SkyDiscover: An Open Framework for LLM-Driven Algorithm Discovery

Shubham Agarwal — Tue, 03 Mar 2026 19:20:51 +0000

We are open-sourcing SkyDiscover, a modular framework for AI-driven algorithm discovery.

Most prior systems (e.g., AlphaEvolve) are closed-source. Existing open implementations are monolithic and hard to extend. SkyDiscover decomposes the discovery loop into four swappable components:

Context Builder
Solution Generator
Evaluator
Selector

On top of this framework, we implemented:

AdaEvolve — adaptive search
EvoX — self-modifying search

Results (200+ Benchmarks)

Across math, systems, programming, and multimodal tasks:

+34% median on 172 Frontier-CS problems (vs prior open methods)
Matched/exceeded AlphaEvolve on several math + systems tasks
41% lower cross-cloud transfer cost
29% lower KV-cache pressure

SkyDiscover provides a clean interface for building, comparing, and extending discovery algorithms.

🌐 Blog: https://skydiscover-ai.github.io/blog.html
💻 GitHub: https://github.com/skydiscover-ai
📄 AdaEvolve: https://arxiv.org/abs/2602.20133
📄 EvoX: https://arxiv.org/abs/2602.23413
🐦 Twitter: https://x.com/shulynnliu/status/2028892335875276919?s=20

DEV Community: Shubham Agarwal

SkyDiscover: An Open Framework for LLM-Driven Algorithm Discovery