💬 NLP

Intention-Adaptive LLM Fine-Tuning for Text Revision Generation

This paper introduces Intention-Tuning, an efficient layer-wise fine-tuning framework that dynamically selects specific LLM layers to learn writer intentions and transfer these representations to text revision generation, effectively addressing the challenges of multi-intent scenarios with limited annotated data.

Original authors: Zhexiong Liu, Diane Litman

Published 2026-03-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhexiong Liu, Diane Litman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Mind-Reading" Robot

Imagine you have a brilliant robot assistant (a Large Language Model, or LLM) that can write stories, summarize news, and answer questions. It's great at following instructions like "Write a poem about a cat."

But now, imagine you ask the robot to edit a paragraph you wrote. You don't just want it to change words randomly; you want it to fix a specific problem. Maybe you want to make it clearer, or maybe you want to change the meaning entirely to sound more aggressive.

The problem is that robots are bad at "mind-reading." If you say, "Fix this," the robot might guess you want it to be shorter, when you actually wanted it to be more polite. In the world of writing, these "fixes" are called Intentions.

Current robots struggle when you give them a messy, complicated request like, "Make this clearer, but also change the meaning to be more positive, and fix the grammar." They get tangled up and often just make small, safe changes that don't really help.

The Old Way: The "Brute Force" Approach

To teach a robot to be a better editor, researchers usually try two things:

Show it examples: "Here is a bad sentence, here is the good version." (This is like showing a student a textbook).
Fine-tune the whole brain: They try to retrain the entire robot brain to understand editing.

The Catch:

Examples are rare: Finding thousands of examples where humans explain exactly why they changed a sentence is hard and expensive. It's like trying to find a library where every book has a sticky note explaining the author's thought process.
Retraining the whole brain is heavy: Retraining a massive robot brain requires huge amounts of computer power (like trying to upgrade a supercomputer's entire operating system just to learn a new recipe).

The New Solution: "Intention-Tuning"

The authors of this paper, Zhexiong Liu and Diane Litman, came up with a clever shortcut called Intention-Tuning.

Think of the robot's brain not as a single block, but as a multi-story skyscraper with many floors (layers).

The Old Way: They tried to renovate every single floor of the skyscraper to learn how to edit. This is expensive and slow.
The New Way: They realized that only specific floors are actually good at understanding "intentions" (the why behind a change), while other floors are just good at general grammar or vocabulary.

How Intention-Tuning Works (The 3-Step Dance)

Step 1: The Detective Phase (Intention Prediction)
First, they take the robot and ask it to play detective. They show it a "Before" and "After" sentence and ask, "What was the writer trying to do here? Were they trying to be clearer? Or change the meaning?"

The Magic: As the robot tries to solve this, the researchers watch which "floors" of the skyscraper light up the most. These are the Important Floors. The other floors stay dark (frozen).
Analogy: It's like watching a construction crew. You notice that only the electricians on the 3rd and 10th floors are working hard to fix the wiring. The plumbers on the 5th floor aren't needed for this job.

Step 2: The Map Phase (Transferring the Knowledge)
Once the robot figures out which floors are the "Important Floors" for understanding intentions, the researchers take a snapshot of that knowledge. They essentially say, "Okay, we know that floors 3 and 10 are the experts on 'Intentions'."

Step 3: The Editor Phase (Revision Generation)
Now, they ask the robot to actually do the editing. But here's the trick: They only let the robot use those specific "Important Floors" (3 and 10) to learn how to write the new text. They freeze the other floors so they don't get confused.

The Result: The robot uses the "Intentions" it learned in Step 1 to guide the "Writing" in Step 3, but it does it using a tiny, efficient slice of its brain.

Why This is a Game-Changer

It's a "Small Data" Hero: Because the robot only needs to learn on a few specific floors, it doesn't need thousands of examples. It can learn effectively from a small library of examples. It's like a student who only needs to study the specific chapters relevant to the test, rather than reading the whole encyclopedia.
It's Super Efficient: Since they aren't retraining the whole brain, it uses much less computer power and memory. It's like upgrading a car's engine by only swapping out the spark plugs and fuel injectors, rather than rebuilding the whole car.
It Handles "Messy" Requests: The paper shows that this method is great at handling Multi-Intent scenarios (e.g., "Make it clearer AND change the meaning"). The robot can juggle multiple goals because it's using the specific "floors" designed to understand complex intentions.

The Bottom Line

The authors built a system that teaches a robot to be a better editor by:

First teaching it to understand the "Why" (the intention) using a small, focused part of its brain.
Then, using that same focused part of the brain to do the "How" (the actual editing).

It's like hiring a specialist consultant who knows exactly which parts of a building need fixing, rather than hiring a crew to renovate the whole thing from the ground up. The result is a smarter, faster, and more efficient robot editor that works well even when you don't have a massive amount of training data.

1. Problem Statement

Text revision is a critical writing process driven by specific, often complex, writer intentions (e.g., improving clarity, changing meaning, fixing fluency). While Large Language Models (LLMs) excel at context-based generation (summarization, reasoning), their application to intention-based revision remains under-explored.

