Efficient Layer-wise LLM Fine-tuning for Revision Intention Prediction

2025-09-30

AI Summary

Fine-grained text revision intent prediction poses two key challenges for large language models (LLMs): (i) direct instruction tuning struggles to capture subtle textual distinctions, and (ii) full-parameter fine-tuning relies heavily on scarce, labor-intensive annotated data. Method: We propose a hierarchical parameter-efficient fine-tuning (PEFT) framework featuring a novel gradient-norm-based dynamic critical layer selection mechanism. This method freezes redundant layers while selectively optimizing only those network layers most sensitive to revision identification. Contribution/Results: The approach significantly enhances discriminative capability under low-data regimes. Experiments demonstrate consistent superiority over existing hierarchical fine-tuning baselines across multi-class revision classification tasks. It achieves faster convergence, reduced GPU memory footprint, and improved generalization—without compromising accuracy—thereby offering an efficient, scalable solution for fine-grained revision intent modeling.

Technology Category

• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Large Language Models (LLMs) have shown extraordinary success across various text generation tasks; however, their potential for simple yet essential text classification remains underexplored, as LLM pre-training tends to emphasize generation over classification. While LLMs with instruction tuning can transform classification into a generation task, they often struggle to categorize nuanced texts. One such example is text revision, which involves nuanced edits between pairs of texts. Although simply fine-tuning LLMs for revision classification seems plausible, it requires a large amount of revision annotations, which are exceptionally expensive and scarce in the community. To address this issue, we introduce a plug-and-play layer-wise parameter-efficient fine-tuning (PEFT) framework, i.e., IR-Tuning, which fine-tunes a subset of important LLM layers that are dynamically selected based on their gradient norm distribution, while freezing those of redundant layers. Extensive experiments suggest that IR-Tuning surpasses several layer-wise PEFT baselines over diverse text revisions, while achieving fast convergence, low GPU memory consumption, and effectiveness on small revision corpora.

Problem

• Fine-tuning LLMs for nuanced text revision classification tasks
• Addressing expensive annotation requirements for text revision datasets
• Optimizing parameter-efficient fine-tuning for limited computational resources

Innovation

• Layer-wise fine-tuning of important LLM layers
• Dynamic layer selection via gradient norm distribution
• Parameter-efficient framework for small revision corpora