来源：公众号“暴躁短腿猫”
原文链接：https://mp.weixin.qq.com/s/sd0kN5yHCIRyn20XA5mAfw
从 6,520 篇录用论文中，看进化算法如何进入 AI 核心系统。
一个很清晰的信号是：进化算法不再只是 NAS、超参调优或传统黑盒优化里的工具，而正在进入 LLM 安全、Agent 系统、科学发现和 AI Systems 的核心链路。
01 三个方向的数量分布
02 优化方法正在渗透哪些 AI 领域？
25 篇｜Optimization Theory
13 篇｜LLM Pretraining & Architecture
12 篇｜LLM Safety & Alignment
8 篇｜Reinforcement Learning
7 篇｜Probabilistic & Bayesian
7 篇｜Time Series Analysis
7 篇｜Agent & Embodied AI
6 篇｜LLM Evaluation & Benchmark
4 篇｜Image Generation
3 篇｜LLM Reasoning & Post-training
2 篇｜ML Systems & Infrastructure
2 篇｜Computer Vision
2 篇｜Data-Centric Learning
2 篇｜AI for Science
2 篇｜Model Compression & Efficiency
2 篇｜Graph & Structured Data
3 篇｜其他
这说明优化方法已经明显跨出 Optimization Theory 的传统边界，进入 LLM Pretraining、LLM Safety、Agent、Time Series、AI for Science、Model Compression 等多个方向。
03 五大范式转向
转向 1：进化算法从“优化工具”升级为“AI 系统组件”
判断：进化算法正在从辅助优化工具，转向 AI 系统自身的设计组件。
转向 2：Bayesian Optimization 进入 LLM 时代
判断：BO 的核心问题正在从“如何建 GP”转向“如何利用强生成模型与语言反馈”。
转向 3：符号回归成为进化算法与 LLM 的交汇战场
判断：LLM 指导的进化搜索，可能成为科学 AI 的重要范式。
转向 4：L2O 进入 MetaBBO 新阶段
判断：Learning to Optimize 不再只学梯度更新规则，而是在学习搜索策略本身。
转向 5：进化算法与 LLM 安全交叉爆发
判断：进化攻击已经出现，进化防御大概率会成为后续蓝海。
04 完整论文清单
? 昂贵优化 / Bayesian Optimization（49 篇）
01 Regular
Taking the GP Out of the Loop
核心贡献：BO 从少观测→多观测场景的范式迁移
02 Regular
Divide and Learn: Multi-Objective Combinatorial Optimization at Scale
核心贡献：大规模多目标组合优化，在线学习重铸
03 Regular
Local Constrained Bayesian Optimization (LCBO)
核心贡献：高维约束 BO 的局部搜索框架
04 Spotlight
Training Diffusion Language Models for Black-Box Optimization
核心贡献：**扩散语言模型用于离线黑盒优化**（材料/DNA/机器人设计）
05 Regular
LILO: Bayesian Optimization with Natural Language Feedback
核心贡献：LLM 作为 BO 中的主观偏好反馈代理
06 Regular
Budgeted Active Experimentation for Treatment Effect Estimation
核心贡献：有限 RCT 预算下的主动实验设计
07 Regular
Many Needles in a Haystack: Active Hit Discovery for Perturbation Experiments
核心贡献：基因扰动实验中的纯探索策略
08 Regular
Regime-Adaptive Bayesian Optimization via Dirichlet Process Mixtures
核心贡献：多区域问题的自适应 GP 混合
09 Regular
Certificate-Guided Pruning for Stochastic Lipschitz Optimization
核心贡献：有噪声评估下的黑盒优化可证明保证
10 Regular
Breaking Multi-Task Curse: Reward-Weighted Evolution for BB-Many-Task Optimization
核心贡献：多任务黑盒优化中的负迁移问题
11 Regular
Meta-Black-Box Optimization Can Do Search Guidance for Expensive Constrained MOP
核心贡献：双层 MetaBBO 框架用于昂贵约束多目标优化
12 Regular
Evolving Interpretable Constitutions for Multi-Agent Coordination
核心贡献：进化框架发现多 Agent 系统行为规范
13 Regular
DecAEvolve: Decompose, Adapt, and Evolve for LLM-based Scientific Equation Discovery
核心贡献：进化搜索 + LLM 科学知识发现方程
14 Regular
Towards Diverse Scientific Hypothesis Search with LLMs
核心贡献：多样性科学假设空间的迭代搜索
15 Regular
MoCo-EA: Exploiting Adversarial Mode Connectivity for Efficient Evolutionary Attacks
核心贡献：模式连接增强进化对抗攻击
16 Regular
AutoMS: Multi-Agent Evolutionary Search for Cross-Physics Inverse Microstructure Design
核心贡献：多物理场逆问题中的进化搜索
17 Regular
KernelFoundry: Hardware-Aware Evolutionary GPU Kernel Optimization
核心贡献：进化算法优化 GPU Kernel
18 Regular
Depth over Fidelity in Fixed-Budget Noisy Evolution Strategies
核心贡献：噪声 ES 中深度与保真度的权衡
