DeltaEvolve: Accelerating Scientific Discovery through Momentum-Driven Evolution
DeltaEvolve employs structured semantic delta and multi-level database to improve program evolution efficiency, reducing token use by 36.79%.
Key Findings
Methodology
This paper formalizes LLM-driven evolutionary agents within a general Expectation-Maximization (EM) framework, where the E-step involves sampling candidate programs conditioned on current context, and the M-step updates the control context based on evaluation feedback. Traditional approaches rely on full-code snapshots, which contain redundant details that dilute core algorithmic ideas, hindering effective transfer and guidance. To address this, the authors introduce semantic delta—structured representations of the logical and strategic changes between successive nodes—that serve as a momentum-like memory signal. They organize these deltas in a multi-level database, including high-level summaries and detailed plans, and employ a progressive disclosure mechanism that adaptively exposes relevant information based on recency and relevance. Empirical results across diverse scientific tasks demonstrate that DeltaEvolve discovers superior solutions with significantly fewer tokens, outperforming baseline full-code methods in both efficiency and solution quality.
Key Results
- In black-box optimization over 100 iterations, DeltaEvolve achieved a 36.79% reduction in total token consumption while improving the best solution score by up to 15% compared to AlphaEvolve.
- Across five scientific domains—such as symbolic regression, PDE solving, and geometric packing—DeltaEvolve consistently matched or exceeded baseline performance, demonstrating strong generalization.
- Ablation studies reveal that the selection mechanism for high-quality solutions plays a more critical role than scalar feedback, emphasizing the importance of effective context construction based on program similarity and evaluation scores.
Significance
This work advances the field of automated scientific discovery by introducing a highly efficient, transfer-friendly memory mechanism—semantic delta—that overcomes the limitations of full-code storage. It enables LLM-based systems to operate with limited context while maintaining or improving solution quality, paving the way for scalable, multi-task, long-horizon AI-driven research. The approach addresses core challenges in knowledge retention, transfer, and resource efficiency, making automated science more practical and accessible for complex, real-world problems.
Technical Contribution
The key technical innovation lies in transforming the context representation from static full programs to structured semantic deltas, which encapsulate the core logical modifications. This is supported by a multi-level database architecture that stores high-level summaries and detailed plans, combined with a progressive disclosure mechanism to optimize token usage. The formalization of the process within an EM framework clarifies the role of the M-step as a gradient-like update driven by semantic changes, providing a solid theoretical foundation. The integration of these components results in a scalable, transfer-capable memory system that significantly enhances the efficiency and effectiveness of LLM-guided program evolution.
Novelty
This paper is the first to formalize program evolution as an EM process and to replace full-code snapshots with structured semantic deltas organized in a multi-level database. The combination of semantic change representation, progressive disclosure, and momentum-inspired updates constitutes a novel paradigm in AI-driven scientific discovery, setting a new standard for resource-efficient, transfer-aware program optimization.
Limitations
- The approach assumes programs are decomposable into reusable components; highly entangled or monolithic programs may not benefit as much from semantic delta extraction.
- Scaling the multi-level database for extremely large or complex tasks could pose storage and retrieval challenges, requiring further engineering solutions.
- Current methods are primarily validated on static, well-structured tasks; dynamic or real-time scientific environments need additional adaptation for robustness.
Future Work
Future research will explore richer semantic representations, incorporating causal reasoning and external knowledge bases like knowledge graphs to enhance the expressiveness of semantic delta. Adaptive multi-level storage strategies and more sophisticated selection policies will be developed to further improve efficiency. Additionally, extending the framework to multi-agent systems and real-time environments could unlock broader applications in scientific automation and industrial AI.
AI Executive Summary
Automated scientific discovery has long been a goal of AI research, promising to accelerate breakthroughs across disciplines. Recent advances leverage large language models (LLMs) to generate, evaluate, and refine programs that solve complex problems. Systems like AlphaEvolve have demonstrated the potential of this approach, but they rely heavily on storing entire program histories, which introduces significant inefficiencies and weakens guidance. As program length and complexity grow, maintaining comprehensive code snapshots becomes impractical, especially under limited context windows.
Jiang et al. address this challenge by proposing DeltaEvolve, a novel framework that replaces full program storage with structured semantic delta—compact, meaningful representations of the core logical and strategic changes between successive nodes. This approach draws inspiration from momentum-based optimization, where accumulated directional changes guide future steps more effectively. The authors organize these semantic deltas in a multi-level database, including high-level summaries and detailed plans, and employ a progressive disclosure mechanism that dynamically adjusts the granularity of stored information based on relevance and recency.
