💻 computer science

Consistency-Aware Editing for Entity-level Unlearning in Language Models

This paper introduces Consistency-Aware Editing (CAE), a novel framework that adapts model editing techniques for efficient and robust entity-level unlearning by jointly optimizing low-rank updates across diverse entity-related prompts to ensure comprehensive knowledge removal while preserving model capabilities.

Original authors: Xiaoqi Han, Víctor Gutiérrez-Basulto, Ru Li, Xiaoli Li, Jiye Liang, Jeff Z. Pan

Published 2026-01-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaoqi Han, Víctor Gutiérrez-Basulto, Ru Li, Xiaoli Li, Jiye Liang, Jeff Z. Pan

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 "Digital Amnesia" Dilemma

Imagine a Large Language Model (LLM) as a super-encyclopedia that has read almost everything on the internet. Sometimes, this encyclopedia accidentally memorizes things it shouldn't, like a celebrity's private home address, copyrighted book chapters, or harmful stereotypes.

We want to teach the encyclopedia to forget these specific things. However, there's a catch:

We don't want to burn the whole library down. We only want to remove the specific pages about "Jackie Chan," not the pages about "Martial Arts" or "Movies" in general.
We can't just erase one sentence. If you ask, "Where was Jackie Chan born?" the model should say, "I don't know." But if you ask, "Who is the actor from Rush Hour?" it should also say, "I don't know."

The Challenge: Existing methods are like trying to erase a name from a list by crossing out one specific line. If someone asks the question in a slightly different way (a "paraphrase"), the model might still remember the answer because it only forgot the specific sentence, not the concept.

The Solution: CAE (Consistency-Aware Editing)

The authors propose a new method called CAE. Think of it as a Master Eraser that works differently than the old ones.

1. The Old Way: The "Random Eraser"

Imagine you have a whiteboard with 100 different ways to ask about "Jackie Chan" (e.g., "Where was he born?", "What is his birthday?", "Who is the guy from Rush Hour?").

Old Method: You take a separate eraser for each question. You erase the answer to Question #1, then Question #2, then Question #3.
The Problem: Because you are erasing them one by one without a plan, you might accidentally rub out the answer to Question #1 when trying to fix Question #2. Or, you might leave Question #5 untouched because your eraser slipped. The result is messy: the model forgets some things but remembers others, or it gets confused and starts hallucinating.

2. The New Way (CAE): The "Synchronized Team"

CAE changes the strategy. Instead of erasing one by one, it treats all 100 questions as a single team.

The Strategy: It gathers all the different ways to ask about "Jackie Chan."
The "Consistency" Rule: It forces all the erasers to move in the exact same direction. It's like a synchronized swimming team; every swimmer moves their arm at the exact same angle and speed.
The Result: Instead of making 100 tiny, messy scratches, the team makes one giant, clean, unified wipe that removes the entire concept of Jackie Chan from the model's brain, regardless of how the question is phrased.

How It Works (The "Under the Hood" Mechanics)

The paper digs into how the model "thinks" to find the best place to apply this eraser.

Finding the Memory: The researchers discovered that the model stores facts about specific people (like Jackie Chan) in a specific part of its brain called the MLP layers (think of these as the model's filing cabinets).
The "Key" Selection: They don't just grab random questions. They use a mathematical tool (SVD) to pick the most important 70 questions that represent the "Jackie Chan" concept best. It's like picking the 70 most representative photos of a person to ensure you recognize them from any angle.
The "Low-Rank" Update: Instead of rewriting the whole encyclopedia (which takes forever and costs a lot of money), they make a tiny, precise adjustment to the filing cabinet. It's like changing a single label on a shelf rather than rebuilding the whole warehouse.

What They Found (The Results)

The team tested this on two major benchmarks (RWKU and ToFU) and compared it to other methods.

Better Forgetting: CAE is much better at making the model say "I don't know" to any question about the target entity, even tricky ones where the name isn't mentioned directly.
Less Collateral Damage: Because the "erasers" move in sync, they don't accidentally wipe out knowledge about related topics. For example, the model still knows how to talk about movies or actors; it just doesn't know about Jackie Chan specifically.
Efficiency: It's incredibly fast. While other methods might need to retrain the whole model (like studying for a whole new degree), CAE just makes a quick, targeted edit. It can do this with just a few dozen examples.
Stability: The paper shows that even if you shuffle the order of the questions or use different models, CAE works consistently. It's robust.

The Bottom Line

The paper argues that to truly "unlearn" a specific person or entity from an AI, you can't just patch individual holes. You need a coordinated, consistent approach that understands how the AI stores that information.

CAE is like a precision laser that targets the entire "Jackie Chan" folder in the model's memory and deletes it cleanly, leaving the rest of the library perfectly intact. It solves the problem of the model remembering the answer when you ask the question in a new way, which was the biggest weakness of previous methods.

Technical Summary: Consistency-Aware Editing for Entity-level Unlearning in Language Models

1. Problem Statement

Large Language Models (LLMs) inadvertently memorize sensitive, copyrighted, or harmful information, posing privacy and security risks. While knowledge unlearning aims to remove specific unwanted knowledge, existing methods face significant limitations when applied to entity-level unlearning.

