AI-Driven Scientific Discovery

SkyDiscover is a modular framework for algorithmic discovery, providing a unified interface for implementing and comparing LLM-powered evolutionary algorithms across 200+ optimization tasks.

Get Started Core Concepts API Reference

Quick Example

Quick Start

Get up and running with SkyDiscover in minutes

Install SkyDiscover

Install using uv (Python 3.10+):

uv sync
export OPENAI_API_KEY="your-key"

Or install extras for specific benchmarks:

uv sync --extra math --extra adrs

Create an Evaluator

Write a scoring function that evaluates program quality:

evaluator.py

def evaluate(program_path):
    score = run_and_grade(program_path)
    return {
        "combined_score": score,
        "artifacts": {
            "feedback": "Performance details here"
        }
    }

Run Discovery

Launch the optimization process:

uv run skydiscover-run initial_program.py evaluator.py \
  --search adaevolve \
  --iterations 100

View example output

Iteration 1/100: score=0.6543
Iteration 2/100: score=0.7012
...
Iteration 100/100: score=0.9847

Best solution saved to: outputs/best_solution.py

Key Features

Everything you need for AI-driven algorithmic discovery

AdaEvolve Algorithm

Multi-island adaptive search with UCB selection, migration strategies, and paradigm breakthroughs

EvoX Algorithm

Self-evolving optimization that co-adapts solution generation and experience management

200+ Benchmarks

Math optimization, systems challenges, and programming problems across diverse domains

Multi-Model Support

OpenAI, Gemini, Anthropic, and local models with weighted sampling strategies

Real-Time Monitoring

Live dashboard with scatter plots, code diffs, and human feedback integration

Modular Architecture

Pluggable search algorithms, context builders, and evaluation pipelines

Explore by Topic

Dive deeper into specific areas

Search Algorithms

Explore AdaEvolve, EvoX, TopK, Beam Search, and other optimization strategies

Learn more

Writing Evaluators

Learn how to create effective scoring functions and provide feedback to guide evolution

Learn more

Example Projects

Real-world examples from math optimization to systems design and algorithm discovery

Learn more

Extending SkyDiscover

Build custom search algorithms, benchmarks, and context builders

Learn more

Research & Publications

SkyDiscover is backed by peer-reviewed research

AdaEvolve Paper

Multi-island adaptive evolutionary search with UCB selection and paradigm breakthroughs

Read on arXiv →

EvoX Paper

Self-evolving optimization that dynamically adapts the evolution strategy using LLMs

Read on arXiv →

Benchmark Performance

State-of-the-art results across diverse optimization tasks

~34%

Score Improvement

Median improvement over OpenEvolve, GEPA, and ShinkaEvolve on Frontier-CS

172

Frontier-CS Problems

Competitive programming challenges with strong open-source results

12/14

SOTA Performance

Matches or exceeds human-designed SOTA on math and systems tasks

Community & Resources

Get help and stay connected

GitHub

Star the repo, report issues, and contribute to the project

Blog

Read about the latest updates, research, and use cases

Contact

Reach out to the team for questions or collaboration

Ready to Start Discovering?

Join researchers and engineers using SkyDiscover to push the boundaries of algorithmic optimization

Get Started Now View on GitHub

Get Started

Core Concepts

Guides

Examples

Extending

AI-Driven Scientific Discovery

Quick Start

Key Features

AdaEvolve Algorithm

EvoX Algorithm

200+ Benchmarks

Multi-Model Support

Real-Time Monitoring

Modular Architecture

Explore by Topic

Search Algorithms

Writing Evaluators

Example Projects

Extending SkyDiscover

Research & Publications

AdaEvolve Paper

EvoX Paper

Benchmark Performance

Community & Resources

GitHub

Blog

Contact

Ready to Start Discovering?