A curated list of frameworks that combine Large Language Models (LLMs) with evolutionary algorithms or search-based optimization techniques (e.g. genetic algorithms, reinforcement learning, MCTS). Categorized by application domain.
-
FunSearch (DeepMind, 2024) – Pairs an LLM with an evaluator to evolve code for open math and algorithm problems. Achieved new records in combinatorics and discovered high-performance heuristics.
-
AlphaEvolve (DeepMind, 2025) – Gemini-based LLM agent that evolves entire programs, outperforming previous algorithm benchmarks, including a breakthrough in matrix multiplication.
-
OpenEvolve – Open-source version of AlphaEvolve. Modular LLM+EA framework with support for multi-language generation, ensemble agents, and custom evaluation loops.
-
Evolution of Heuristics (EoH) – Evolves natural-language 'thoughts' and corresponding heuristics with an LLM loop. Beats previous methods on combinatorial optimization problems.
-
ReEvo – Reflective evolution loop where the LLM both mutates and critiques heuristic strategies, leading to general-purpose solvers.
-
LLaMEA – Uses GPT-4 to iteratively evolve optimization algorithm code; beats classic optimizers on benchmark suites.
-
MPaGE – Evolves diverse multi-objective heuristics guided by a Pareto grid. Achieves strong performance across several domains.
-
SOAR – Self-improving LLM that fine-tunes itself from successful generations. Dominates ARC symbolic reasoning benchmark.
-
LAEA – Zero-shot LLMs act as surrogate models to rank solutions, replacing learned predictors in classic EA.
- LLM4EC – Community-sourced list of all papers at the intersection of LLMs and Evolutionary Computation.
-
LLM-MCTS – GPT-4 guides Monte Carlo Tree Search for high-level planning in robotics and games.
-
MC-DML – Dynamic memory + GPT integrated into MCTS for interactive fiction. Significantly boosts single-shot win rates.
-
LERO – Evolves reward shaping functions and partial observation strategies in multi-agent RL with GPT assistance.
-
Voyager – GPT-4 powered autonomous Minecraft agent that iteratively improves its own skills and codebase.
-
Tree-of-Thoughts – Search-based reasoning framework that treats LLM generations as a search tree of "thoughts".
-
ASI-Arch – Gemini-based autonomous AI researcher that discovered 100+ transformer architectures outperforming human designs.
-
LLMatic – GPT-4 powered neural architecture mutation guided by quality-diversity search. Efficient CIFAR and NAS-Bench exploration.
-
DesignGPT – Framework where GPT-4 recommends network architecture improvements iteratively. Early-stage AutoML system.
-
MOLLEO – GPT-4 proposes chemical modifications as evolutionary mutations. Strong results on drug design and property optimization.
-
MOLLM – LLM framework for multi-objective molecule generation using in-context experience replay.
-
MultiMol – Two-agent LLM architecture: one learns from data, one reads literature to guide mutation of drug candidates.
-
LLM-Evolver for Polymers – Claude-3.5 driven optimizer outperforms traditional Bayesian methods for designing polymer sequences.
-
ChemLatica – Family of small chemistry LLMs trained on 100M property-labeled molecules. Combines with evolutionary prompts.
-
VALID-Mol – Incorporates chemical validation filtering into LLM-driven evolutionary molecular design.
-
EvoPrompt – Uses evolutionary search with LLM-informed mutation to discover superior prompts for >30 NLP tasks.
-
Promptbreeder – Self-mutating prompt generation. LLM breeds new prompts using self-descriptive meta prompts.
-
Vision-Language EvoPrompt – GA optimized prompts induced emergent tool-use behavior in multimodal LLMs.
-
Bayesian Prompt Optimization – BO in prompt space, using embeddings + surrogate models for search.
-
LLM Policy Evolvers (various) – Early-stage research into using LLMs to generate, mutate, or hybridize RL policies.
Have a framework to add? Submit a PR with the canonical link, a 1–3 sentence summary, and a domain tag.
MIT.