🤖 machine learning

Standard vs. Modular Sampling: Best Practices for Reliable LLM Unlearning

This paper critically evaluates conventional LLM unlearning practices, revealing that relying on single neighbor sets and standard sampling methods is suboptimal, and proposes a new framework of diverse neighbor sets combined with a Modular Entity-Level Unlearning (MELU) strategy to achieve more reliable and effective unlearning.

Original authors: Praveen Bushipaka, Lucia Passaro, Tommaso Cucinotta

Published 2026-06-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Praveen Bushipaka, Lucia Passaro, Tommaso Cucinotta

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

Imagine you have a very smart, well-read librarian (the LLM) who has read millions of books. Sometimes, this librarian remembers things they shouldn't—like a specific person's private address or a copyrighted story they weren't supposed to share. You want to teach the librarian to "unlearn" these specific facts without making them forget everything else or turning them into a confused mess.

This paper is like a guidebook for the best way to run this "re-education" class for the librarian. The authors found that the standard ways people have been doing this are actually quite messy and inefficient. They tested new methods and found much better ways to do it.

Here is the breakdown of their findings using simple analogies:

1. The Problem with the "Standard Class"

Usually, when trying to make the librarian forget something, researchers set up a class with two types of students:

The "Forget" Group: Students holding the cards with the secret info you want erased.
The "Retain" Group: Students holding cards with safe, general knowledge to keep the librarian sharp.

The Old Way (The Flawed Setup):

The "Single Neighbor" Mistake: Traditionally, the "Retain" group only included students who were directly related to the secret (e.g., if you want to forget "John lives in Paris," the retain group only knows about Paris). The authors found this is like trying to learn a language by only talking to one person; it's too narrow. They discovered you need a mix of Direct Neighbors (people directly linked to the secret) and Indirect Neighbors (people with similar jobs or themes) to keep the librarian balanced.
The "1:1 Shuffle" Mistake: The old method paired one "Forget" card with exactly one "Retain" card, over and over. The authors found this is like trying to clean a room by only picking up one sock and putting it down, then picking up the next sock. It's inefficient and doesn't clean the room well.

2. The New Solution: The "Modular" Approach

The authors propose a new strategy called MELU (Modular Entity-Level Unlearning).

The Analogy:
Imagine the librarian has to forget facts about three different people: Alice, Bob, and Charlie.

The Old "Cyclic" Way: The librarian tries to forget Alice while looking at Bob's photos, then forgets Bob while looking at Charlie's photos. It's a chaotic mix-up. The brain gets confused because the "forget" signal is paired with unrelated "remember" signals.
The New MELU Way: The librarian focuses on Alice only while looking at Alice's related photos (and safe photos about Alice). Then, they switch to Bob and look only at Bob's related photos.

Why it works:
By keeping the "Forget" and "Remember" groups strictly matched to the same person or topic, the learning signal is clear and stable. It's like studying for a math test by only looking at math problems, rather than mixing math problems with history questions.

3. What They Discovered

The authors ran experiments to see which method worked best. Here are their main takeaways:

Mix it up: Using only one type of "Retain" data (like only direct neighbors) makes the librarian either forget too little or become too confused. You need a diverse mix of related and loosely related data to get the balance right.
Stop the 1:1 Shuffle: The common practice of pairing one forget item with one retain item is a waste of time. It doesn't work well.
MELU is the Winner: The new "Modular" method (grouping by topic/person) was much more stable. It successfully erased the bad memories (high "Forget Efficacy") without making the librarian forget how to speak or answer other questions (high "Model Utility").

4. The Bottom Line

The paper argues that if you want to reliably make an AI forget specific things without breaking its brain, you need to:

Curate your data better: Don't just use one type of related information; use a mix of direct and indirect connections.
Change your training schedule: Stop pairing items randomly or in a simple 1-to-1 loop. Instead, group your training so that the AI focuses on one specific topic at a time, pairing the "forget" data strictly with the "retain" data for that same topic.

The authors conclude that this "Modular" approach provides a clear, stable path to effective unlearning, whereas the old standard methods often lead to unstable results or poor performance.

Technical Summary: Standard vs. Modular Sampling: Best Practices for Reliable LLM Unlearning

Problem Statement

Large Language Models (LLMs) trained on vast web-scraped datasets often memorize sensitive information, including personal data and copyrighted content. While LLM Unlearning has emerged as a technique to remove specific knowledge while preserving general model utility, the field lacks standardized, evidence-based best practices for dataset construction and sampling strategies.

Current benchmarks and methodologies suffer from two primary deficiencies:

Retain Set Composition: Most existing works rely on a single type of "neighbor set" (either direct or indirect connections to the forget target) combined with a general knowledge pool. This fails to reflect the complex, multi-faceted relationships in real-world data.
Sampling Strategies: The field relies heavily on de facto standards such as 1:1 sampling (matching one forget sample to one retain sample per epoch) or cyclic sampling (iterating through all retain samples while cycling through forget targets). The efficacy and stability of these heuristics have not been critically examined, potentially obscuring performance trade-offs and introducing hidden risks like catastrophic forgetting or instability.

