NeuroGenesis: LLM-Driven Architecture Search
Reasoning-guided Neural Architecture Search (NAS) using LLMs for systematic hypothesis formation across audio, video, images, and time-series (IEEE QRS 2026).
NeuroGenesis: LLM-Driven Neural Architecture Search
Stack: Python, PyTorch, LLMs, AutoML, Neural Architecture Search (NAS)
Venue: IEEE International Conference on Software Quality, Reliability and Security (QRS 2026)
Authors: Muhammad Anas Raza, Niful Islam, Mohammad Wardat
Overview
Traditional Neural Architecture Search (NAS) approaches rely heavily on brute-force random search, reinforcement learning, or genetic algorithms, which require massive compute clusters and evaluate hundreds of degenerate candidates without semantic reasoning.
NeuroGenesis introduces a reasoning-guided NAS paradigm that replaces blind trial-and-error with systematic hypothesis formation, design justification, and iterative architectural refinement using LLMs.
Key Innovations
- Multimodal Search Space: Evaluated across 4 distinct data modalities: audio waveforms, raw video, spatial images, and multivariate time series (MTS).
- 20 Candidate Backbones: Supports modern backbones across classification, regression, and sequence forecasting.
- Rapid Prototyping: Uses calibrated synthetic and toy datasets to explore model configurations at minimal GPU cost before transferring to full-scale datasets.
- Empirical Validation: Benchmarked across 8 large real-world user datasets and 22 specialized benchmarks, matching or exceeding manually hand-crafted architectures and established AutoML tools (AutoKeras, AutoTS) while slashing total search duration.