Stop coding AI: Use Runtime Topological Self-Assembly (UC, DeepMind)

2026-02-25 · Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

Two new research papers, "Alpha Evolve" from Google DeepMind and "Open Sage" from a consortium including UC Berkeley and UCLA, propose a paradigm shift in AI development by using Large Language Models (LLMs) as optimizers for discrete, non-differentiable symbolic graphs rather than end-to-end solvers. Alpha Evolve focuses on micro-architectural optimization, using LLMs as genetic operators to mutate Abstract Syntax Trees (ASTs) of solver code, enabling the discovery of novel mathematical optimization algorithms, such as a dynamic annealing MASA formula. Open Sage addresses macro-architectural optimization by dynamically generating Topological Execution Graphs (TEGs) at runtime, allowing LLMs to construct and manage multi-agent systems, tools, and memory hierarchies autonomously. This approach moves beyond human-coded, static agent development kits like LangChain, demonstrating superior performance and cost-efficiency in benchmarks like Cyber Gym and Terminal Bench, often outperforming GPT-5 with lower operational costs by optimizing LLM collaboration.

Key takeaway

For AI Engineers and Research Scientists designing complex multi-agent systems, you should explore shifting from continuous, human-coded optimization to LLM-driven discrete graph optimization. This approach, exemplified by Alpha Evolve and Open Sage, allows for the discovery of novel algorithms and dynamic system architectures, potentially leading to more robust, adaptive, and cost-efficient AI solutions than traditional methods. Consider how to integrate LLMs as optimizers for your system's underlying logic and topological structure.

Key insights

LLMs can optimize discrete symbolic graphs to discover new algorithms and dynamically architect multi-agent systems.

Principles

• AI logic exists in discrete, non-differentiable spaces.
• LLMs can act as genetic operators for code graphs.
• Dynamic graph generation enhances agent adaptability.

Method

Alpha Evolve uses LLMs to mutate Abstract Syntax Trees (ASTs) for algorithmic discovery. Open Sage employs LLMs to dynamically generate Topological Execution Graphs (TEGs) for multi-agent system architecture and tool orchestration.

In practice

• Explore LLMs for optimizing discrete system logic.
• Implement dynamic TEG generation for adaptive agents.
• Combine LLMs for cost-effective performance optimization.

Topics

• LLM Optimization
• Discrete Symbolic Graphs
• Multi-Agent Systems
• Alpha Evolve
• Open Sage

Best for: AI Engineer, AI Scientist, Research Scientist, AI Researcher, Machine Learning Engineer, AI Architect

Related on AIssential

• Papers You Should Know About — AI research is advancing LLM efficiency and agent autonomy through novel architectures, self-improvement mechanisms, and real-world applications.
• DW-Bench: Benchmarking LLMs on Data Warehouse Graph Topology Reasoning — LLMs excel at orchestrating graph algorithms via tools but struggle with compositional multi-hop graph reasoning.
• LLM-as-Code Agentic Programming for Agent Harness — Agentic Programming reassigns control flow from LLMs to the program, treating LLMs as "LLM-as-Code" for reliable, context-efficient reasoning.
• Exploring LLM-based Verilog Code Generation with Data-Efficient Fine-Tuning and Testbench Automation — Multi-agent LLMs can automate testbench generation, creating high-quality data for efficient Verilog code generation.
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — LLM agent memory should be actively reconstructed and dynamically adapted to reasoning context, not statically retrieved.
• Teaching Machines to Code: A Social Scientist’s Guide to LLM-Based Information Extraction — Multi-agent LLM pipelines excel at complex information extraction by mimicking human collaborative coding processes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.