Evolutionary program search guided by Large Language Models (LLMs) has emerged as a powerful paradigm for automated scientific discovery. However, current approaches are fundamentally constrained by three bottlenecks: structurally blind parent selection, sparse whole-program evaluation rewards, and static mutation operators that fail to adapt during search. We present GAE (Graph-Augmented Evolution), a framework that resolves these limitations through a tightly coupled, three-pillar architecture. First, a relational graph neural network (GNN) parses programs into typed computation graphs, producing structure-aware embeddings. Second, an RL-optimized meta-controller leverages these embeddings to replace blind evolutionary sampling with a directed policy, dynamically selecting optimal parents and mutation directions based on reward history. Third, an online GRPO fine-tuning loop continuously updates the LLM mutation operator at test-time using group-normalized evaluation rewards, directly aligning the model's generation distribution with high-fitness structural edits. We evaluate GAE on a challenging scientific discovery task: symbolic regression for complex nonlinear oscillator systems. By transforming stochastic search into a directed, self-improving trajectory, GAE efficiently discovers closed-form physical equations, consistently matching or outperforming static LLM-driven baselines and achieving state-of-the-art out-of-distribution performance.
Large Language Models (LLMs) are increasingly used for agent-based virtual cell modeling, yet existing frameworks rely on unstructured retrieval or static prompt engineering, injecting noisy evidence and wasting inference budget on redundant tool-use trajectories. We propose VCAgent, a self-evolving framework that optimizes how biological knowledge from external APIs is structured and integrated into LLM reasoning. VCAgent introduces a schema-based slot filling mechanism that transforms raw API responses into compact, typed representations, replacing unstructured text concatenation. To efficiently search the high-dimensional instruction space, we develop Mutation-Guided Monte Carlo Tree Search (MG-MCTS), which combines relevance-based priors derived from inter-API semantic similarity with an adaptive gating strategy to balance exploration and exploitation under a constrained evaluation budget. Experiments on single-cell perturbation prediction across five cell lines show that VCAgent improves both reasoning accuracy and search efficiency across multiple backbone LLMs, substantially outperforming existing prompt optimization methods. Codes are available at https://github.com/LZYBUPT/VCAgent.
Zhiyun Li, Rong Han, Xiao-Yong Wang et al.· Proceedings of the 32nd ACM...· 0 citations
This work proposes a general LES framework that incorporates feature-assisted niche construction within abstract search spaces, enabling the seamless integration of niche-based search strategies from evolutionary computation and introduces PartEvo (Partition to Evolve), an LES method that combines niche collaborative search and advanced prompting strategies to improve algorithm discovery efficiency.
Qinglong Hu, Qingfu Zhang· Neural Information Processin...· 9 citations· ⚡3
Neural Architecture Search (NAS) aims to automate the design of neural network architectures, reducing the need for manual expert-driven engineering. Evolutionary approaches, such as Cartesian Genetic Programming (CGP), provide a flexible graph-based representation for evolving neural structures but suffer from high computational costs due to expensive fitness evaluations. In this paper, we propose a Surrogate-Assisted Dynamic Cartesian Genetic Programming (SA-DCGP) framework for automated neural architecture design in graph clustering tasks. The framework evolves architectures composed of FastKAN-based nonlinear blocks and SGCC-style linear normalized layers using dynamic mutation and multiple crossover operators. To reduce evaluation cost, we introduce a pair-wise surrogate model that predicts whether an offspring architecture will outperform its parent based on genotype-derived features and cheap training signals. The surrogate guides selection, enabling full training only for promising candidates. We employ a two-stage evaluation protocol with a cheap training phase for surrogate feature extraction and a full training phase for selected architectures. Clustering performance is evaluated using Accuracy, NMI, ARI, and F1 scores. Experimental results on benchmark graph datasets show that SA-DCGP discovers compact and high-performing architectures while significantly reducing computational overhead, demonstrating the effectiveness of surrogate-assisted evolutionary search for graph-based representation learning.
