Evolution needs parents to mix and mutate. Can an AI do that with plain-English plans instead of code?
Can LLM-based crossover and mutation work in unstructured natural language spaces?
This explores whether an LLM can stand in for the crossover and mutation steps of evolutionary search, blending and tweaking candidate solutions written in plain language, without the problem first being converted into a formal structure.
This explores whether an LLM can stand in for the crossover and mutation steps of evolutionary search, blending and tweaking candidate solutions written in plain language, without the problem first being converted into a formal structure. The corpus has one note that answers this directly, and it says yes. Mind Evolution has the LLM generate the crossover and mutation, and it keeps separate 'island' populations so the candidates don't all converge on one idea. It solves over 98% of the planning tasks tested, and it beats both best-of-N (many independent tries) and sequential revision (polishing one draft) Can evolutionary search beat sampling and revision at inference time?.
This is surprising because classic evolutionary methods need a representation you can cut and splice, such as a bit string or a parameter vector. A paragraph-long travel plan has no obvious seam to cut along. An LLM can read two partial solutions, keep what works in each, and rewrite the result as a coherent new candidate, so the model supplies the missing operator. Best-of-N never lets candidates learn from each other, and sequential revision only ever improves a single lineage. A population with recombination does both, and the islands guard against the whole population settling on one mediocre answer.
Other notes in the corpus suggest why this shouldn't work, and that makes the result more interesting. LLMs plateau at roughly 55–60% constraint satisfaction on real optimization problems regardless of model size Do larger language models solve constrained optimization better?. They can't run iterative procedures inside their own forward pass and fall back on emitting plausible-looking numbers Do large language models actually perform iterative optimization?. Even RL-tuned models lean on memorized templates and drop sharply on unfamiliar variants Do fine-tuned language models actually learn optimization procedures?. The corpus doesn't test this explanation, but one reading is that evolutionary search moves the iteration out of the model. The LLM only has to produce a plausible variant on each call, which is what it's good at. Selection across a population then does the optimizing that the model can't do internally.
The evidence is thin. There is one evolutionary paper, on planning tasks where candidates can be scored. Nothing here shows whether LLM crossover is doing meaningful recombination or just producing varied samples that selection then filters, and nothing tests open-ended text like essays or strategies where 'better' is fuzzy. So the corpus supports 'yes, when you can judge candidates', and it leaves open how far that extends.
Sources 4 notes
Mind Evolution, an evolutionary search strategy using LLM-generated crossover and mutation with island model diversity, solves 98%+ of planning tasks and significantly outperforms best-of-N and sequential revision strategies while working directly in natural language without task formalization.
Across constrained-optimization tasks, LLMs converge to ~55–60% constraint satisfaction independent of architecture, parameter count, or training regime. Reasoning models do not systematically outperform standard models, suggesting a fundamental ceiling rather than a scaling gap.
Research shows LLMs cannot perform iterative procedures in latent space. They recognize optimization problems as template-similar and emit plausible-looking but incorrect values, a failure mode that persists across model scale and training approaches.
Even GRPO-trained models show sharp performance drops on out-of-distribution variants (N-1 test sets) compared to in-distribution problems, indicating RL optimizes template-matching rather than genuine problem-solving procedures.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Can Large Language Models Reason and Optimize Under Constraints?
- Branch-Solve-Merge Improves Large Language Model Evaluation and Generation
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- Chain of Thoughtlessness? An Analysis of CoT in Planning
- Evolving Deeper LLM Thinking
- Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining
- Train Long, Think Short: Curriculum Learning for Efficient Reasoning
- ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning