Technical Summary: Forget Narrowly, Retain Broadly

Problem Statement

Machine unlearning in Large Language Models (LLMs) aims to remove specific knowledge (e.g., for privacy or safety) while preserving all other capabilities. However, existing evaluation benchmarks fail to reliably measure success due to an asymmetric generalization problem.

Current benchmarks suffer from two primary failure modes:

1. Under-forgetting: Models appear to forget target facts when queried directly but resurface the information under paraphrased, indirect, or multi-hop reasoning queries. Existing evaluations often lack the diverse query formulations necessary to test intensive generalization across the "forget" set.
2. Over-forgetting: Unlearning methods inadvertently degrade performance on knowledge disjoint from the target, including semantically related concepts, syntactically similar queries, and lexically overlapping terms. Existing benchmarks fail to define the "forget-retain" boundary at a fine-grained level, lacking probes for semantic proximity, syntactic structure, and lexical sensitivity.

The core challenge is that the "retain" set is implicitly defined as "everything else," making it difficult to distinguish genuine forgetting from suppression, or genuine retention from collateral damage.

Methodology

SUITE: Selective Unlearning of Isolated Topics and Events

To address these gaps, the authors introduce SUITE, a fine-grained evaluation protocol and training corpus designed to capture the asymmetric nature of unlearning. SUITE is constructed around four real-world factual topics (e.g., the Challenger disaster, Salem witch trials) and defines the forget-retain boundary across three axes:

• Forget Set Construction ($D_F$):

  • Direct, Reverse, and Indirect Queries: Beyond standard direct questions, SUITE includes reverse questions (providing the answer to identify the topic) and, crucially, indirect questions requiring multi-hop reasoning that omit the target entity name entirely.
  • Linguistic Variation: Every question is augmented with paraphrases and fill-in-the-blank variants generated by different LLMs to prevent style overfitting.
  • Training/Evaluation Split: The training and evaluation splits use disjoint rephrasings to ensure generalization is tested on unseen query forms.

• Retain Set Construction ($D_R$):

  • Semantic Proximity: The retain set is organized into 16 tiers of decreasing semantic proximity (Tier 0 to Tier 15). Tier 0 covers the general concept of the forget topic (e.g., the Space Shuttle program without the disaster), while higher tiers cover increasingly distant topics.
  • Syntactic and Lexical Control: The set includes questions that share the syntactic structure of forget queries but use unrelated entities, and questions that use terms from the forget set in unrelated contexts (e.g., "Challenger Deep" vs. "Challenger disaster").
  • General Knowledge: Includes broad factual queries from unrelated domains to test overall model utility.

JensUn++: An Unlearning Algorithm

Building on the insights from SUITE, the authors propose JensUn++, an unlearning algorithm that improves upon the previous state-of-the-art (JensUn) by focusing on generating valid refusals rather than suppressing knowledge (which often leads to gibberish). Key components include:

1. Distribution Matching Objective: The model is trained to match the original pretrained distribution on the retain set (using Jensen-Shannon divergence) while learning to generate a specific refusal string on the forget set.
2. Stochastic Prefix-Mixing: To ensure the model refuses even when conditioned on partial answers, the training data is augmented by sampling a random prefix length from the original answer and appending the refusal string. This forces the model to transition to a refusal from any point in the generation sequence.
3. Adaptive Loss Balancing: A dynamic scaling factor is introduced to balance the forget and retain losses based on their gradient norms, preventing one objective from dominating the other during different training stages.
4. Query Pairing: Forget and retain samples are paired by augmentation type to encourage the model to rely on semantic content rather than superficial correlations.

Key Contributions

1. Reframing Unlearning Evaluation: The paper reframes unlearning as an asymmetric generalization problem, motivating the need for benchmarks that test both intensive generalization (forget side) and extensive generalization (retain side).
2. SUITE Dataset and Protocol: The introduction of SUITE, which provides the first fine-grained annotation of the forget-retain boundary, including indirect queries for under-forgetting and graded semantic/syntactic/lexical probes for over-forgetting.
3. Empirical Analysis of Training Data: A comprehensive study demonstrating that the choice of training data is as critical as the algorithmic design. Methods trained on SUITE show substantial improvements over those trained on previous benchmarks (like LKF*), revealing that prior evaluations significantly overestimated unlearning performance.
4. JensUn++ Algorithm: A new unlearning method that achieves the best trade-off between forgetting and retaining across three different LLMs (Llama-3.2-3B, Ministral-3-3B, and Qwen3.5-9B) in both sequential and joint unlearning settings. It uniquely produces natural, consistent refusal strings, avoiding the gibberish and hallucinations common in suppression-based methods.

Results

• Evaluation Discrepancy: When evaluated on SUITE, methods that appeared successful under LKF* (e.g., PDU, RMU, NPO) exhibited severe under-forgetting (forget rates >90%) and over-forgetting (significant drops in retain accuracy on semantically close tiers).
• Training Impact: Training on SUITE significantly improved the performance of all methods except RMU, which failed to benefit from the fine-grained data, exposing a limitation in its ability to handle indirect and reverse queries.
• JensUn++ Performance:
  • Forget Performance: Achieved near-complete forgetting (worst-case forget knowledge rate of 0–3%) across all models and settings, with 0% gibberish outputs.
  • Retain Performance: Maintained high retain accuracy, with minimal degradation (1–3% drop) compared to pretrained models, even on semantically close tiers (Tier 0).
  • Robustness: Demonstrated resilience against adversarial queries (e.g., role-playing, language switching) and "benign relearning" attacks, where forgotten knowledge did not resurface during subsequent training on unrelated data.
  • Utility: Preserved general downstream capabilities (MMLU scores, generation diversity, and response quality) better than competing methods.

Significance

The paper argues that unlearning is ill-defined without a fine-grained forget-retain boundary. SUITE provides the necessary structure to diagnose both under-forgetting and over-forgetting, serving as a foundation for more reliable unlearning. The results indicate that training data quality is at least as important as algorithmic design; without data that explicitly encodes the boundary between what to forget and what to retain, even sophisticated algorithms fail to generalize correctly. JensUn++ demonstrates that with the right data and a refusal-based approach, it is possible to achieve effective unlearning that is robust, safe (producing valid refusals), and preserves model utility.