PACEvolve: Enabling Long-Horizon Progress-Aware Consistent Evolution

TL;DR

PACEvolve addresses instability in LLM-in-the-loop evolutionary search by identifying Context Pollution, Mode Collapse, and Weak Collaboration as core failure modes. It introduces a principled framework with Hierarchical Context Management, Momentum-Based Backtracking, and Self-Adaptive Collaborative Evolution to manage context, momentum, and cross-island coordination. Through experiments on Symbolic Regression (LLM-SR), KernelBench, and Modded NanoGPT, it achieves state-of-the-art results and demonstrates robust long-horizon self-improvement. The work provides a practical, scalable recipe for building robust, knowledge-guided evolutionary agents using LLMs.

Abstract

Large Language Models (LLMs) have emerged as powerful operators for evolutionary search, yet the design of efficient search scaffolds remains ad hoc. While promising, current LLM-in-the-loop systems lack a systematic approach to managing the evolutionary process. We identify three distinct failure modes: Context Pollution, where experiment history biases future candidate generation; Mode Collapse, where agents stagnate in local minima due to poor exploration-exploitation balance; and Weak Collaboration, where rigid crossover strategies fail to leverage parallel search trajectories effectively. We introduce Progress-Aware Consistent Evolution (PACEvolve), a framework designed to robustly govern the agent's context and search dynamics, to address these challenges. PACEvolve combines hierarchical context management (HCM) with pruning to address context pollution; momentum-based backtracking (MBB) to escape local minima; and a self-adaptive sampling policy that unifies backtracking and crossover for dynamic search coordination (CE), allowing agents to balance internal refinement with cross-trajectory collaboration. We demonstrate that PACEvolve provides a systematic path to consistent, long-horizon self-improvement, achieving state-of-the-art results on LLM-SR and KernelBench, while discovering solutions surpassing the record on Modded NanoGPT.

Figures (5)

• Figure 1: We show the overall workflow of $\mathop{\mathtt{PACEvolve}}\limits$. More details about each module can be found in Figure \ref{['fig:idea_gen']}.
• Figure 2: This figure demonstrates the core components of $\mathop{\mathtt{PACEvolve}}\limits$. We decouple idea generation from idea selection to enable easy hierarchical management of idea memory (§\ref{['sec:memory']}). We also design momentum-based self-adaptive backtracking (§\ref{['sec:momentum']}) and crossover sampling mechanisms (§\ref{['sec:crossover']}) to foster long-horizon reasoning in evolutionary search and escaping local minima.
• Figure 3: We show three prototypical trajectories from 10 independent trials. If the search process does not converge quickly to a good answer in the first few iterations, it remains in a local minima for the rest of the search. Variance across runs is also large.
• Figure 4: Cumulative Boxplot Comparison of $\mathop{\mathtt{PACEvolve}}\limits$ Techniques. The distribution of performance across 10 runs is shown, starting with vanilla append-only context management and progressively adding each optimization technique.
• Figure 5: Head-to-head win rate comparison. Win rate percentages are shown for all method pairs, indicating the proportion of kernels where the row method outperformed the column method. Equal counts as a win for both methods; therefore, the heatmap is not strictly symmetric.