19 Regular
EvoMAS: Heuristics in the Loop — Evolving Smarter Agentic Workflows
核心贡献：进化 MAS 工作流设计
20 Regular
MFH-NAS: A Hybrid Neural Architecture Search Framework
核心贡献：进化 + NAS 多模态融合架构搜索
21 Regular
Towards Execution-Grounded Automated AI Research
核心贡献：自动 AI 研究中的执行反馈优化
22 Regular
Breaking the Simplification Bottleneck in Amortized Neural Symbolic Regression
核心贡献：符号回归超越遗传规划
23 Regular
Contrastive Symbolic Regression
核心贡献：进化符号回归 + 对比学习
24 Regular
Iterated Population Based Training with Task-Agnostic Restarts
核心贡献：PBT 超参数优化的重启策略
25 Regular
Neural-Evolutionary Symbolic Regression with Global Constraints
核心贡献：神经+进化的约束符号回归
26 Regular
HyMTRL: A Hybrid Multi-Task RL Framework via Phased Policy Evolution
核心贡献：多任务 RL 的分阶段策略进化
27 Regular
A₂DEPT: LLM-Driven Automated Algorithm Design via Evolutionary Program Trees
核心贡献：LLM + 进化程序树的自动算法设计
28 Regular
Towards A Generative Protein Evolution Machine with DPLM-Evo
核心贡献：蛋白质进化生成模型
29 Regular
Dream-MPC: Gradient-Based MPC with Latent Imagination
核心贡献：基于梯度的 vs 基于种群的规划对比
30 Regular
Inducing Overthink: Hierarchical GA-based DoS Attack on Black-Box Reasoning Models
核心贡献：遗传算法攻击大推理模型
31 Regular
Reasoning as an Attack Surface: Adaptive Evolutionary CoT Jailbreaks
核心贡献：进化 CoT 越狱攻击
32 Regular
ECCO: Evidence-Driven Causal Reasoning for Compiler Optimization
核心贡献：编译器自动调优的因果推理
33 Regular
Order Matters: Proxy-Guided LLM Evolution for Macro Placement
核心贡献：VLSI 布局中的 LLM 进化搜索
34 Regular
Flow for Future: SE(3)-Equivariant Flow Matching for 3D Trajectory Prediction
核心贡献：流匹配优化物理轨迹预测
35 Regular
AutoQRA: Joint Optimization of Mixed-Precision Quantization and Low-rank Adapters
核心贡献：量化+微调的联合优化
36 Regular
CoGenCast: Coupled Autoregressive-Flow Generative Framework for Time Series
核心贡献：自回归+流模型联合时间序列预测
37 Regular
PULSE: Generative Phase Evolution for Non-Stationary Time Series
核心贡献：非平稳时序的相位进化
38 Regular
FlowCloud: Learning Continuous Spatiotemporal Dynamics
核心贡献：最优传输的连续动力学学习
39 Regular
Hierarchical Anchor Graph Learning for Multi-View Clustering
核心贡献：多层锚点图优化多视角聚类
40 Regular
PipeSD: Cloud-Edge Collaborative Pipeline Inference with Speculative Decoding
核心贡献：投机解码的云端-边缘协同优化
41 Regular
Sparse Relaxed-Lasso Steering: SAE Feature Selection for Image Editing
核心贡献：稀疏特征选择的 Lasso 松弛
42 Regular
ScoreMix: Synthetic Data Generation by Score Composition
核心贡献：扩散模型分数合成优化
43 Spotlight
Diffuse to Detect: Bi-Level Sample Rebalancing for Point-Supervised IR Detection
核心贡献：二值层优化解决样本不均衡
44 Spotlight
On the origin of neural scaling laws
核心贡献：随机图→自然语言中的标度律起源
45 Spotlight
Learning Randomized Reductions
核心贡献：学习随机自归约
46 Spotlight
Latent Laplace Diffusion for Irregular Multivariate Time Series
核心贡献：不规则时序的拉普拉斯扩散
47 Regular
Predicting What Matters: Robust Generalist Robot Policy
核心贡献：世界模型策略优化
48 Regular
MalTree: Tracing Malware Evolution using Embeddings at Scale
核心贡献：恶意软件进化谱系追踪
49 Regular
Ambient Dataloops: Generative Models for Dataset Refinement
核心贡献：数据集质量的迭代精炼
? 进化算法 / Evolutionary Algorithms（52 篇）
01 Regular
MoCo-EA: Exploiting Adversarial Mode Connectivity for Efficient Evolutionary Attacks
核心贡献：模式连接增强进化攻击效率
02 Regular
DecAEvolve: Decompose, Adapt, and Evolve for Scientific Equation Discovery
核心贡献：进化搜索+LLM科学发现
03 Regular
EvoMAS: Heuristics in the Loop — Evolving Smarter Agentic Workflows
核心贡献：进化多Agent系统工作流
04 Regular
KernelFoundry: Hardware-Aware Evolutionary GPU Kernel Optimization
核心贡献：硬件感知的GPU Kernel进化
05 Regular
Depth over Fidelity in Fixed-Budget Noisy Evolution Strategies
核心贡献：噪声ES预算分配策略
06 Regular
Breaking Multi-Task Curse: Reward-Weighted Evolution for BB-Many-Task
核心贡献：多任务进化优化的负迁移
07 Regular
Iterated Population Based Training with Task-Agnostic Restarts
核心贡献：PBT通用重启策略
08 Regular
Contrastive Symbolic Regression: Aligned Representations + Diverse Ensembles
核心贡献：对比学习+进化符号回归
09 Regular
Neural-Evolutionary Symbolic Regression with Global Constraints
核心贡献：神经进化符号回归带约束
10 Regular
A₂DEPT: LLM-Driven Automated Algorithm Design via Evolutionary Program Trees
核心贡献：LLM+进化程序树自动算法设计
11 Regular
Inducing Overthink: Hierarchical GA-based DoS Attack on LLMs
核心贡献：遗传算法DoS攻击推理模型
12 Regular
Reasoning as an Attack Surface: Adaptive Evolutionary CoT Jailbreaks
核心贡献：进化式思维链越狱
13 Regular
Towards A Generative Protein Evolution Machine
核心贡献：蛋白质序列的进化生成模型
14 Regular
AutoMS: Multi-Agent Evolutionary Search for Microstructure Design
核心贡献：多Agent进化材料设计
15 Regular
HyMTRL: Phased Policy Evolution for Multi-Task RL
核心贡献：多任务RL分阶段进化
16 Regular
On the Generalization Gap in Self-Evolving LLM Reasoning
核心贡献：LLM自我进化推理的泛化差距
17 Regular
Towards Execution-Grounded Automated AI Research
核心贡献：AI研究的进化式执行反馈
18 Regular
Evolving Interpretable Constitutions for Multi-Agent Coordination
核心贡献：多Agent协调规范的进化
19 Regular
Order Matters: Proxy-Guided LLM Evolution for Macro Placement
核心贡献：VLSI布局的LLM进化搜索
20 Regular
ECCO: Evidence-Driven Causal Reasoning for Compiler Optimization
核心贡献：编译器优化的因果推理+进化
21 Regular
Flow for Future: SE(3)-Equivariant Flow Matching for 3D Trajectory
核心贡献：三维轨迹进化预测
22 Spotlight
On the origin of neural scaling laws: from random graphs to natural language
核心贡献：标度律起源的演化视角
23 Spotlight
Learning Randomized Reductions
核心贡献：学习随机自归约发现
24 Regular
Dream-MPC: Gradient-Based MPC with Latent Imagination
核心贡献：梯度 vs 群体规划方法对比
25 Regular
Evolving Interdependent Operators with Large Language Models for Multi-Objective Combinatorial Optimization
核心贡献：LLM进化算子协同
26 Spotlight
A Unifying View of Variational Generative Wasserstein Flows
核心贡献：分布进化流统一
27 Regular
DiffCrossGait: Trajectory-Level Alignment for 2D-3D Cross-Modal Gait Recognition via Latent Diffusion
核心贡献：扩散轨迹演化对齐
28 Regular
Position: Safe AI Should be Resistant and Resilient in an Evolving World
核心贡献：安全AI协同进化
29 Regular
TSFAdv: Frequency-Guided Black-Box Adversarial Attacks on Time Series Forecasting
核心贡献：频域进化黑盒攻击
30 Regular
LineageFlow: Flow Matching for High-Fidelity Family-Aware Protein Sequence Generation
核心贡献：蛋白序列进化生成
31 Regular
Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
核心贡献：ES扩展LLM微调
32 Regular
Evolution Strategies at the Hyperscale
核心贡献：低秩ES超大规模
33 Regular
Bridging Spherical Black-Box Optimizers
核心贡献：ES黑盒优化统一
34 Regular
