A Flexible Framework for AI-Driven Scientific and Algorithmic Discovery
SkyDiscover is a modular framework for AI-driven scientific and algorithmic discovery, providing a unified interface for implementing, running, and fairly comparing discovery algorithms across 200+ optimization tasks.
SkyDiscover introduces two new adaptive optimization algorithms:
- AdaEvolve, which dynamically adjusts its optimization behavior based on observed progress.
- EvoX, which dynamically evolves the optimization (evolution) strategy itself using LLMs on the fly.
SkyDiscover also supports using OpenEvolve, ShinkaEvolve and GEPA to quickly benchmark these algorithms using their own source code. SkyDiscover also hosts native versions of OpenEvolve and GEPA under openevolve_native and gepa_native algorithms using the modular interface.
SkyDiscover natively supports Harbor-format benchmarks, so you can run external benchmark suites out of the box, including AlgoTune, EvoEval, HumanEvalFix, BigCodeBench, LiveCodeBench, USACO, CRUSTBench, and CodePDE.
🚧 This project is under active development.
Across ~200 optimization benchmarks, AdaEvolve and EvoX achieve the strongest open-source results — matching or exceeding AlphaEvolve and human SOTA, and outperforming OpenEvolve, GEPA, and ShinkaEvolve under identical generation budgets.
- Frontier-CS (172 problems): ~34% median score improvement over OpenEvolve, GEPA, and ShinkaEvolve
- Math + Systems Optimization (12 tasks): Matches or exceeds AlphaEvolve and human-designed SOTA on 6/6 systems and 6/8 math tasks
- Real-world systems impact: 41% lower cross-cloud transfer cost, 14% better GPU load balance for MoE serving, and 29% lower KV-cache pressure via GPU model placement
For a detailed breakdown of results for each algorithm, see the respective papers.
Task breakdown across math, systems, and programming challenges
| Benchmark | Domain | Tasks | Description | |
|---|---|---|---|---|
| 🔢 | math/ | Math | 14 | Circle packing, Erdos problems, geometric optimization |
| 🖥️ | ADRS/ | Systems | 5 | Cloud scheduling, load balancing, MoE expert placement |
| ⚡ | gpu_mode/ | Systems | 4 | GPU kernel optimization |
| 🧩 | frontier-cs-eval/ | Algorithms | 172 | Frontier-CS competitive programming |
| 🧠 | arc_benchmark/ | Reasoning | — | ARC-AGI visual reasoning |
| 💻 | ale_bench/ | Algorithms | 10 | Algorithmic programming contests |
| 🎨 | image_gen/ | Creative | 1 | AI image generation evolution |
| 💬 | prompt_optimization/ | NLP | 1 | HotPotQA prompt evolution |
See Dependency extras for install commands per benchmark.
Prerequisites: Python >= 3.10, uv
# Install
uv sync
export OPENAI_API_KEY="<your-key>"
# Try the circle packing benchmark
uv sync --extra math
uv run skydiscover-run benchmarks/math/circle_packing/initial_program.py \
benchmarks/math/circle_packing/evaluator.py \
--config benchmarks/math/circle_packing/config.yaml \
--search evox \
--iterations 100
uv run skydiscover-run benchmarks/math/circle_packing/initial_program.py \
benchmarks/math/circle_packing/evaluator.py \
--config benchmarks/math/circle_packing/config.yaml \
--search adaevolve \
--iterations 100
# Or run on your own problem
# algo can be "evox", "adaevolve", "openevolve", "gepa", "shinkaevolve"
uv run skydiscover-run initial_program.py evaluator.py \
--search <algo> \
--model gpt-5 \
--iterations 100
# initial_program is optional — omit it to let the LLM start from scratch
uv run skydiscover-run evaluator.py \
--search <algo> \
--model gpt-5 \
--iterations 100
# Run a Harbor benchmark (e.g. AlgoTune) — no seed program needed
pip install harbor
harbor datasets download [email protected] -o /tmp/algotune
uv run skydiscover-run /tmp/algotune/<id>/algotune-set-cover \
--model anthropic/claude-sonnet-4-6 \
--search best_of_n -i 10Or use the Python API:
from skydiscover import run_discovery
result = run_discovery(
initial_program="initial_program.py",
evaluator="evaluator.py",
search=[algo], # algo can be "adaevolve", "evox", "openevolve", "gepa", "shinkaevolve"
model="gpt-5",
iterations=100,
)
print(result.best_score, result.best_solution)SkyDiscover supports three evaluator formats — pick whichever fits your use case:
| Format | When to use | What you point evaluation_file at |
|---|---|---|
| Python function | Simple tasks, no system deps | evaluator.py |
| Containerized | Custom deps, data files, isolation | evaluator/ directory (must contain Dockerfile + evaluate.sh) |
| Harbor task | External benchmark suites (AlgoTune, EvoEval, HumanEvalFix, BigCodeBench, LiveCodeBench, USACO, CRUSTBench, CodePDE, and more) | Task directory (must contain instruction.md + tests/ + environment/Dockerfile) |
SkyDiscover auto-detects the format. See benchmarks/README.md for full setup instructions.
