LEMUR 2: Unlocking Neural Network Diversity for AI

Automated Machine Learning (AutoML) frameworks increasingly leverage Large Language Models (LLMs) for tasks such as hyperparameter optimization and neural architecture code generation. However, current LLM-based approaches focus on generative outputs and evaluate them by training the produced artifacts. Whether LLMs can learn to reason about neural network performance across datasets remains underexplored. We present a classification task integrated into the NNGPT framework, in which a fine-tuned LLM predicts which of two image classification datasets a given neural network architecture achieves higher accuracy on. The task is built on the LEMUR dataset, which provides standardized PyTorch implementations with reproducible performance metrics. Three prompt configurations of increasing difficulty are evaluated: a normalized-accuracy baseline (trivially reaching 100%), a metadata-enriched prompt replacing accuracies with dataset properties, and a code-only prompt presenting only architecture source code and dataset names. Using DeepSeek-Coder-7B-Instruct fine-tuned with LoRA, the code-only prompt reaches 80% peak accuracy over 15 epochs, while the metadata prompt peaks at 70%. Perdataset analysis reveals complementary strengths: metadata excels for datasets with distinctive properties (CelebAGender at 90.9%) but degrades for overlapping characteristics, whereas the code-only prompt shows more balanced performance. A comparison with DeepSeek-Coder1.3B confirms that model capacity affects this form of architectural reasoning. The results establish that LLMs can be fine-tuned to predict cross-dataset suitability from neural network code, suggesting that architecture source code contains richer discriminative signal than dataset metadata alone.

Large language models (LLMs) show strong potential for neural architecture generation, yet existing approaches produce complete model implementations from scratch -- computationally expensive and yielding verbose code. We propose Delta-Code Generation, where fine-tuned LLMs generate compact unified diffs (deltas) to refine baseline architectures rather than synthesizing entire models. Our pipeline iteratively fine-tunes the LLM via LoRA on curated architectures from the LEMUR dataset, with MinHash-Jaccard novelty filtering for structural diversity. We evaluate three 7B-class LLMs -- DeepSeek-Coder-7B, Qwen2.5-Coder-7B, and Mistral-7B -- across six datasets (CIFAR-10, CIFAR-100, MNIST, SVHN, ImageNette, CelebA) using a 22-cycle protocol (1,100 candidates per LLM). All three substantially surpass the full-generation baseline (50.6% valid rate, 42.3% mean first-epoch accuracy): DeepSeek-Coder reaches 75.3% valid rate and 65.8% mean accuracy; Qwen2.5-Coder 72.1%/64.6%; Mistral 66.6%/66.1%. On CIFAR-10, best first-epoch accuracies reach 85.5% (Mistral), 85.2% (DeepSeek), 80.6% (Qwen) -- well above 63.98% full generation and 71.5% for the concurrent approach of Gu et al. Output lengths are 30-50 lines versus 200+ for full generation (75-85% reduction). A 50-epoch study confirms the 1-epoch proxy preserves rankings (Mistral: Spearman $ρ$ = 0.926). Delta-based generation is a token-efficient, multi-domain, LLM-agnostic alternative to full-model synthesis for LLM-driven NAS.

Introduce a MinHash-based similarity scheduling framework that constructs a progressive curriculum over neural architecture code for LLM-based neural architecture search (NAS). Using 128-permutation MinHash signatures over normalised 7-gram source code shingles, we partition the reference pool into similarity bands and present them in increasing architectural heterogeneity, with the best LoRA adapter from each stage merged cumulatively into the backbone. We evaluate the framework on OlympicCoder-7B within the LEMUR benchmark on CIFAR-10 image classification, generating N =15 candidate architectures per epoch across six progressive fine-tuning steps. The curriculum achieves 60% peak success rate at the high-similarity level without post-processing repair. A 2*2 ablation at the most diverse level curriculum versus base model, with versus without partial interface repair reveals that without repair the base model (47% peak SR) substantially outperforms the curriculum model (7% SR), while adding partial repair brings both to 53% SR. This pattern is consistent with merge-level weight drift progressively erasing evaluator-interface priors, and suggests that interface repair and curriculum scheduling target distinct failure modes. We further report a cross-dataset transfer observation on SVHN, where direct base-model generation without curriculum warmup yields 27% peak SR at substantially lower accuracy (60.5%) than the CIFAR-10 equivalent, consistent with the increased synthesis difficulty of the unq-family anchor architecture.