OpenLoopX · LLM4AD Next

From problem description to runnable evolutionary algorithm search — in one command.
LLM-driven automated algorithm design with evolutionary optimization

Website · Quickstart · Docs · Wiki

English | 中文

---

⭐ Star us on GitHub to earn 10$ bonus tokens for the Online Website!

---

🔥 News

• 🔬 [2026.09][AutoResearch]: LLM4AD_Next × AutoResearchClaw adds algorithm evolution before code is finalized in stage 13 of the 23-stage research workflow. Candidates are checked with the original evaluator; only better algorithms replace the baseline, while unsuccessful attempts leave the original code intact. Try AutoResearch with the llm4ad experiment mode.
• 🧮 [2026.09][New Dataset]: The AlphaEvolve Mathematics Benchmark Suite adds 11 independently runnable mathematical optimization cases, case-local evaluators, evolved implementations, and reusable experience artifacts.
• 🏝️ [2026.09][New Search Method]: Diverse Island GA is now available, assigning a continuous spectrum of exploitation, correction, and independent-exploration behaviors across any number of islands while coordinating migration and memory use.
• 🎯 [2026.08][New Feature]: Algorithm Design Skills — Modular skill definitions (EoH, FunSearch, ReEvo, MEoH, MOEA/D) that enable coding agents to autonomously design algorithms. See Algorithm Design Skills.
• 🔬 [2026.07][New Feature]: Search methods migrated — EoH, MEoH, ReEvo, and MCTS-AHD are now available as standalone orchestrators. See Search Methods.
• 🧠 [2026.07][New Feature]: MindMemOS-backed long-term memory is now available, with global, project, and task memory scopes plus configurable Chat and Embedding model bindings. See the Memory Guide.
• 🚀 [2026.07][New Release]: LLM4AD_Next Online Trial is now available at https://llm4ad-next.cn/ — try the full problem-to-algorithm workflow directly in your browser with no local setup.
• ✨ [2026.07][New Feature]: Introducing an interactive problem-to-project workflow that turns natural-language problem descriptions into runnable evolutionary algorithm search projects.
• 🐳 [2026.07][New Feature]: Versioned Docker Hub deployment images are now aligned with GitHub Release tags for reproducible local deployment.

🚀 Why LLM4AD_Next?

Traditionally, using Large Language Models for Automated Algorithm Design (LLM4AD) required a tedious, multi-step configuration pipeline. LLM4AD_Next destroys this entry barrier.

With LLM4AD_Next, after creating your directory, all of these painful steps are fully automated through an interactive conversational terminal. Just run:

uv run llm4ad chat

Our built-in AI-powered consultant will interview you, instantly understand your requirements, and automatically generate a ready-to-run pipeline (evaluator, algorithm skeleton, configuration, and debugger) so you can leap straight into producing Useful Algorithms.

🎯 Key Features Overview

• 🧠 LLM-Powered Design & 🧬 Evolutionary Optimization combined to automatically evolve top-performing code.
• 💬 Interactive Configuration (llm4ad chat) — Your conversational AI consultant that generates the entire runnable app framework.
• 🔍 Evolve-Block Advisor & Recommender — Point LLM4AD_Next at any repository, and it will scan, score, and recommend exactly which blocks of code are most promising to evolve to hit your goals.
• 🔬 AutoResearch — Run an AutoResearchClaw research workflow; choose llm4ad experiment mode to evolve algorithms and retain the original code when candidates do not improve their evaluation scores.

Search Methods (Automatic Heuristic Design)

Overview of the Automatic Heuristic Design (AHD) search methods from the original LLM4AD platform. Impl = whether the method has a working orchestrator implementation in code; Skill = whether an algorithm design skill is available for coding agents.

Using the migrated methods

Set evolution.type in your config and run llm4ad run <config.yaml>. See examples/config/config.complete.yaml for full examples.

evolution:
  type: "eoh"  # options include "diverse_island_ga", "island_ga", "eoh", "meoh", "reevo", "mcts_ahd", "dyca"

Algorithm Design Skills

Modular skill definitions that enable coding agents to autonomously design algorithms. Give a coding agent this prompt:

I want you to design a [PROBLEM] solver using the [SKILL] method.

Skill: https://github.com/Optima-CityU/LLM4AD_Next/blob/develop/skills/algo-design/[SKILL]/SKILL.md
Task: /path/to/your/task/

Read the skill, read the task package, run [N] generations, give me the best algorithm.

See use_example for a complete TSP + EoH example.

🏆 Featured Cases

AlphaEvolve Mathematics Benchmark

Quick Start

Instruction Video

Run LLM4AD Next

Option A: Online Demo (No Installation Required)

Use the online demo from Quick Start, or open it directly: Launch Online Demo.

