RE-MMNAS: LLM Reasoning-Guided Evolutionary Architecture Search for Multi-Modal Feature Fusion

Multi-modal neural architecture search faces two fundamental challenges: semantic misalignment across heterogeneous feature scales and modality imbalance during fusion. While traditional evolutionary methods rely on blind heuristic search, recent LLM-assisted NAS approaches treat language models as static code generators, failing to exploit their capacity for reflective reasoning. We propose RE-MMNAS, a reasoning-guided evolutionary framework that embeds LLMs as active optimization agents throughout the search loop. By performing contrastive analysis on historical architecture populations, RE-MMNAS drives two novel operators. The modality-aware crossover identifies and corrects modality imbalance via per-modality F1 scores, while the context-aware mutation is conditioned on iteratively refined evolutionary experience encoded in dynamic chain-of-thought prompts. Evaluated on MM-IMDB and EgoGesture, RE-MMNAS achieves 64.22% weighted-F1 and 95.44% accuracy, outperforming state-of-the-art baselines while discovering architectures with up to 35% fewer parameters. Ablation studies confirm that the synergy between the proposed operators and dynamic prompting is essential—neither component alone surpasses the non-LLM baseline.