- Rating
- 2
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper frames LLM unlearning as an asymmetric generalization problem and introduces SUITE, a benchmark designed to evaluate both under-forgetting and over-forgetting. The authors also propose JensUn++, which combines stochastic prefix-mixing and dynamic loss balancing. Experiments across three LLMs and four forget topics demonstrate competitive forget-retain trade-offs while producing coherent refusal responses.
Critical strengths
-
Fine-grained forget-retain evaluation. SUITE evaluates under-forgetting and over-forgetting through diverse question formulations and semantic, syntactic, and lexical retain probes. The finding that training data substantially affects unlearning performance is particularly valuable.
-
Comprehensive empirical evaluation. The paper evaluates three LLMs under sequential and joint unlearning, adversarial queries, relearning, and quantization. JensUn++ achieves competitive forget-retain trade-offs while substantially reducing incoherent generations.
Critical weaknesses
-
JensUn++ explicitly trains models to produce refusals, while the primary evaluation measures whether target knowledge appears in generated responses. This does not establish whether the underlying knowledge has been removed or merely suppressed. The NLL, logit-lens, and relearning analyses do not fully resolve this distinction.
-
SUITE covers only four forget topics, leaving its generalizability to broader knowledge domains unclear. Moreover, JensUn++ still degrades performance on the semantically closest retain tier (s0), suggesting that preserving fine-grained forget-retain boundaries remains challenging.
Questions for authors
-
A recent unlearning position paper [1] argues that knowledge suppression and refusal should be distinguished from genuine unlearning. How do the authors view this distinction in the context of JensUn++?
-
Could the authors report membership inference attack (MIA) results to provide additional evidence that JensUn++ removes knowledge rather than merely suppressing its generation?
-
What explains the remaining performance degradation on the closest retain tier (s0)? Given this degradation, to what extent does SUITE successfully define and preserve the intended forget-retain boundary?
[1] Position: The Term "Machine Unlearning" Is Overused in LLMs.
3 comments
The pretrained model prefers the correct answer as expected. NPO ranks the GT answer last for ~70% of the facts, which makes the GT easier to identify, while RMU, JensUn, and JensUn++ are closest to chance level, so they don’t seem to suppress only the generation of the GT answer, at least on average. In our worst case forget set evaluation, RMU and JensUn still reveal 40% and 14% of the facts in generation, vs. 3% for JensUn++ (Tab. 3), showing why behavioral evaluation is important.
If the reviewer had a different MIA setup in mind, please clarify, and we will try to test it.
W2+Q3. What explains the remaining performance degradation on the closest retain tier (s0)?
We manually verified that the retain set does not contain facts from the forget topic, so the degradation on (the retain tier semantically closest to the forget set) reflects over-forgetting, not an overlap between the two sets. SUITE defines this boundary precisely, and training on it is what improves for all methods (in Table 1 in the paper, we show that training on LKF* instead leads to full over-forgetting on ).
To understand why , shows the largest drop relative to the pretrained model, we compare the retain performance on the training and evaluation questions of with that of the more distant tiers . In this way, we can check if missing generalization is the problem.
The table below shows the percentage of retain facts each method answers correctly, out of those the pretrained model answered correctly, on the retain training set (Train) and evaluation set (Eval), averaged over all models and settings, together with the forget knowledge rate .
