AlphaResearch: Accelerating New Algorithm Discovery with Language Models
Zhaojian Yu, Kaiyue Feng, Yilun Zhao, Shilin He, Xiao-Ping Zhang, Arman Cohan
2025-11-14
Summary
This paper introduces AlphaResearch, a system built using a large language model that attempts to independently discover new algorithms for solving problems, essentially acting as an automated researcher.
What's the problem?
While large language models are good at solving problems where the answer is easy to check, they struggle with truly *new* discoveries – finding algorithms for problems where we don't already know the solution. It's hard for them to come up with genuinely innovative approaches and then prove those approaches work.
What's the solution?
The researchers created a system called AlphaResearch that works in a loop. First, it proposes new ideas for algorithms. Then, it tests those ideas in a special environment that both checks if they work correctly and simulates how well they'd be received by other researchers. Finally, it uses the results of those tests to refine its ideas and try again. They also created a set of challenging problems, called AlphaResearchComp, to specifically test this kind of algorithm discovery.
Why it matters?
AlphaResearch managed to discover an algorithm for one problem, 'packing circles,' that actually outperformed existing solutions developed by human researchers and other AI systems. This shows that large language models *can* potentially accelerate the process of discovering new algorithms, which could be important for advancements in many fields. Even where it didn't succeed, analyzing the failures provides valuable information for improving these kinds of systems in the future.
Abstract
Large language models have made significant progress in complex but easy-to-verify problems, yet they still struggle with discovering the unknown. In this paper, we present AlphaResearch, an autonomous research agent designed to discover new algorithms on open-ended problems. To synergize the feasibility and innovation of the discovery process, we construct a novel dual research environment by combining the execution-based verify and simulated real-world peer review environment. AlphaResearch discovers new algorithm by iteratively running the following steps: (1) propose new ideas (2) verify the ideas in the dual research environment (3) optimize the research proposals for better performance. To promote a transparent evaluation process, we construct AlphaResearchComp, a new evaluation benchmark that includes an eight open-ended algorithmic problems competition, with each problem carefully curated and verified through executable pipelines, objective metrics, and reproducibility checks. AlphaResearch gets a 2/8 win rate in head-to-head comparison with human researchers, demonstrate the possibility of accelerating algorithm discovery with LLMs. Notably, the algorithm discovered by AlphaResearch on the ``packing circles'' problem achieves the best-of-known performance, surpassing the results of human researchers and strong baselines from recent work (e.g., AlphaEvolve). Additionally, we conduct a comprehensive analysis of the remaining challenges of the 6/8 failure cases, providing valuable insights for future research.