The theoretical foundation of DeltaEvolve is grounded in formalizing the evolution process as an Expectation-Maximization (EM) framework. In this setting, the E-step involves sampling candidate programs conditioned on the current context, while the M-step updates the context based on evaluation feedback to maximize the expected reward. Unlike traditional methods that rely on static code snapshots, this formulation emphasizes the importance of semantic changes as the driving force behind program improvement.
Empirical evaluations across five diverse scientific domains—such as symbolic regression, PDE solving, and geometric packing—demonstrate that DeltaEvolve consistently outperforms baseline approaches. Notably, it achieves comparable or superior solution quality while reducing token consumption by approximately 36.79%. These results highlight the efficiency and transferability of semantic delta as a memory mechanism, enabling scalable, resource-efficient AI systems for scientific discovery.
Furthermore, ablation studies reveal that the selection mechanism for high-quality solutions plays a more vital role than scalar feedback alone, emphasizing the importance of intelligent context construction. Overall, DeltaEvolve offers a significant step forward in resource-efficient, scalable, and effective AI-driven scientific research. Future directions include integrating richer semantic representations, adaptive multi-level storage, and broader multi-agent collaboration, aiming to realize fully autonomous, long-horizon scientific exploration.
Deep Analysis
Background
The pursuit of automated scientific discovery has evolved significantly over the past decades, driven by advances in optimization algorithms, symbolic reasoning, and machine learning. Early systems relied on symbolic AI and heuristic search methods like genetic algorithms, which could handle well-structured problems but struggled with scalability and complexity. The advent of deep learning, especially large language models (LLMs) such as GPT-3 and Codex, revolutionized program synthesis and code generation, enabling systems like AlphaEvolve to generate and refine programs iteratively. These systems demonstrated promising results in domains like mathematical optimization, physical simulations, and molecular design.
However, a common limitation persisted: reliance on full program snapshots as context, which leads to exponential growth in input size, redundant information, and difficulty in extracting transferable knowledge. As programs become longer and more complex, the models face the challenge of maintaining effective memory and guidance over multiple iterations. Recent research has explored various context engineering techniques, including prompt compression, memory augmentation, and selective retrieval, but these are often generic and do not explicitly focus on the core logical changes that drive program improvement. Consequently, there remains a pressing need for more efficient, structured memory representations that can capture the essence of program evolution and facilitate transfer across tasks.
Core Problem
Despite the successes of LLM-driven program evolution, existing systems like AlphaEvolve are constrained by their dependence on full code snapshots, which contain大量冗余信息，导致上下文容量限制和效率瓶颈。程序的复杂性和长度不断增加，使得存储完整代码变得不切实际，尤其是在多轮、多任务环境中，模型难以有效利用历史信息。此外，完整代码中包含大量与核心算法无关的实现细节，干扰模型提取有用的迁移模式，限制了算法的指导能力。如何在有限的上下文空间内，提取、存储和利用程序演化中的关键逻辑变化，成为提升自动科学发现系统性能的关键。解决这一问题，不仅关系到系统的资源效率，也直接影响其在复杂科学任务中的应用潜力。
Innovation
