Too much apply diff fail

[xwqsl]

Currently, openEvolve use the search/replace format for llm to generate patch for code. However, in practice, we found when prompt gets longer and more complex, such a format will cause many apply diff fails, (for example , the search part is not exactly match with the original code ，e.g., missing space, put unnecessary extra comment，...), is there any solutions to deal with such a problem?

[codelion]

You can try with a better model, or a coding variation like gpt-codex. You can set diff_based_evolution as false to let the LLM update the full program.