LEMUR 2: Unlocking Neural Network Diversity for AI

Overview

LEMUR 2 addresses the limitations of existing Neural Architecture Search (NAS) benchmarks, which are often restricted to narrow, task-specific search spaces and lack real-world deployment data. By unifying generative, evaluative, and deployment pipelines, LEMUR 2 provides a comprehensive resource for studying neural network diversity, transferability, and performance across heterogeneous hardware.

Generative Methodologies

The framework achieves architectural diversity by employing a multi-pronged generation strategy. This includes AST-based code mutation, genetic algorithms, reinforcement learning, and fractal-inspired construction. A key innovation is the NN-RAG (Retrieval-Augmented Generation) system, which mines public repositories to extract and validate over 900 reusable PyTorch modules, allowing the framework to synthesize new architectures from proven, real-world motifs.

Deployment and Evaluation

Unlike traditional benchmarks that focus solely on theoretical accuracy, LEMUR 2 incorporates deployment-aware metadata. The framework includes the NN-Lite and NN-VR pipelines, which automatically convert and benchmark models on Android mobile devices and Unity-based virtual reality platforms. This provides researchers with empirical latency measurements, bridging the gap between abstract architectural design and practical, resource-constrained implementation.

LEMUR 2 serves as a foundational dataset for the emerging paradigm of LLM-driven AutoML. By linking diverse architectural origins with cross-platform empirical validation, it enables researchers to identify structural regularities and generalizable design principles. This data-driven approach facilitates more efficient model discovery and provides the necessary metadata to optimize neural networks for specific hardware targets, advancing the state of reproducible AI design.