Spectral Evolution Search: Efficient Inference-Time Scaling for Reward-Aligned Image Generation
核心贡献：频谱进化图像搜索
35 Spotlight
Latent Laplace Diffusion for Irregular Multivariate Time Series
核心贡献：扩散演化时序建模
36 Regular
RSTR: Reducing SpatioTemporal Redundancy in Diffusion Transformers
核心贡献：扩散演化冗余削减
37 Regular
PULSE: Generative Phase Evolution for Non-Stationary Time Series Forecasting
核心贡献：相位进化时序预测
38 Regular
FlowCloud: Learning Continuous Spatiotemporal Dynamics from Unpaired Sparse Point Cloud Snapshots
核心贡献：点云时空演化学习
39 Regular
Predicting What Matters: Robust Generalist Robot Policy Learning via Future Semantic Mask
核心贡献：机器人策略演化鲁棒
40 Regular
Towards Diverse Scientific Hypothesis Search with Large Language Models
核心贡献：LLM进化假设搜索
41 Regular
Meta-Black-Box Optimization Can Do Search Guidance for Expensive Constrained Multi-Objective Optimization
核心贡献：元进化搜索引导
42 Regular
MalTree: Tracing Malware Evolution using Embeddings at Scale
核心贡献：恶意软件进化追踪
43 Regular
DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models
核心贡献：扩散演化多模推理
44 Regular
Future Dynamic 3D Reconstruction: A 3D World Model with Disentangled Ego-Motion
核心贡献：3D世界演化重建
45 Regular
Hierarchical Anchor Graph Learning for Multi-View Clustering
核心贡献：层次图演化聚类
46 Regular
Unfolding Generative Flows with Koopman Operators: Trajectory-Preserving Linearization
核心贡献：生成流演化线性化
47 Regular
MFH-NAS:A Hybrid Neural Architecture Search Framework for Multimodal Fusion Object Detection
核心贡献：进化NAS融合搜索
48 Regular
Breaking the Simplification Bottleneck in Amortized Neural Symbolic Regression
核心贡献：进化符号回归化简
49 Regular
AutoQRA: Joint Optimization of Mixed-Precision Quantization and Low-rank Adapters for Efficient LLM Fine-Tuning
核心贡献：进化量化秩搜索
50 Regular
CoGenCast: A Coupled Autoregressive–Flow Generative Framework for Time Series Forecasting
核心贡献：生成流演化预测
51 Regular
Ambient Dataloops: Generative Models for Dataset Refinement
核心贡献：数据模型协同进化
52 Regular
Marrying Generative Model of Healthcare Events with Digital Twin of Human-Environment Interaction for Disease Reasoning
核心贡献：医疗孪生演化推理
? Learning to Optimize（6 篇）
01 Regular
Meta-Black-Box Optimization Can Do Search Guidance for Expensive Constrained MOP
核心贡献：MetaBBO 双层学习搜索引导
02 Spotlight
Training Diffusion Language Models for Black-Box Optimization
核心贡献：扩散LM学习黑盒优化策略
03 Regular
Iterated Population Based Training with Task-Agnostic Restarts
核心贡献：学习通用PBT重启
04 Regular
Breaking Multi-Task Curse: Reward-Weighted Evolution for BB-Many-Task Optimization
核心贡献：学习多任务间的知识迁移
05 Regular
A₂DEPT: LLM-Driven Automated Algorithm Design
核心贡献：LLM学习生成优化算法
06 Regular
LILO: Bayesian Optimization with Natural Language Feedback
核心贡献：LLM学习提供BO反馈
05 未来热门方向预测
Tier 1：2027 年即可爆发
LLM + 进化算法协同
扩散模型替代 GP 的生成式 BBO
进化多 Agent 系统设计
LLM 辅助 Bayesian Optimization
Tier 2：1-3 年逐步成熟
面向 LLM 安全的进化攻击与防御
科学 AI 中的进化方程发现
多目标昂贵优化
MetaBBO 自动化优化器设计
Tier 3：3-5 年前沿探索
自进化 AI 系统
进化式 AI-Native 科学发现闭环
超低预算进化策略
06 给进化计算研究者的五条建议
数据说明：本文基于 ICML 2026 全量 6,520 篇录用论文（OpenReview 数据）进行关键词筛选与方向归类。
原文链接: https://www.leiphone.com/category/private/BNk7EnQQMSPaca49.html