Python evaluator — a file with an evaluate(program_path) function:
def evaluate(program_path):
score = run_and_grade(program_path)
return {
"combined_score": score, # primary optimization target (maximized)
"artifacts": { # optional — stored with the solution for future context
"feedback": "Off by one in the loop boundary",
},
}Containerized evaluator — a directory with a Dockerfile and evaluate.sh that writes JSON to stdout. Runs in Docker, so it can have arbitrary dependencies.
Harbor task — a directory following the Harbor format (instruction.md, environment/Dockerfile, tests/test.sh). Works out of the box with 8+ tested benchmark suites (see benchmarks/README.md for the full list).
- combined_score drives evolution. If omitted, SkyDiscover averages all numeric values in the dict.
- artifacts is optional — entries are injected into the next LLM prompt as context.
For search.type: adaevolve, you can also enable explicit Pareto optimization by configuring search.database.pareto_objectives and returning those objective metrics directly from the evaluator. In that mode, combined_score becomes optional and is only used as a scalar fallback/proxy when configured.
The initial program is optional. When omitted, the LLM generates a solution from scratch. If provided, it marks the region to mutate with EVOLVE-BLOCK markers. Everything outside is left untouched.
# EVOLVE-BLOCK-START
def solve(input_data):
return input_data # baseline — SkyDiscover will improve this
# EVOLVE-BLOCK-ENDIf no markers are present, the entire file is treated as mutatable.
| Algorithm | Flag | Description |
|---|---|---|
| ⭐ AdaEvolve | --search adaevolve |
Multi-island adaptive search with UCB, migration, and paradigm breakthroughs |
| 🧠 EvoX | --search evox |
Self-evolving paradigm that co-adapts solution generation and experience management |
| 📊 Top-K | --search topk |
Selects top-K solutions to refine |
| 🔍 Beam Search | --search beam_search |
Breadth-first expansion of a beam of top solutions |
| 🎲 Best-of-N | --search best_of_n |
Generates N variants per iteration, keeps the best |
| 🧪 GEPA Native | --search gepa_native |
Pareto-efficient search with reflective prompting and LLM-mediated merge |
| 🗺️ OpenEvolve Native | --search openevolve_native |
MAP-Elites + island-based evolutionary search |
Install with uv sync --extra external, then use the corresponding flag:
| Backend | Flag | Source |
|---|---|---|
| OpenEvolve | --search openevolve |
codelion/openevolve |
| GEPA | --search gepa |
gepa-ai/gepa |
| ShinkaEvolve | --search shinkaevolve |
SakanaAI/ShinkaEvolve (manual install) |
ShinkaEvolve manual install
git clone --depth 1 https://github.com/SakanaAI/ShinkaEvolve.git external_repos/ShinkaEvolve
uv pip install -e external_repos/ShinkaEvolvePass a YAML config with -c. See configs/ for full annotated templates.
max_iterations: 100
llm:
models: [{ name: "gemini/gemini-3-pro-preview", weight: 1.0 }]
search:
type: "adaevolve" # or "evox", "topk", "beam_search", "best_of_n"
prompt:
system_message: |
You are an expert at optimizing algorithms.API keys (OPENAI_API_KEY, GEMINI_API_KEY, etc.) are resolved from environment variables automatically.
