Intention-Adaptive LLM Fine-Tuning for Text Revision Generation

TL;DR

This work addresses revision generation that must reflect writer intentions, a problem made difficult by entangled multi-intent goals and scarce labeled data. It introduces Intention-Tuning, an intention-adaptive layer-wise PEFT framework that first learns intention representations via an intention-prediction task and then transfers them to revision generation by selecting and sharing a subset of LLM layers through LoRA adapters, trained sequentially to reduce computation. Layer probing uses gradient norms to identify important layers for intention prediction and updating-frequency to finalize layers for revision generation, enabling cross-task alignment measured by a layer-alignment ratio $r$. Across three LLMs and two revision corpora (ITERATER and ArgRevision), Intention-Tuning achieves strong performance and faster convergence with lower GPU memory than full fine-tuning or other PEFT baselines, while qualitatively reflecting writer intentions and offering a practical path for low-resource, intention-aware text revision. The framework demonstrates the potential of integrating intention-aware signals into large-model fine-tuning, with implications for education tech and other generation tasks, while highlighting areas for future work such as more robust span-level edits and larger-scale models.

Abstract

Large Language Models (LLMs) have achieved impressive capabilities in various context-based text generation tasks, such as summarization and reasoning; however, their applications in intention-based generation tasks remain underexplored. One such example is revision generation, which requires the generated text to explicitly reflect the writer's actual intentions. Identifying intentions and generating desirable revisions are challenging due to their complex and diverse nature. Although prior work has employed LLMs to generate revisions with few-shot learning, they struggle with handling entangled multi-intent scenarios. While fine-tuning LLMs using intention-based instructions appears promising, it demands large amounts of annotated data, which is expensive and scarce in the revision community. To address these challenges, we propose Intention-Tuning, an intention-adaptive layer-wise LLM fine-tuning framework that dynamically selects a subset of LLM layers to learn the intentions and subsequently transfers their representations to revision generation. Experimental results suggest that Intention-Tuning is effective and efficient on small revision corpora, outperforming several PEFT baselines.

Figures (11)

• Figure 1: Revision examples: the original text is revised differently based on a single (meaning change) intention and multiple (clarity and fluency) intentions. The examples are from the ITERATER corpus du-etal-2022-understanding-iterative.
• Figure 2: The Intention-Tuning framework. In the intention prediction task (predictor), the important LLM layers are fine-tuned while the redundant LLM layers are frozen. Upon completion, (I) its layer-wise importance frequency is used to finalize the important layers for the revision generation task (generator); (II) the learned layer-wise representations (LoRA weights) in the predictor are shared with the generator.
• Figure 3: The layer-wise importance score alignment between the intention prediction (green boxes) and the revision generation (brown boxes) tasks while fine-tuning LLMs on the ITERATER-sent. All PEFT uses LoRA. The dark colors indicate high scores (important layers) while the light colors indicate low scores (redundant layers).
• Figure 4: Llama3.1-8B fine-tuning loss on the training set of the datasets. All PEFT uses LoRA.
• Figure 5: The example of the human and generated revisions by Llama3.1-8B with Intention-Tuning on the ITERATER-sent dataset. The yellow and green colors denote additions, and the purple denotes deletions.