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

Zhexiong Liu, Diane LitmanEMNLP 2025

TL;DR

This paper introduces IR-Tuning, a novel layer-wise parameter-efficient fine-tuning (PEFT) method for Large Language Models (LLMs) that dynamically selects and fine-tunes only the most important LLM layers for a given task. This approach significantly reduces computational costs and improves efficiency, especially for tasks like revision intention prediction where labeled data is scarce.

What Problem Does This Paper Address?

Large Language Models (LLMs) have shown remarkable capabilities in text generation, but their application to text classification tasks, particularly those involving nuanced understanding like predicting revision intentions, remains less explored. While instruction tuning can adapt LLMs for classification, they often struggle with complex, subtle texts. Fine-tuning LLMs for such tasks is computationally expensive and requires large amounts of labeled data, which are difficult and costly to obtain for specific domains like text revision. Existing parameter-efficient fine-tuning (PEFT) methods, while reducing computational load, often fine-tune entire layers or a fixed subset, potentially overlooking the dynamic nature of layer importance during training. This paper addresses the challenge of efficiently fine-tuning LLMs for revision intention prediction on small datasets by intelligently selecting and updating only the most relevant layers.

Methods

The core innovation is Importance Redundancy Tuning (IR-Tuning), a plug-and-play PEFT framework. IR-Tuning works by:

  1. Layer-wise Importance Score Calculation: It dynamically computes an "importance score" for each LLM layer, based on the norm of its gradients during fine-tuning. Layers with higher gradient norms are considered more important as they contribute more significantly to the loss function's optimization.
  2. Dynamic Layer Selection: Instead of a fixed selection, IR-Tuning uses an algorithm to dynamically select a subset of "important" layers and freezes the "redundant" ones. This selection is based on a distribution divergence problem, aiming to find a threshold that separates layers with significantly different importance scores. This process is repeated periodically during fine-tuning to adapt to changing layer importance.
  3. Parameter-Efficient Fine-tuning: The selected important layers are then fine-tuned using LoRA (Low-Rank Adaptation), a popular PEFT technique, while the redundant layers remain frozen.

This approach contrasts with prior methods like IST (Importance-aware Sparse Tuning) which fix the number of important layers, potentially leading to overlooking crucial layers or including unnecessary ones.

Datasets

The authors use two primary datasets:

  • ArgRevision Corpus: A newly collected corpus of essays written by elementary and middle school students. This corpus focuses on argumentation and is used to study revisions in less skilled writers.
  • ITERATER Corpus: A publicly available corpus that annotates text revisions from Wikipedia, ArXiv, and news articles, representing revisions by experienced authors, editors, and researchers.

Key Insights and Ideas

  • Dynamic Layer Importance: LLM layers do not have static importance for a given task. Their relevance can change during the fine-tuning process, and IR-Tuning's dynamic selection captures this.
  • Efficiency Gains: By selectively fine-tuning only a subset of layers, IR-Tuning significantly reduces computational costs (GPU memory) and speeds up convergence compared to full fine-tuning and other PEFT baselines.
  • Effectiveness on Small Corpora: The method demonstrates strong performance even with limited annotated data, making it practical for tasks where data annotation is a bottleneck.
  • Instruction Tuning Nuance: While instruction tuning can be beneficial, its effectiveness for revision intention prediction is not universal and depends on the data's complexity and the LLM's capabilities. For simpler revisions, it might not be necessary or could even be detrimental.

Key Results

  • Performance: IR-Tuning consistently outperforms several layer-wise PEFT baselines (LISA-Baseline, IST-Baseline) and even matches or surpasses full fine-tuning on various metrics (Precision, Recall, F1-Score, AUPRC) across both the ArgRevision and ITERATER corpora.
  • Efficiency: IR-Tuning shows significantly faster convergence and lower GPU memory allocation compared to full fine-tuning and other PEFT methods. For example, it uses the smallest memory footprint among the tested methods.
  • Generalizability: The method shows effectiveness across different LLMs (Mistral-7B, Llama3.1-8B, Deepseek-R1-8B) and varying model sizes.

Limitations and Future Work

  • Importance Metric: The current importance scoring relies on gradient norms. The authors suggest exploring more data-driven and task-specific metrics, such as evaluation loss, for scoring layer importance.
  • Layer Splitting: The current implementation uses a single split for selecting important layers. Future work could explore automated methods for hierarchical splitting to achieve more fine-grained layer selection.
  • Model Size: The study was constrained to relatively lightweight LLMs due to computational limitations. Evaluating IR-Tuning on larger LLMs (e.g., Llama3.3-70B) is a promising direction.
  • PEFT Techniques: The research focused on LoRA. Exploring IR-Tuning with other PEFT techniques (e.g., prompt-tuning) and diverse LLMs could lead to more robust strategies.
  • Generalization: While promising for revision intention prediction, the authors plan to evaluate IR-Tuning on more revision tasks to generalize their findings.

Summary updated

BibTeX
@inproceedings{liu2025efficient,
  author = {Zhexiong Liu and Diane Litman},
  title = {{Efficient Layer-wise LLM Fine-tuning for Revision Intention Prediction}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure