HypoExplore: A Hypothesis-Driven Agent Framework for Neural Architecture Discovery

This article introduces HypoExplore, an agent framework that formalizes neural architecture discovery as hypothesis-driven scientific inquiry. Its core idea is to simulate the research process of human scientists, using key components such as evolutionary branching, hypothesis memory bank, and confidence tracking. On the CIFAR-10 dataset, it achieves a performance leap from the initial architecture (18.91% accuracy) to the optimal architecture (94.11% accuracy), and has generalization capabilities across datasets (e.g., CIFAR-100, Tiny-ImageNet) and domains (e.g., MedMNIST medical imaging).

Neural architecture design has evolved from manual design to automated search stages: early architectures like AlexNet and ResNet relied on researchers' intuition; later Neural Architecture Search (NAS) methods attempted automation but faced issues of high computational cost and lack of interpretability. The rise of Large Language Models (LLMs) brings new possibilities for architecture discovery. HypoExplore redefines architecture discovery as a hypothesis-driven scientific inquiry process, enabling knowledge accumulation and understanding.

Core Idea of the Framework

Simulate the research process of human scientists: propose hypotheses → design experiments → validate hypotheses → iterate improvements.

Key Components

1. High-level Research Directions: Retain human experts' experience; after specifying directions, the system automatically fills in details;
2. Evolutionary Branching Mechanism: Form a traceable architecture tree based on improvements to parent architectures;
3. LLM-driven Hypothesis Generation: Select parent hypotheses based on research status and propose modification plans;
4. Dual-strategy Guidance: Balance exploitation (optimize existing successes) and exploration (address high uncertainty);
5. Trajectory Tree: Record architecture lineage, supporting traceability, knowledge inheritance, and failure analysis;
6. Hypothesis Memory Bank: Track hypothesis confidence, update through experiments, and guide subsequent selections;
7. Multi-perspective Feedback Agent: Analyze experimental results from perspectives like performance, efficiency, stability, and comparison to update confidence.

Performance Improvement on CIFAR-10

The initial random architecture has an accuracy of 18.91%, and the optimal architecture after evolution reaches 94.11%, an improvement of over 75 percentage points.

Generalization Capability Validation

• Cross-dataset: Performs well on CIFAR-100 (100-class classification) and Tiny-ImageNet (large-scale image recognition);
• Cross-domain: Achieves state-of-the-art performance on the MedMNIST medical imaging dataset, proving its versatility.

Predictiveness of Hypothesis Confidence

With experimental accumulation, the correlation between confidence and actual performance increases, making it a reliable predictor of architecture quality.

Knowledge Transfer

Learned design principles (e.g., depthwise separable convolution improves efficiency) can spread across evolutionary lineages, building a true understanding of the design space.

Comparison with Traditional NAS

• Interpretability: Trajectory tree and memory bank provide complete decision history;
• Knowledge Accumulation: Accumulate knowledge across experiments, making the system gradually 'smarter';
• Sample Efficiency: Find high-performance architectures with fewer experiments;
• Human-machine Collaboration: Humans inject prior knowledge, and the system automatically handles details.

Limitations

• Computational Cost: Architecture training and evaluation still require significant resources;
• LLM Dependence: Insufficient domain knowledge may lead to unreasonable hypotheses;
• Exploration Depth: Scalability to large-scale models (e.g., Transformer variants) needs verification;
• Theoretical Understanding: The theoretical mechanism behind the predictiveness of hypothesis confidence requires in-depth analysis.

Future Directions

• Explore more efficient proxy evaluation methods;
• Combine domain expert models to compensate for LLM shortcomings;
• Extend to large-scale model exploration.

Implications for AI Research

• Treat architecture discovery as scientific discovery rather than mere optimization;
• LLM agents can serve as scientific research assistants;
• The shift from 'finding good architectures' to 'understanding why architectures are good' is key to AI intelligence.