Maciej Krzywda, Szymon Łukasik, Amir H. Gandomi· Proceedings of the Genetic a...· 0 citations
Large language models (LLMs) have emerged as a powerful tool for automated evolutionary optimization, but existing methods remain limited in pattern reuse, error-aware refinement, and retrieval robustness across diverse tasks. To address these limitations, we propose OptGraph, the first optimization agentic workflow that introduces graph retrieval-augmented generation (GraphRAG). Specifically, OptGraph first constructs reusable experience as a typed graph, capturing the relationships among modeling patterns, problem formalization, implementation details, and error corrections. In the inference stage, OptGraph leverages graph neighborhood information to enrich retrieved knowledge, providing structured context to improve modeling, verification, and iterative refinement. Moreover, OptGraph supports adaptive knowledge updates, enabling the distillation of execution traces and verification feedback into reusable graph knowledge without ndertaking LLM parameter tuning. Extensive experiments on benchmark datasets show that our proposed OptGraph achieves an average exact accuracy 8.9% higher than the state-of-the-art prompt-based automated optimization frameworks. Our code has been made available at https://github.com/xianchaoxiu/OptGraph.
Xianchao Xiu, Jianhao Li, Huangyue Chen et al.· 1 citation
Configuring genetic algorithms (GAs) for new optimization problems typically requires substantial expertise in representation design, decoder construction, and constraint handling. We present a framework that combines large language models (LLMs) for problem understanding with a structured Biased Random-Key Genetic Algorithm (BRKGA) configurator for algorithm realization, allowing users to describe optimization problems in natural language and receive executable GPU-accelerated GA implementations. We compare two workflows: LLM-Only, where the LLM generates a complete standalone GA from the problem description; and LLM+Config, where the LLM is restricted to problem parsing and decoder synthesis against a fixed brkga3 interface that provides the evolutionary architecture, Adaptive Large Neighborhood Search (ALNS) warm-start, GPU-parallel segmented sort, and bottleneck-focused local search. We evaluate on three problem families (the Traveling Salesperson Problem (TSP), the Traveling Salesperson Problem with Job-times (TSPJ), and the Vehicle Routing Problem with Resource Pick-up and Delivery (VRP-RPD)) on instances up to 999 nodes. LLM+Config outperforms or matches LLM-Only on every instance where both workflows produced a result, and additionally solves four TSP instances on which LLM-Only failed (Wilcoxon signed-rank with Holm correction across the three problem families, all adjusted p ≤ 0.012); on TSPJ it matches CPLEX optimal on 3 of 10 benchmarks and establishes first reported results for 6 large instances (n = 202–1,000); on VRP-RPD it wins all 13 instances against LLM-Only with 51–329 % advantage on large instances. GPU-parallel BRKGA evolution runs 25–35× faster per generation than a single-threaded CPU baseline. We characterise LLM stochasticity through two complementary studies: 15 sessions across 3 problems with one model (80% success rate), and 100 sessions across 10 LLMs from 3 providers on TSPJ/eil101 (Friedman and Wilcoxon-Holm tests on within-tier and within-provider blocks), finding that capability tier dominates over provider at the large tier and that small-tier non-Anthropic models contribute zero usable sessions.
Harishjitu Seesandrn, M. Sodhi, Resit Sendag· Proceedings of the Genetic a...· 0 citations
Parent selection significantly affects exploration, exploitation, and complexity control in genetic programming (GP) for symbolic regression. It is unclear whether large language models (LLMs) can synthesize effective operators in a zero-shot setting without iterative meta-evolution. Here, zero-shot means that the model receives only the task description, with no reference operators or iterative feedback. In this work, we benchmark zero-shot synthesis of parent-selection operators across eight LLMs within a standard GP framework for symbolic regression. Each model receives the same natural-language prompt to generate a parent-selection operator, which is then evaluated in a standard GP framework with only the parent-selection operator replaced, while all other components and the evolutionary-search budget are held constant. For each LLM, ten independent zero-shot operators are evaluated on twelve OpenML regression benchmarks and compared against automatic lexicase and tournament selection baselines. Claude Sonnet~4.6 and Gemini~3.1 Pro stand out for consistently strong performance on both training and held-out test $R^2$. The strongest operator in our benchmark---a Kimi~K2.5 zero-shot synthesis---surpasses the automatic lexicase and tournament baselines in search effectiveness. These results suggest that zero-shot LLM synthesis is a viable approach to generating competitive GP selection operators. Analysis shows that many generated operators use semantics to guide selection, suggesting that LLMs can produce non-trivial search heuristics from the task description alone. We also examine the relationship between public LLM leaderboard rankings and GP performance. Widely used benchmarks, such as Humanity's Last Exam and SWE-bench Verified, strongly correlate with training $R^2$, while their relationship to held-out test $R^2$ is weaker and less clear.
Hengzhe Zhang, Qi Chen, Bing Xue et al.· 1 citation