本文的核心创新在于引入结构化的语义变化（semantic delta），作为程序演化的高效记忆单元。具体而言，作者将连续节点之间的变化描述为高层次的策略总结（delta summary）和详细的逻辑变更（delta plan），避免存储冗余的实现细节。这些变化被组织在多层数据库中，结合渐进披露机制，根据节点的相关性和新颖性动态调整信息的详细程度，从而在保证信息丰富的同时，大幅度降低输入令牌数。理论上，将演化过程形式化为期望最大化（EM）框架，明确了在有限上下文条件下，利用语义变化引导程序优化的路径。这一创新不仅提升了记忆的迁移能力，也为程序演化提供了更具指导性的逻辑基础。
Methodology
- �� 将程序演化过程建模为期望最大化（EM）框架：在E步中，模型根据当前上下文采样候选程序；在M步中，系统根据评估反馈优化上下文。
- �� 传统方法依赖完整程序快照，存储大量冗余信息，影响效率和迁移。
- �� 引入结构化语义变化（semantic delta），描述连续节点间的核心逻辑变更，作为记忆单元，避免冗余。
- �� 设计多层数据库（pyramid structure），存储高层策略总结（delta summary）和详细逻辑变更（delta plan），以及完整程序。
- �� 使用渐进披露机制，根据节点的相关性和新颖性，动态调整信息的详细程度，优化上下文内容。
- �� 在每次迭代中，模型结合历史变化信息和评估反馈，更新上下文，指导下一轮采样。
- �� 通过在多个科学任务中的实验验证，系统实现了高效搜索和优质解的发现。
Experiments
作者在符号回归、偏微分方程求解、几何优化、神经网络结构搜索和化学分子设计等五个科学领域设计了丰富的实验，比较DeltaEvolve与AlphaEvolve的性能差异。采用的指标包括最高目标值、平均目标值、令牌消耗和收敛速度。实验中，模型在不同的上下文规模（如Top-k、多样性比例）下运行，验证了语义delta在提升效率和效果方面的优势。还进行了消融研究，分析去除数值评分或变化选择机制的影响。结果显示，DeltaEvolve在保持或超越最优解的同时，令牌消耗平均降低36.79%，在复杂任务中表现出更强的泛化能力和稳定性。
Results
具体数据表明，DeltaEvolve在黑箱优化任务中，100轮迭代后，最高目标值提升了15%，令牌消耗减少36.79%，显著优于AlphaEvolve。符号回归任务中，平均目标值提升了10%以上。偏微分方程求解中，解的精度提高了8%，同时减少了约40%的计算资源。消融实验进一步验证了变化机制的重要性，去除数值评分后性能下降不明显，但仅依赖高质量程序选择策略即可保持优异表现。这些结果充分证明了语义变化作为记忆单元的有效性和迁移能力，为未来大规模自动科学探索提供了技术基础。
Applications
该方法适用于多种科学研究场景，包括自动优化算法设计、复杂系统模拟、材料科学中的分子结构搜索等。系统要求具备良好的程序可分解性和评估机制，能在有限的上下文中高效引导搜索。其优势在于减少资源消耗、提升搜索效率，特别适合资源有限或需要长时间跨度探索的科研项目。未来，结合知识图谱和因果推理，有望实现更智能的科学发现辅助工具，推动AI在基础科学和工业应用中的深度融合。
Limitations & Outlook
目前方法主要依赖程序的可分解性；对于高度耦合或非结构化程序，语义变化的提取可能不够准确。此外，存储和检索多层变化信息在极大规模任务中仍存在规模瓶颈，需进一步优化存储结构。系统未充分考虑动态环境和多任务场景下的适应性，未来需引入自适应机制以增强鲁棒性。同时，模型在极端复杂任务中的表现仍有待提升，需结合更强的推理和知识整合能力。
Plain Language Accessible to non-experts
想象你在厨房里做菜。每次你尝试一道菜，都会根据味道调整配料和烹饪方法。你记住了哪些调整带来了更好的味道，哪些失败了，然后在下一次做菜时，参考这些经验，逐步改进。DeltaEvolve就像是厨房里的智能助手，它不是简单地记住每次做菜的完整步骤，而是专注于那些真正改变味道的关键调整——比如放多了盐或少了糖。通过记录这些“关键变化”，它可以更快地找到做出美味菜肴的方法，而不用每次都重新记住所有细节。这就像是你在不断学习，逐渐变成厨艺大师，效率更高，效果更好。它还会用不同的层次来存储信息——有时候只记住大概的改动，有时候记住具体的步骤细节，就像用不同的笔记本记录不同的内容一样。这样，整个过程既省时又高效，能帮助你在厨房里做出越来越棒的菜肴。
ELI14 Explained like you're 14
想象你在学校里参加一个科学比赛，你要设计一个能自动解决数学题的机器人。刚开始，你可能会写一大堆代码，把所有可能的解法都写进去。可是，这样的代码太长太复杂，机器人很难理解，也很慢。后来，你发现其实只需要记住一些关键的变化，比如“把这个公式改成这个样子”，就能让机器人变得更聪明。DeltaEvolve就像是这个过程中的聪明助手，它不记住所有的代码细节，而是专注于那些真正让机器人变得更厉害的关键变化。它会把这些变化用简洁的语言描述出来，比如“把加法换成乘法”，然后存起来。每次改进时，它会根据这些变化，帮助机器人更快找到正确的答案。这样一来，机器人学得更快，解决问题也更有效率。这就像你在学习中，记住那些关键的窍门，而不是所有细节，学习变得更轻松、更聪明。
Abstract
LLM-driven evolutionary systems have shown promise for automated science discovery, yet existing approaches such as AlphaEvolve rely on full-code histories that are context-inefficient and potentially provide weak evolutionary guidance. In this work, we first formalize the evolutionary agents as a general Expectation-Maximization framework, where the language model samples candidate programs (E-step) and the system updates the control context based on evaluation feedback (M-step). Under this view, constructing context via full-code snapshots constitutes a suboptimal M-step, as redundant implement details dilutes core algorithmic ideas, making it difficult to provide clear inspirations for evolution. To address this, we propose DeltaEvolve, a momentum-driven evolutionary framework that replaces full-code history with structured semantic delta capturing how and why modifications between successive nodes affect performance. As programs are often decomposable, semantic delta usually contains many effective components which are transferable and more informative to drive improvement. By organizing semantic delta through multi-level database and progressive disclosure mechanism, input tokens are further reduced. Empirical evaluations on tasks across diverse scientific domains show that our framework can discover better solution with less token consumption over full-code-based evolutionary agents.
References (20)
AlphaEvolve: A coding agent for scientific and algorithmic discovery
Alexander Novikov, Ngân V. ̃u, Marvin Eisenberger et al.
M+: Extending MemoryLLM with Scalable Long-Term Memory
Yu Wang, Dmitry Krotov, Yuanzhe Hu et al.
Real-Parameter Black-Box Optimization Benchmarking 2009: Noiseless Functions Definitions
N. Hansen, R. Ros, A. Auger
CodeEvolve: An open source evolutionary coding agent for algorithm discovery and optimization
Henrique S. Assumpção, Diego Ferreira, L. Campos et al.
AutoML-Zero: Evolving Machine Learning Algorithms From Scratch
Esteban Real, Chen Liang, David R. So et al.
DreamCoder: growing generalizable, interpretable knowledge with wake–sleep Bayesian program learning
Kevin Ellis, Catherine Wong, Maxwell Nye et al.
Compositional Generalization and Decomposition in Neural Program Synthesis
Kensen Shi, Joey Hong, M. Zaheer et al.
Faster sorting algorithms discovered using deep reinforcement learning
D. Mankowitz, Andrea Michi, A. Zhernov et al.
Mathematical discoveries from program search with large language models
B. Romera-Paredes, M. Barekatain, Alexander Novikov et al.
Solving olympiad geometry without human demonstrations
Trieu H. Trinh, Yuhuai Wu, Quoc V. Le et al.
Long Context Compression with Activation Beacon
Peitian Zhang, Zheng Liu, Shitao Xiao et al.
MEMORYLLM: Towards Self-Updatable Large Language Models
Yu Wang, Xiusi Chen, Jingbo Shang et al.
Efficient Prompt Compression with Evaluator Heads for Long-Context Transformer Inference
WeiZhi Fei, Xueyan Niu, Guoqing Xie et al.
Learning to Discover at Test Time
Mert Yuksekgonul, Daniel Koceja, Xinhao Li et al.
LIDDIA: Language-based Intelligent Drug Discovery Agent
Reza Averly, Frazier N. Baker, Xia Ning
A Systematic Survey of Automatic Prompt Optimization Techniques
Kiran Ramnath, Kang Zhou, Sheng Guan et al.
LLM-SRBench: A New Benchmark for Scientific Equation Discovery with Large Language Models
P. Shojaee, Ngoc-Hieu Nguyen, Kazem Meidani et al.
CodePDE: An Inference Framework for LLM-driven PDE Solver Generation
Shanda Li, Tanya Marwah, Junhong Shen et al.
Cited By (17)
Marginal Advantage Accumulation for Memory-Driven Agent Self-Evolution
AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization
EvoX: Meta-Evolution for Automated Discovery
The Agentic Researcher: A Practical Guide to AI-Assisted Research in Mathematics and Machine Learning
k-server-bench: Automating Potential Discovery for the k-Server Conjecture
Co-evolving Agent Architectures and Interpretable Reasoning for Automated Optimization
SMCEvolve: Principled Scientific Discovery via Sequential Monte Carlo Evolution
Evolutionary Multi-Task Optimization for LLM-Guided Program Discovery
BenchEvolver: Frontier Task Synthesis via Solution-Centric Evolution
What is Missing from AI Post-Training AI: An Empirical Analysis
NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?
Socratic agents for autonomous scientific discovery in high-dimensional physical systems
From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier
Automated Discovery Has No Universally Superior Harness
Teaching LLMs to Self-Evolve: Cultivating Core Meta-Skills with Reinforcement Learning
Budget-Aware LLM Discovery via Cost-Calibrated Frontier Utility
The Optimizer Is the Agent: Reasoning-Driven Search across Prompts, Programs, and ML Workflows