Current approaches primarily focus on instance-level unlearning, which targets the deletion of specific facts or expressions from a curated forget set. However, entity-level unlearning requires the comprehensive elimination of all information associated with a specific entity (e.g., "Jackie Chan"), regardless of phrasing or context. Existing techniques often fail in this domain because:

Overfitting to Surface Forms: They tend to overfit to the specific surface form of training prompts, struggling to generalize across diverse linguistic realizations (paraphrases) of the same entity.
Inconsistency in Editing: Location-based model editing methods, which modify parameters for specific facts, typically treat each input independently. When applied to multiple prompts related to the same entity, these independent optimizations result in inconsistent edit directions. Updates for different facts may conflict or only partially erase the target knowledge, leading to "leakage" where the model still answers correctly to some paraphrased queries.
Collateral Damage: Expanding the forget set to improve coverage often results in substantial degradation of the model's performance on unrelated facts or closely related entities.

2. Methodology: Consistency-Aware Editing (CAE)

The authors propose Consistency-Aware Editing (CAE), a framework that adapts model editing techniques for effective entity-level unlearning. The core insight is that entity-level knowledge is stored in the MLP (Multi-Layer Perceptron) modules of the transformer, specifically in the output weights ( $W_{out}$ ). CAE addresses the inconsistency problem by jointly optimizing updates across diverse prompts associated with a target entity.

Key Components:

Data Aggregation and Selection:
- The method retrieves a diverse set of facts about a target entity from Wikidata, converting them into natural language prompts ( $T_e$ ).
- These prompts are further augmented via paraphrasing using a text generation model to create an augmented set ( $T_g$ ).
- SVD-based Selection: To avoid redundancy and conflicting updates, the key vectors ( $K_f$ ) derived from these prompts are processed via Singular Value Decomposition (SVD). The method selects the top- $r$ key vectors that capture the most significant projection energy, ensuring a compact and informative subset for editing.
Consistency-Aware Optimization:
- Instead of editing each prompt independently, CAE introduces a consistency constraint.
- For each prompt $x_j$ , a residual vector $\delta_j$ is optimized to suppress the entity's factual associations.
- A consistency loss ( $L_{Cons}$ ) is added to the objective function, which minimizes the variance of the updated hidden states ( $h_j^L + \delta_j$ ) across the prompt set. This forces all residuals to align in a common direction, ensuring that the parameter update is coherent across all paraphrases of the entity.
- The final objective combines the negative log-likelihood loss (to induce uncertainty, e.g., "I don't know") with the consistency regularizer.
Multi-Layer Propagation:
- The optimized residuals are distributed across subsequent layers to ensure the suppression effect propagates through the network.
- A low-rank update (rank-one modification) is applied to the $W_{out}$ matrices of the MLP layers, effectively deleting the entity's knowledge without retraining the entire model.

3. Key Contributions

Identification of Limitations: The paper reveals that existing location-based editing methods fail at entity-level unlearning due to the inconsistency introduced by independently editing multiple prompts associated with the same entity. It also identifies that entity-level knowledge is primarily stored in MLP modules.
Novel Framework (CAE): The authors introduce a consistency-aware editing framework that jointly optimizes valuable facts under a consistency constraint. This enables coherent and comprehensive unlearning of an entire entity.
SVD-based Fact Selection: A strategy is proposed to select the most valuable facts for an entity using SVD, reducing the number of required prompts while maximizing unlearning efficacy.
Empirical Validation: The method is rigorously evaluated on two challenging benchmarks, RWKU (100 entities) and ToFU (20 entities), demonstrating superior performance over existing unlearning and editing baselines.

4. Experimental Results

The authors evaluated CAE on LLaMA3-Instruct (8B) and LLaMA3.1-Instruct (8B), comparing it against gradient-based methods (GA, DPO, NPO), prompt-based methods (ICU, RepE), and editing-based methods (MEMIT, EMMET, AlphaEdit).

Forgetting Performance: CAE achieved the lowest scores on forget sets (indicating successful removal of knowledge) across Fill-in-the-Blank (FB), Question-Answer (QA), and Adversarial Attack (AA) probes. For example, on RWKU with LLaMA3, CAE achieved an "All" forget score of 15.9, outperforming the next best editing method (EMMET at 29.0) and training-based methods.
Robustness and Generalization: CAE maintained high performance on Neighbor Sets (unrelated but semantically close facts) and Utility Sets (general capabilities like MMLU, TruthfulQA). It achieved the highest Mean FN (a metric balancing forgetting and neighbor preservation), indicating a superior trade-off between erasing target knowledge and retaining useful context.
Efficiency: CAE requires significantly fewer examples (as few as 100, often fewer) and less computational resources (single 40GB GPU) compared to full fine-tuning approaches.
Sequential Unlearning: The method demonstrated stability when removing multiple entities sequentially, without degrading the forgetting performance of previously unlearned entities.
Ablation Studies:
- Token Selection: Editing the Last Subject Token (LST) proved more effective and stable than editing the Last Token (LT).
- Consistency Weight: The method remained robust across different weights for the consistency constraint.
- Data Source: The method generalized well when using LLM-generated synthetic data instead of Wikidata.

5. Significance and Claims

The paper claims that Consistency-Aware Editing represents a significant advancement in the field of knowledge unlearning by bridging the gap between instance-level editing and the more complex requirement of entity-level unlearning.

Scalability: CAE offers a scalable solution for removing entities using only tens of carefully selected prompts, avoiding the computational cost of full retraining.
Precision: By aligning edit directions through a consistency constraint, CAE prevents the "partial forgetting" and knowledge leakage common in independent editing approaches.
Insight: The work provides insights into how entity-level knowledge is internally represented in LLMs (primarily in MLPs) and how it can be systematically deleted.

The authors conclude that CAE provides a more favorable and stable trade-off between forgetting and retention, making it a promising direction for the safe and responsible deployment of LLMs where specific entities must be comprehensively unlearned.