Methodology

The authors conducted a systematic evaluation to establish best practices for LLM Unlearning, focusing on Entity-Level Unlearning (removing all knowledge of a specific entity, e.g., a person).

1. Dataset Construction (Extended WPU)

The study utilized the Wikipedia Person Unlearning (WPU) dataset, which features less popular figures to simulate realistic unlearning scenarios. The authors extended this dataset to include:

Direct Neighbor Set ( $N_d$ ): Entities directly linked to the forget target (e.g., birthplace, family).
Indirect Neighbor Set ( $N_{ind}$ ): Entities sharing semantic or contextual relationships (e.g., same profession, historical period), generated using an LLaMA 3.3 70B model to find similar professions for the target entities.
Syntactic Similarity ( $N_s$ ): Ensuring Q&A pairs in the retain set shared interrogative syntactic structures with forget targets to control for syntactic influence.
Test Set ( $D_t$ ): A balanced mix of remaining indirect neighbors and general knowledge samples to evaluate Model Utility.

2. Experimental Setup

The authors fine-tuned LLaMA 3.1 8B Instruct and evaluated three unlearning algorithms:

Gradient Difference (GD) and Negative Preference Optimization (NPO) for un-targeted unlearning.
Direct Preference Optimization (DPO) for targeted unlearning.

They compared seven distinct settings combining data practices and sampling strategies:

Data Practices: Restricting the retain set to only Direct Neighbors, only Indirect Neighbors, or a Balanced mix of both.
Sampling Practices:
- 1:1 Sequential: Matching top retain samples to forget samples.
- 1:1 Random: Randomly selecting an equal number of retain samples per epoch.
- Cyclic: Cycling through the full retain pool while iterating through forget targets.
- Modular Entity-Level Unlearning (MELU): A proposed strategy where each forget target is paired only with retain samples belonging to the same specific entity, avoiding unrelated pairings.

3. Evaluation Metrics

Performance was measured using:

Forget Efficacy (FE): Calculated as $1 - \text{Arithmetic Mean}$ of ROUGE-L, Conditional Probability, and Cosine Similarity on the forget set.
Model Utility (MU-T): Harmonic mean of the same metrics on the test set.
Token Diversity: Measured via Distinct-N to detect model collapse or degeneration.
General Knowledge: MMLU scores to assess broader model integrity.

Key Contributions

Critical Analysis of Data Practices: The study demonstrates that relying on a single neighbor set is suboptimal. A diverse retain set (combining direct and indirect neighbors) is crucial for balancing Forget Efficacy and Model Utility.
Inefficiency of 1:1 Sampling: The authors found that standard 1:1 sampling (both sequential and random) is inefficient, yielding poor forgetting results and failing to utilize the full potential of the retain set.
Proposal of MELU: The paper introduces Modular Entity-Level Unlearning (MELU), a structured sampling strategy where forget targets are paired exclusively with their respective entity-specific retain samples. This approach is shown to be more stable and effective than conventional cyclic sampling.

Results

Data Composition:
- Direct vs. Indirect: Gradient-based methods (GD) showed higher Forget Efficacy with Direct Neighbors but suffered significant drops in Model Utility. Conversely, Indirect Neighbors preserved Model Utility better but sometimes resulted in lower Forget Efficacy for gradient-based methods.
- Balanced Sets: While a balanced set did not always maximize both metrics simultaneously, it provided a necessary trade-off, preventing the extreme utility loss seen in Direct-only sets.
- Token Diversity: Indirect and Balanced sets generally maintained higher token diversity compared to Direct sets, which often led to model degeneration (especially with GD).
Sampling Strategies:
- 1:1 Sampling: Consistently failed to produce meaningful forgetting across all algorithms, regardless of the number of epochs.
- Cyclic vs. MELU: Both Cyclic and MELU significantly outperformed 1:1 sampling.
- MELU Superiority: MELU demonstrated superior stability. Under DPO, it boosted Forget Efficacy by 12% compared to Cyclic sampling while maintaining high Model Utility. For difficult targets (e.g., "Ted Kooser"), MELU achieved Forget Efficacy > 0.50, whereas 1:1 sampling failed (< 0.10).
- Stability: MELU provided a more consistent learning signal by reducing the variance caused by unrelated forget-retain pairs, leading to stable MMLU scores and token diversity.

Significance and Claims

The paper claims that the composition of the retain set and the choice of sampling strategy are critical, yet often overlooked, factors in LLM Unlearning. The authors assert that:

Relying on single neighbor sets or standard 1:1 sampling introduces hidden risks and performance ceilings.
The proposed MELU strategy, combined with robust algorithms, offers a clear and stable path toward effective unlearning.
The work provides an initial set of evidence-based best practices, urging the community to move toward more diverse, realistic benchmark datasets and structured sampling techniques rather than relying on simple heuristics.

The authors remain modest regarding their claims, acknowledging limitations such as the use of a "sister model" for dataset generation (which may limit generalizability across model families) and the lack of a deep dive into the theoretical reasons behind the stability of MELU versus Cyclic sampling. They position this work as a step toward establishing rigorous protocols for future unlearning research.