A Framework for AI-Driven System Synthesis and Scientific Discovery
SkyDiscover is an open source project from UC Berkeley providing state-of-the-art tooling for scientific discovery and end-to-end system synthesis.
SkyDiscover provides tools for two broad classes of tasks:
- Optimization: given an evaluator and problem description, apply evolutionary techniques to incrementally discover better solutions.
- Synthesis: starting from a formal specification or text description of a system, full end-to-end synthesis of compliant and highly-performant systems.
See the getting started page for quickstart guides for each.
🚧 This project is under active development.
SkyDiscover Optimize 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.
We ship with two SOTA adaptive optimization algorithms built by the SkyDiscover team, AdaEvolve and EvoX. SkyDiscover Optimize also supports OpenEvolve, ShinkaEvolve and GEPA to quickly benchmark these algorithms using their own source code.
SkyDiscover Optimize 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.
Get started with our quick start guide here.
SkyDiscover Synthesize is a set of tooling that enables full, end-to-end synthesis of compliant and highly-performant systems, such as LLM inference systems, model routing, or key-value stores.
Our synthesis tooling is built on two published techniques introduced by the SkyDiscover team:
- Inductive-Deductive Synthesis: when a formal specification in Rocq for the target system is provided, Synthesize jointly generates implementation and corresponding proofs in stages, optimizing for performance in the process.
- Test-Driven Synthesis: when the system is described by natural language and tests, Synthesize iterates on implementations with a collection of planning, coding, critiquing, and auditing agents.
SkyDiscover Synthesize is provided as a Claude Code and Codex-compatible skill, and has been used to produce a wide range of systems.
See the Quick Start guide here.
SkyDiscover Optimize has been used across industry including Google, Uber, and more; the algorithms released by the SkyDiscover team, AdaEvolve and EvoX, achieve the strongest open-source results across ~200 optimization benchmarks: 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 (14 tasks evaluated): 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
📊 Complete results of AdaEvolve and EvoX (100 iterations)
AdaEvolve and EvoX are complementary: AdaEvolve adapts search parameters for fast early gains; EvoX evolves the search strategy itself for stronger long-horizon gains. Both are built on SkyDiscover.
📈 Scaling behavior of AdaEvolve and EvoX
The scaling behavior of AdaEvolve and EvoX shows a complementary crossover. AdaEvolve's per-iteration parameter adaptation yields fast early gains in low-budget runs (T≤50), while EvoX's demand-driven strategy evolution unlocks step-change improvements in longer runs (T≥50).
Best-so-far score vs. iteration for Signal Processing, Heilbronn Convex, Prism, and Cloudcast (500 iterations, GPT-5).
🔗 Evolving AdaEvolve's policy with EvoX (coming soon)
The two methods are composable: EvoX can evolve using AdaEvolve as its starting strategy, achieving the best results on 3 out of 4 benchmarks (100 iterations, GPT-5). This combined mode will be available in SkyDiscover soon.
| Benchmark | AdaEvolve | EvoX (Random Init) | EvoX (AdaEvolve Init) |
|---|---|---|---|
| Signal Proc. (↑) | 0.718 | 0.721 | 0.760 |
| Heilbronn Cvx. (↑) | 0.0290 | 0.0270 | 0.0291 |
| Cloudcast (↓) | 640.5 | 637.1 | 623.4 |
| Prism (↑) | 26.37 | 30.52 | 26.27 |
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 |
| 🔧 | kernelbench/ | Systems | 250+ | KernelBench GPU kernel speedup 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 algotune@1.0 -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.
See Benchmark Performance for a detailed comparison of AdaEvolve and EvoX against other algorithms.
| 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.
@inproceedings{liu2026skydiscover,
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 Zaharia, Matei and Dimakis, Alexandros G. and Stoica, Ion},
title = {SkyDiscover: A Flexible, Adaptive Framework for AI-Driven Scientific and Algorithmic Discovery},
booktitle = {Proceedings of the ACM Conference on AI and Agentic Systems},
series = {CAIS '26},
year = {2026},
pages = {1223--1227},
publisher = {Association for Computing Machinery},
doi = {10.1145/3786335.3813221},
url = {https://doi.org/10.1145/3786335.3813221}
}For questions or feedback, reach out to us: lshu@berkeley.edu · akrentsel@berkeley.edu · mert_cemri@berkeley.edu · shubham3@berkeley.edu