Similarity-Guided Curriculum Fine-Tuning of LLMs for Neural Architecture Synthesis

2026-07-13 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new MinHash-based similarity scheduling framework is introduced for fine-tuning Large Language Models (LLMs) in neural architecture synthesis (NAS). This framework constructs a progressive curriculum using 128-permutation MinHash signatures over normalized 7-gram source code shingles to partition neural architecture code into similarity bands. These bands are presented in increasing architectural heterogeneity, with the best LoRA adapter from each stage cumulatively merged into the backbone. Evaluated on OlympicCoder-7B within the LEMUR benchmark for CIFAR-10 image classification, the curriculum generates N=15 candidate architectures per epoch across six fine-tuning steps, achieving a 60% peak success rate at the high-similarity level without post-processing repair. Ablation studies reveal that at the most diverse level, the base model (47% peak SR) outperforms the curriculum model (7% SR) without repair, but both reach 53% SR with partial interface repair, indicating distinct failure modes. Cross-dataset transfer to SVHN shows a 27% peak SR at 60.5% accuracy for the base model without curriculum warmup.

Key takeaway

For Machine Learning Engineers developing LLM-based Neural Architecture Search systems, you should integrate similarity-guided curriculum fine-tuning to enhance synthesis success rates, particularly at higher architectural diversity. Implement partial interface repair alongside curriculum scheduling, as these techniques address distinct failure modes and collectively improve performance, mitigating issues like merge-level weight drift. Consider the increased synthesis difficulty for anchor architectures when transferring across datasets.

Key insights

MinHash-guided curriculum fine-tuning improves LLM-based neural architecture synthesis by progressively increasing architectural heterogeneity.

Principles

• Curriculum scheduling and interface repair address distinct failure modes in NAS.
• Merge-level weight drift can erase evaluator-interface priors.

Method

A MinHash-based similarity scheduling framework partitions neural architecture code into increasing heterogeneity bands, cumulatively merging the best LoRA adapter from each stage into the backbone.

In practice

• Apply 128-permutation MinHash for code similarity.
• Use LoRA adapter merging for progressive fine-tuning.
• Implement partial interface repair to mitigate weight drift.

Topics

• Neural Architecture Search
• Large Language Models
• Curriculum Learning
• MinHash Signatures
• LoRA Adapters
• Fine-tuning

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

Related on AIssential

• Convergence Theory for Iterative LLM-Based Neural Architecture Search: A Parametric Cross-Entropy Framework with Closed-Form Proxy Reliability — Formal convergence theory validates iterative LLM-NAS, ensuring quality improvement and providing reliability diagnostics.
• Enough is as good as a feast: A Comprehensive Analysis of How Reinforcement Learning Mitigates Task Conflicts in LLMs — RL training significantly improves LLM model merging by reducing task conflicts and performance degradation.
• Multi-step Large Language Model for Fine-Grained Feedback in Stepwise Linear Equation Solutions — Multi-step LLMs and hybrid architectures significantly improve fine-grained error classification in algebraic problem solving for educational feedback.
• Granite 4.1 LLMs: How They’re Built — Granite 4.1 LLMs achieve strong performance through rigorous multi-stage data curation and reinforcement learning, even with smaller dense architectures.
• Multilingual Fine-Tuning via Localized Gradient Conflict Resolution — Localized gradient conflict resolution via parameter buckets significantly improves multilingual LLM fine-tuning efficiency and performance.
• Goal-Conditioned Supervised Learning for LLM Fine-Tuning — GCSL fine-tunes LLMs offline by treating graded feedback as explicit, beyond-threshold goals, leveraging natural language for improved performance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.