// the find
ai4co/reevo
[NeurIPS 2024] ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution
ReEvo uses LLMs as hyper-heuristics — instead of hand-crafting heuristics for combinatorial optimization problems like TSP or VRP, it has the LLM generate, evaluate, and iteratively refine heuristic code through an evolutionary loop. NeurIPS 2024 paper with results showing it can match or beat hand-tuned algorithms on several standard benchmarks. Aimed at researchers in combinatorial optimization who want to automate the algorithm design process.
The reflective evolution loop is the real contribution: separate 'generator' and 'reflector' LLM roles mean the system can critique its own output and explain why a heuristic failed before generating the next candidate, which beats naive mutation. The black-box vs white-box split is well thought out — black-box mode works when you can only observe objective values, white-box when you can inspect intermediate state. Problem extensibility is genuinely straightforward: three files (config, eval, prompts) and you're plugged in. LiteLLM integration means you're not locked to OpenAI and can run it locally with Ollama or similar.
The evaluation loop runs subprocesses and parses stdout to get objective values, which is fragile — a print statement in the wrong place silently corrupts results. LLM API costs are real and undocumented: running 20 heuristic evaluations with GPT-4o on a non-trivial problem could easily cost $5-20 per run, and there's no cost tracking or budget guard. The 289-star count suggests limited adoption outside the paper authors, meaning the problems directory is the full extent of what's been validated in practice; extending to your own domain means you're on your own. No reproducibility tooling — results will drift as model providers update their APIs.