SkyDiscover: An Open Framework for LLM-Driven Algorithm Discovery

Source: Dev.to

Framework Overview

SkyDiscover decomposes the discovery loop into four interchangeable components:

• Context Builder – constructs the problem context for the LLM.
• Solution Generator – produces candidate algorithms or solutions.
• Evaluator – assesses the quality or performance of each candidate.
• Selector – chooses the most promising candidates for the next iteration.

Implementations

Built on top of the framework are two discovery algorithms:

• AdaEvolve – an adaptive search strategy.
• EvoX – a self‑modifying search strategy.

Results (200+ Benchmarks)

Across mathematics, systems, programming, and multimodal tasks:

• +34 % median improvement on 172 Frontier‑CS problems compared with prior open methods.
• Matched or exceeded AlphaEvolve on several math and systems tasks.
• 41 % lower cross‑cloud transfer cost.
• 29 % lower KV‑cache pressure.

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

Resources

• Blog:
• GitHub:
• AdaEvolve paper:
• EvoX paper:
• Twitter announcement: