AlphaGo Moment for Model Architecture Discovery

arxiv:2507.18074

Original source · Paper JSON

Agents, open-endedness, AGI · 2025-07-24

Cites (3) · Cited by (2)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• AlphaGo Moment for Model Architecture Discovery
• Abstract
• 1 Introduction
• 2 Related Work
• AI For AI Research
• Efficient Architecture
• 3 Methodology
• 3.1 The Fitness Function
• 3.2 Researcher: Propose New Architecture
• Seed Selection
• Model Design
• Program Implementation
• Novelty and Sanity Check
• 3.3 Engineer: Train and Evaluate Models
• Real Code Environment
• LLM-as-Judge Scoring
• 3.4 Analyzer: Mine Experimental Insights
• Cognition Base
• Contextual Analysis
• 3.5 Exploration-then-Verification Strategy
• 4 Experiments
• 4.1 Experimental Setup
• Fitness Function
• Parallel Search Architecture and Policy
• Efficiency-Oriented Framework Settings
• 4.2 Main results
• 5 Analysis
• 5.1 Effectiveness of LLM-Driven Architecture Search
• 5.2 Architectural Design Patterns
• Model Complexity Stability
• Architectural Component Preferences
• 5.3 Where Do Good Designs Come From?
• Discussion and Future Work
• Multi-Architecture Initialization
• Component-wise Analysis
• Engineering Optimization
• Appendix A Experimental Setup
• A.1 Pipeline Configuration
• Framework Overview
• Multi-Model Integration
• Data Management and Retrieval
• A.2 Experimental Configuration
• Progressive Evaluation Strategy
• Model Architecture
• Training Protocol
• Data Configuration
• Evaluation Protocol
• Appendix B Prompts
• B.1 Planner
• B.2 Checker
• B.3 Debugger
• B.4 Analyser
• B.5 Cognition
• Instructions for reporting errors