Optima-CityU/LLM4AD_Next

A next-generation automatic algorithm design platform, making automated algorithm design more accessible and easier to use

📦 What is LLM4AD Next?

LLM4AD Next (the successor to the original LLM4AD project) is a Python‑based platform that lets you describe a problem in natural language and instantly get a complete, runnable evolutionary‑algorithm search pipeline. It combines:

  • a conversational AI “consultant” (llm4ad chat) that interviews you and writes the scaffolding code, configuration files and an evaluator;
  • a collection of search/orchestrator methods (IslandGA, Diverse Island GA, EoH, MEoH, ReEvo, MCTS‑AHD, etc.) that evolve algorithm implementations;
  • Algorithm‑design skills that encode how to build specific heuristics (e.g., EoH, MOEA/D) which the AI can invoke automatically.

The end‑to‑end flow is: problem description → AI‑generated project → evolutionary optimisation → best algorithm code – all triggered from a single command.

🚀 Key Features (as described in the README)

Feature What it does
LLM‑powered interactive configuration (llm4ad chat) Conversational terminal that asks you about the task and creates a ready‑to‑run project (evaluator, algorithm skeleton, config, debugger).
Automatic heuristic design (AHD) search methods Built‑in orchestrators such as IslandGA, Diverse Island GA, EoH, MEoH, ReEvo, MCTS‑AHD, DyCA, etc. (some already have Python implementations, others are marked pending).
Algorithm‑design skills Modular skill definitions (e.g., EoH, FunSearch, MOEA/D) that let the coding agent autonomously design algorithms for a given problem.
Evolve‑Block Advisor & Recommender Point the tool at any code repository; it scans, scores, and suggests which blocks of code are most promising to evolve.
AutoResearch workflow An experimental mode (llm4ad experiment) that runs the AutoResearchClaw pipeline, evolving algorithms and only keeping improvements.
Long‑term memory (MindMemOS) Global/project/task‑scoped memory with optional chat/embedding model bindings.
Web demo & Docker images A hosted online demo (https://llm4ad-next.cn/) that requires no local install, plus versioned Docker images for reproducible local deployment.
Extensive examples Benchmarks such as the AlphaEvolve Mathematics suite, TSP, lunar‑lander RL, sorting benchmarks, etc., with ready‑to‑run configs and result artifacts.

🛠️ Installation & Quick Start

Step Command / Action
Clone git clone https://github.com/Optima-CityU/LLM4AD_Next.git && cd LLM4AD_Next
Dependencies Install uv (recommended) then run uv sync (or pip install -r requirements.txt).
Configure LLM provider Set environment variables or create ~/.llm4ad/settings.yaml with provider info (OpenAI, Anthropic, etc.).
Run interactive builder llm4ad chat – the AI will ask you about the problem and generate a full project.
Run a prepared example llm4ad run examples/applications/tsp_benchmark_python/config.yaml
Online demo (no install) Open https://llm4ad-next.cn/ and start designing algorithms in the browser.

The package is published on PyPI (llm4ad‑next) and requires Python 3.12+.

📂 Project Layout (high‑level)

  • src/llm4ad/ – core library: agents, advisors, builders, planners, coders, evaluators, orchestrators, memory, infra.
  • skills/ – reusable “skill” definitions for algorithm design, paper‑to‑task conversion, research‑stage writing, etc.
  • examples/ – ready‑to‑run benchmark tasks (math optimisation, TSP, RL, sorting) and auto‑application demos.
  • docker/, frontend/, backend/ – Docker compose, FastAPI backend, React/Vite UI for the web demo.
  • docs/ – MkDocs site with full user guide, installation, configuration, and contribution docs.

🎯 Typical Use‑Case Workflow

  1. Describe your problem (e.g., “find a high‑quality traveling‑salesperson route for 200 cities”).
  2. Run llm4ad chat; the AI asks clarifying questions and creates a project folder containing:
    • an evaluator that can score candidate solutions,
    • a skeleton algorithm (e.g., a GA template),
    • a configuration file (config.yaml) selecting a search method.
  3. Optionally tweak the config or add custom evaluation code.
  4. Execute llm4ad run <config.yaml> – the selected evolutionary search runs, repeatedly generating and evaluating code variants.
  5. When the evolution stops, the best‑performing algorithm code is saved and can be inspected or deployed.

📚 Documentation & Help

  • Online docs: docs/en/index.md (covers installation, quick‑start, configuration, writing evaluators, etc.).
  • Video tutorial: https://youtu.be/x47kEosu0jk
  • Community: GitHub Discussions, Issues, and a WeChat QR‑code community group.
  • Contributing: Follow the guide in docs/en/contributing/guidelines.md; run tests with pytest and format with black/ruff.

📄 License

BSD‑3‑Clause (see LICENSE).

TL;DR – LLM4AD Next is a genuine, actively maintained open‑source tool that fuses conversational LLM assistance with evolutionary algorithm search, letting users go from a plain English problem statement to a working, optimized algorithm in a single command.

Related

  • Project
    Trampoline-AI/predict-rlm

    predict-rlm is a Python library that gives large language models a sandboxed, stateful runtime. Models write and execute Python code, keep mutable state, call focused sub‑models, and use tools/skills (e.g., PDF handling, spreadsheets). The framework records a full trace for observability and can feed those traces to optimisation tools like GEPA. It’s aimed at complex, multi‑step workflows such as document analysis, audit, or spreadsheet automation, and supports optional Docker sandboxing and rich callbacks.

  • Project
    hiyouga/LlamaFactory

    LlamaFactory is an open‑source library that provides a zero‑code CLI and Gradio UI for fine‑tuning over 100 large language models (LLaMA, Qwen, Mistral, Mixtral‑MoE, etc.) using a wide range of techniques (LoRA, QLoRA, FlashAttention‑2, GaLore, etc.) and supports NVIDIA, AMD, and Ascend hardware. It includes built‑in experiment tracking, OpenAI‑style serving, and Docker images, and is used by Amazon, NVIDIA, and Alibaba.

  • Project
    OEvortex/llm4free

    A Python toolkit that provides a unified, OpenAI-compatible interface for 40+ free and paid AI models, including text, image, and voice generation.

  • Project
    usail-hkust/LLM-MM-Agent

    MM‑Agent is an open‑source system that lets LLMs act as a full‑stack mathematical‑modeling assistant. It parses unstructured problem statements, selects a formulation from a curated library (HMML), auto‑generates and runs Python code, and produces a LaTeX report—all via a web UI or a single CLI command.

  • Project
    sophgo/LLM-TPU

    LLM-TPU is an open‑source toolkit that compiles HuggingFace LLM/VLM checkpoints into SOPHGO TPU‑specific bmodels and provides ready‑to‑run Python/C++ demos, supporting quantization, KV‑cache, dynamic shapes, and multi‑chip parallelism.