Add monitor: { enabled: true } to your config. The dashboard URL prints at run start — scatter plot of all programs, code diffs, metrics, and AI summaries. A Human Feedback panel lets you steer evolution in real time.
Replay a completed run:
uv run skydiscover-viewer /path/to/checkpoints/checkpoint_100CLI flags
uv run skydiscover-run [INITIAL_PROGRAM] EVALUATOR [options]
| Flag | Description |
|---|---|
-c, --config FILE |
Config YAML |
-i, --iterations N |
Number of iterations |
-m, --model MODEL |
LLM model (overrides config) |
-s, --search TYPE |
Search algorithm |
-o, --output DIR |
Output directory |
--api-base URL |
Override LLM API endpoint |
--checkpoint DIR |
Resume from checkpoint |
--agentic |
Enable agentic mode (LLM can read your files) |
-l, --log-level LEVEL |
DEBUG, INFO, WARNING, or ERROR |
Python API — discover_solution() (convenience wrapper)
discover_solution() is a convenience wrapper around run_discovery() (shown in Quick Start) for inline string solutions and callable evaluators:
from skydiscover import discover_solution
result = discover_solution(
initial_solution="def solve(x): return x", # optional — omit to start from scratch
evaluator=lambda path: {"combined_score": run_tests(path)},
iterations=50,
search="evox",
)Model providers
Any LiteLLM-compatible model works using provider/model format:
--model gpt-5 # OpenAI (default)
--model gemini/gemini-3-pro-preview # Gemini
--model anthropic/claude-sonnet-4-20250514 # Anthropic
--model ollama/llama3 --api-base http://localhost:11434/v1 # Local (Ollama, vLLM, etc.)Multi-model pools with weighted sampling are supported in config:
llm:
models:
- name: "gpt-5-mini"
weight: 0.7
- name: "gemini/gemini-2.0-flash"
weight: 0.3Benchmark dependency extras
uv sync # Base install
uv sync --extra math # Math benchmarks (SciPy, JAX, PyWavelets, …)
uv sync --extra adrs # ADRS systems benchmarks
uv sync --extra frontier-cs # Frontier-CS benchmark tooling
uv sync --extra external # OpenEvolve / GEPA / ShinkaEvolve backends
uv sync --extra prompt-optimization # HotPotQA prompt optimizationCombine extras as needed: uv sync --extra external --extra math
If a benchmark ships its own requirements.txt, also run: uv pip install -r path/to/requirements.txt
- New benchmark →
benchmarks/README.md - New search algorithm →
skydiscover/search/README.md - New context builder →
skydiscover/context_builder/README.md
SkyDiscover is inspired by AlphaEvolve and incorporates useful code components from open-source efforts such as OpenEvolve. Its interface is compatible with the optimize_anything API.
@misc{skydiscover2026,
title = {SkyDiscover: A Flexible Framework for AI-Driven Scientific and Algorithmic Discovery},
author = {Liu, Shu and Cemri, Mert and Agarwal, Shubham and Krentsel, Alexander and Naren, Ashwin and Mang, Qiuyang and Li, Zhifei and Gupta, Akshat and Maheswaran, Monishwaran and Cheng, Audrey and Pan, Melissa and Boneh, Ethan and Ramchandran, Kannan and Sen, Koushik and Dimakis, Alexandros G. and Zaharia, Matei and Stoica, Ion},
year = {2026},
url = {https://skydiscover-ai.github.io/blog.html}
}For questions or feedback, reach out to us: [email protected] · [email protected] · [email protected]