Challenges:
- Entangled Multi-Intent Scenarios: Revisions often involve multiple simultaneous intentions (e.g., fixing grammar while changing meaning), which existing models struggle to disentangle.
- Data Scarcity: High-quality, human-annotated revision corpora are small and expensive to produce. Standard fine-tuning or Parameter-Efficient Fine-Tuning (PEFT) methods typically require large-scale data, making them unsuitable for this domain.
- Layer Utilization: Existing PEFT methods often apply identical strategies across all layers or use static heuristics, failing to leverage the fact that different layers contribute differently to specific tasks (intention prediction vs. text generation).

2. Methodology: Intention-Tuning

The authors propose Intention-Tuning, a novel intention-adaptive layer-wise PEFT framework. The core hypothesis is that the LLM layers responsible for understanding revision intentions are highly aligned with those needed to generate revisions.

The framework operates in two sequential phases using a shared LLM backbone (e.g., Mistral, Llama, Qwen) and LoRA adapters:

Phase 1: Probing Important Layers (Intention Prediction)

Task: The model is fine-tuned on an Intention Prediction task ( $T_{pre}$ ), where it classifies the intention(s) behind a pair of original and revised texts.
Layer Selection: During this phase, the system calculates layer-wise importance scores based on gradient norms. Layers with high gradient norms are deemed "important" for capturing intention signals.
Dynamic Selection: Instead of a fixed set, the system tracks the frequency of layer selection over $K$ fine-tuning steps to identify a stable subset of important layers ( $S_{pre}$ ) and redundant layers ( $\bar{S}_{pre}$ ).

Phase 2: Transferring Representations (Revision Generation)

Task: The model is fine-tuned on the Revision Generation task ( $T_{gen}$ ), generating revised text given the original text and the target intention(s).
Adaptive Alignment: The importance frequency ( $G$ $G$ ) derived from Phase 1 is used to finalize the subset of layers ( $S_{gen}$ $S_{g e n}$ ) to be fine-tuned for generation.
- Shared Weights: The LoRA weights learned during the intention prediction phase are shared with the generation phase for the selected important layers.
- Freezing: Redundant layers are frozen to reduce computational cost.
Objective: This sequential transfer allows the model to leverage the "intention understanding" learned in Phase 1 to guide the "text generation" in Phase 2, specifically through the most relevant neural pathways.

3. Key Contributions

First Intention-Adaptive Layer-wise PEFT: The paper introduces the first framework to dynamically select LLM layers based on intention prediction to optimize revision generation.
Representation Transfer: It demonstrates a mechanism to align two distinct tasks (prediction and generation) by sharing LoRA weights across selected layers, effectively transferring intention representations to generation.
Efficiency on Small Corpora: The method is specifically designed for low-resource scenarios, proving effective on small, annotated revision datasets where full fine-tuning or standard PEFT fails.
Empirical Validation: Extensive experiments across three datasets (ArgRevision, ITERATER-sent, ITERATER-doc) and three LLMs (Mistral-7B, Llama3.1-8B, Qwen2.5-14B).

4. Experimental Results

Performance: Intention-Tuning consistently outperforms several baselines, including:
- Full Fine-Tuning: Often achieves comparable or better results with significantly fewer parameters updated.
- Standard PEFT (LoRA, DoRA, Bottleneck): Outperforms standard full-layer PEFT.
- Layer-wise Baselines: Beats random layer selection (LISA), static top-layer selection (IST), and gradient-based selection (IR-Baseline).
- In-Context Learning (ICL/CoT): Significantly outperforms non-finetuned LLMs.
Metrics: Achieved state-of-the-art results in SARI (addition/deletion/modification quality), Update-ROUGE, and GLEU across datasets. Notably, it showed strong performance on multi-intent revisions (ITERATER-doc).
Efficiency:
- Convergence: The method converges faster than standard PEFT baselines.
- Memory: Reduces peak GPU memory usage by 5% to 14% compared to standard PEFT by freezing redundant layers.
Layer Alignment: Analysis revealed a high layer alignment ratio (often >50-80%) between the intention prediction and revision generation tasks, validating the core hypothesis that the same layers drive both processes.

5. Significance and Impact

Bridging the Data Gap: Intention-Tuning offers a viable solution for NLP tasks with limited annotated data by leveraging the transfer of representations from a related auxiliary task (intention prediction).
Task-Aware Efficiency: It moves beyond "one-size-fits-all" PEFT by dynamically identifying and utilizing only the most relevant neural layers for specific sub-tasks, setting a precedent for more efficient LLM adaptation.
Educational and Practical Application: The framework is particularly valuable for automated writing evaluation and student feedback systems (e.g., ArgRevision), where understanding the why (intention) behind a revision is as important as the revision itself.
Future Direction: The work highlights that while LLM layers are generally aligned for these tasks, future research should explore task-specific layer variations and improve the identification of specific text spans for complex, multi-intent revisions.

In summary, Intention-Tuning provides a robust, memory-efficient, and data-scarce-friendly framework for text revision by intelligently aligning the neural pathways used for understanding intent with those used for generating text.