Integrate PACEvolve's Core Mechanisms to Enhance OpenEvolve's Long-Horizon Evolutionary Search Capabilities

[Cying212Jack]

I recently came across the PACEvolve framework (arXiv:2601.10657, https://arxiv.org/abs/2601.10657), a state-of-the-art progress-aware consistent evolution solution for LLM-in-the-loop evolutionary search, and it addresses key pain points that current evolutionary search systems (like OpenEvolve) face—Context Pollution, Mode Collapse, and Weak Collaboration.

PACEvolve’s three core innovative modules bring systematic improvements to evolutionary search management, which I believe could be game-changing for OpenEvolve if integrated:

1. Hierarchical Context Management (HCM): Decouples abstract "ideas" from concrete "solutions" and prunes low-quality experimental history to keep the context clean, avoiding the model’s thinking being overwhelmed by failed cases and eliminating Context Pollution.
2. Momentum-Based Backtracking (MBB): Monitors search progress in real time; when stagnation in local optima (momentum stall) is detected, it forces backtracking to previous states, effectively breaking Mode Collapse and enabling escape from local minima.
3. Adaptive Sampling for Collaborative Evolution (CE): Intelligently decides between internal backtracking exploration and cross-trajectory crossover collaboration, striking an optimal balance between exploration of new directions and exploitation of existing knowledge, and solving the Weak Collaboration issue from rigid crossover strategies.

Notably, PACEvolve has demonstrated significant performance gains over OpenEvolve in benchmarks, and also achieved SOTA results on LLM-SR and KernelBench, even discovering solutions that surpass the record on Modded NanoGPT—proving its practical value and effectiveness for long-horizon self-improvement in evolutionary search.

I’m curious to know if the maintainers have any plans to adapt and integrate these PACEvolve core features into OpenEvolve to boost its evolutionary search capabilities for LLMs? Would love to hear thoughts on the feasibility and potential roadmap for this integration!