LLM-Driven Neural Network Generation with Same-Family Architecture Guidance: Disentangling Transfer and Adaptation

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

A new protocol for Large Language Model (LLM)-driven neural network generation demonstrates that guiding LLMs with stronger, same-family source models significantly improves weak target models. This method, which uses a four-arm candidate generation protocol including non-source controls and source-conditioned candidates, reports validity and accuracy separately. On CIFAR-10, the best source-guided candidate achieved 0.5049 accuracy, a +0.2651 advantage over the best non-source candidate (0.2398), improving a target from 0.1254. For SVHN AlexNet with DeepSeek-Coder-6.7B, source-guided transfer reached 0.7880, a +0.5626 advantage over 0.2254. The study identifies two regimes: \"recipe-transfer,\" where direct source recipe copying is effective, and \"recipe-adaptation,\" where the LLM modifies the source recipe for the weak target. AlexNet and alt_nn1 architectures show the most consistent positive results.

Key takeaway

For machine learning engineers seeking to improve weak neural networks or explore neural architecture search, you should integrate source-guided LLM candidate generation into your workflow. By utilizing stronger, same-family models to condition LLM outputs, particularly for hyperparameter and transform transfer, you can achieve significant accuracy gains. Always evaluate source-guided and non-source candidates concurrently, accepting the best valid option only if it surpasses your original target's performance.

Key insights

LLMs improve weak neural networks by adapting or transferring recipes from stronger, same-family source models.

Principles

• Source-guided LLM generation outperforms non-source baselines.
• LLMs can adapt or transfer training recipes for performance gains.
• Report validity and accuracy separately for code generation tasks.

Method

Select weak target and strong same-family source, generate non-source and source-guided candidates, then train and evaluate under equal budgets.

In practice

• Employ hp_transfer for recipe adaptation.
• Evaluate source-guided and target-only arms.
• Focus on AlexNet or alt_nn1 families.

Topics

• LLM Code Generation
• Neural Architecture Search
• Hyperparameter Optimization
• Recipe Transfer
• Recipe Adaptation
• AlexNet Architecture

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• ZeroCoder: Can LLMs Improve Code Generation Without Ground-Truth Supervision? — Co-evolving code and test generation with self-generated execution feedback significantly improves LLM performance without ground-truth supervision.
• LLM4RTL: Tool-Assisted LLM for RTL Generation — Tool-assisted LLMs and refined datasets significantly enhance RTL code generation performance.
• AWS Generative AI Model Agility Solution: A comprehensive guide to migrating LLMs for generative AI production — A structured framework for LLM migration and optimization ensures continuous performance improvement and operational stability.
• Bridging Functional Correctness and Runtime Efficiency Gaps in LLM-Based Code Translation — LLM-translated code often lacks runtime efficiency; SwiftTrans improves both correctness and speed through multi-perspective exploration and difference-aware selection.
• GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models — GIFT enhances instruction-tuned models by guiding base model adapter training with instruction model confidence signals.
• No Resource, No Benchmarks, No Problem? Evaluating and Improving LLMs for Code Generation in No-Resource Languages — Instruction transferring on pre-trained base models effectively specializes LLMs for no-resource languages without costly instruction fine-tuning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.