No setup, no API key needed — just open the link and start designing algorithms.

Option B: Local Installation

Requires Python 3.12+ (pinned in .python-version) and uv (recommended) or pip. A plain uv sync sets up everything, including the chatv2 AI build agent, out of the box.

# Clone the repository
git clone https://github.com/Optima-CityU/LLM4AD_Next.git
cd LLM4AD_Next

# Install dependencies
uv sync

# Configure your LLM provider (see Global Settings section below)
# Or set environment variables directly:
export LLM_BASE_URL="https://api.openai.com/v1"
export LLM_API_KEY="your-api-key"
export LLM_MODEL="gpt-4o"

# Option 1: Interactive configuration (recommended for new users)
llm4ad chat

# Option 2: Run with an existing config file
llm4ad run examples/applications/tsp_benchmark_python/config.yaml

For optional dependency groups (infra, providers, eval, dev, docs, all) and uv installation, see the Installation Guide.

Global Settings

Create ~/.llm4ad/settings.yaml to configure shared providers across all projects:

providers:
  - name: default
    type: openai
    api_key: ${OPENAI_API_KEY}
    model: gpt-4o
  - name: anthropic
    type: anthropic
    api_key: ${ANTHROPIC_API_KEY}
    model: claude-sonnet-4-20250514

Task configs then only need the provider name — credentials and model are resolved from global settings automatically.

For CLI commands, the interactive chat workflow, the Evolve-Block Advisor / Recommender, and the Python API, see the Documentation.

Documentation

• Documentation Home
• Quick Start Guide
• Configuration Guide
• Writing Evaluators

Local Development

# Serve documentation with live reload
mkdocs serve

# Build static documentation
mkdocs build

Project Structure

LLM4AD_Next/
├── src/
│   ├── llm4ad/           # Core Python package and CLI
│   │   ├── agent/        # Conversational task-building agent
│   │   ├── advisor/      # Evolve-block advisor and recommender
│   │   ├── builder/      # Runnable task-package builder
│   │   ├── config/       # Configuration schemas and settings
│   │   ├── planner/      # Algorithm planning
│   │   ├── coder/        # Code generation
│   │   ├── evaluator/    # Candidate evaluation
│   │   ├── orchestrator/ # Evolution workflows
│   │   ├── memory/       # Memory integration
│   │   └── infra/        # Providers and shared infrastructure
│   ├── backend/          # FastAPI, workers, migrations, and API tests
│   └── frontend/         # React/Vite web app and UI tests
├── skills/               # Agent skills
│   ├── algo-design/      # Algorithm design methods
│   ├── autodiscovery/    # Paper-to-evolvable-task workflow
│   ├── autorebuttal/     # Reviewer response and AC summary workflow
│   ├── document-knowledge-organizer/ # Markdown knowledge organization
│   ├── llm4ad-task-builder/ # Runnable task-package creation
│   ├── openair-proposal/ # Staged research proposal writing
│   ├── research-stage-publication/ # Stage result publication
│   └── typst-author/    # Typst document authoring
├── docker/               # Compose stacks and runtime support
├── third_party/          # Integrated upstream submodules
│   ├── CloudCLI/         # Cloud development workspace
│   └── MindMemOS/        # Memory service and SDK
├── examples/             # Example tasks and benchmarks
│   ├── applications/    # Runnable tasks and benchmarks (examples below)
│   │   ├── alphaevolve_math_benchmark/ # Mathematical optimization cases
│   │   ├── lunarlander_python/ # Reinforcement learning example
│   │   ├── sorting_benchmark/ # Sorting algorithm benchmark
│   │   └── tsp_benchmark_python/ # Traveling-salesperson benchmark
│   ├── auto_applications/ # Automatic task-building examples
│   │   ├── from_code/   # Build tasks from existing code
│   │   └── from_description/ # Build tasks from a description
│   └── config/          # Sample configuration files
├── tests/                # Core Python test suite
├── docs/                 # English and Chinese documentation
└── scripts/              # Repository utilities

Contributing

Contributions are welcome! Please read our Contributing Guide for details.

# Set up development environment
uv sync --extra all

# Run tests
pytest

# Format code
black src/ tests/
ruff check src/ tests/ --fix

License

This project is licensed under the BSD 3-Clause License - see the LICENSE file for details.

Acknowledgements

The AutoResearch module is based on / adapted from AutoResearchClaw (MIT License). Its original copyright and license notice are retained in THIRD_PARTY_LICENSES.md.

Support

• Documentation
• Discussions
• Issue Tracker

Join the Community

Scan the QR code with WeChat to join the LLM4AD_Next community group.