically, and preserve only improvements.

Why LLM search became complicated

Verifiable problems give every candidate a dependable automatic score. Examples include optimization programs, mathematical constructions, and code whose output can be checked without human review. This feedback allows an inference system to spend additional compute generating, testing, and refining answers.

Systems such as DeepMind’s AlphaEvolve maintain populations of programs, preserve diverse candidates, and coordinate several evolutionary stages. Other approaches perturb an LLM’s weights during inference, score the resulting models, and update the parameters toward better-performing perturbations.

Beck, Ogren, and Kobren compare those approaches across circle packing in a unit square, sums and differences of finite sets, and Erdős’s minimum-overlap problem. Each benchmark converts a proposed construction into a numeric fitness score, allowing the search loop to accept or reject candidates automatically.

One incumbent, one acceptance rule

Each Hill Sampling iteration stores the highest-scoring program found so far and uses it as the context for the next edit. The model’s parameters remain fixed throughout the run.