| Method | ↓ | Train ↑ | Eval ↑ | Gap | Train ↑ | Eval ↑ | Gap |
|---|---|---|---|---|---|---|---|
| GradDiff | 38.7 | 65.4 | 40.9 | −24.5 | 80.1 | 68.4 | −11.7 |
| NPO | 35.0 | 77.6 | 41.8 | −35.8 | 83.7 | 61.5 | −22.2 |
| PDU | 35.8 | 80.1 | 49.5 | −30.6 | 86.7 | 65.8 | −20.9 |
| RMU | 36.5 | 67.4 | 66.7 | −0.7 | 80.7 | 81.5 | +0.8 |
| JensUn | 8.7 | 72.5 | 62.4 | −10.1 | 83.5 | 77.2 | −6.3 |
| JensUn++ | 4.5 | 75.5 | 70.4 | −5.1 | 87.2 | 82.5 | −4.7 |
For JensUn++, the gap between Train and Eval is about the same for and (5.1 and 4.7). The gap is similar to or better than for other methods, except RMU, which has a high forget knowledge rate , so missing generalization of JensUn++ is not the problem. As discussed in Sec. 3 of the paper, methods that try to learn the retain answers during training (GradDiff, NPO, PDU) generalize less, while methods that stay close to the pretrained model generalize better.
Thus, the main problem is the difficulty of keeping the knowledge in already on the training questions (75.5% of the questions answered correctly by the pretrained model vs. 87.2% for ). The reason for this gap seems to be the difficulty of separating facts that share an entity with the forget set, e.g., the sickness of Steve Jobs (forget set) vs. other events in his life (). This creates a trade-off between under- and over-forgetting on these facts, and when choosing hyperparameters for all methods, we favored checkpoints with better forget quality as long as the drop in retain performance is limited (App. B.6), which in turn led to a bias towards models which perform well on the forget set while sacrificing more on .
As stated in the discussion in Section 5, we think that improving performance on the retain set while keeping a low forget knowledge rate is an important area for future research.
W2. SUITE covers only four forget topics, leaving its generalizability to broader knowledge domains unclear.
It is true that SUITE covers only four topics, but they are diverse, and the results between them are overall consistent. We focused on a high-quality dataset, where the forget set and the forget-retain boundary were manually verified for all topics. The design choices behind SUITE (e.g., using the worst-case over indirect, reverse, and direct questions, defining the forget boundary on the semantic, lexical, and syntactic sets) already expose failure cases missed by previous benchmarks. At the same time, training on SUITE shows that methods that seem broken under previous benchmarks actually perform well. The dataset itself can be expanded in future work using the same principles and the provided code.
Thank you for the detailed rebuttal and additional experiments. It seems that the manuscript has not yet been updated. Do the authors plan to revise the paper to incorporate these clarifications and additional results? I would like to review the updated manuscript before making my final rating decision.
We thank the reviewer for finding SUITE and its findings valuable and for appreciating the comprehensive evaluation of the paper. We are also happy that the reviewer appreciates JensUn++ performance while substantially reducing incoherent generations. We address the weaknesses and questions below.
Q1. How do the authors view the distinction of genuine unlearning and knowledge suppression/refusal in the context of JensUn++? In classical machine unlearning, an unlearned model should be indistinguishable from one that was trained without the forget set, and [1] suggests using this definition in LLMs, or, when a different objective is used, to state it explicitly (their Sec. 6). We believe that in most LLM scenarios the objective should be knowledge removal, and more practically, that the forget facts can not be extracted from the model with any kind of query, and also under benign relearning, which can happen unintentionally during standard fine-tuning. This is also the goal of JensUn++.
We focus on this setting for 3 reasons:
We will state explicitly in the paper that our goal is not indistinguishability from a retrained model, and cite [1].
However, we are not sure if this is the point you had in mind. Please clarify if this is not the case.
W1 and Q2. Has the underlying knowledge been removed or merely suppressed. Could the authors report membership inference attack (MIA) results to provide additional evidence that JensUn++ removes knowledge rather than merely suppressing its generation?
Since we try to remove knowledge that is already present in the model, we don’t have an “oracle” (a model that was never trained on the forget set) as in MUSE or TOFU. We therefore follow RWKU and use a loss-based MIA comparing the mean NLL per token on the GT answers of forget and retain questions, extending the evaluation we did in App. B.9. Unlike RWKU, which uses short texts from Wikipedia, we use QA pairs. Forget questions are 5 direct and 5 indirect paraphrases (unseen during training) of each forget fact that the pretrained model answered correctly in at least one paraphrase. Retain questions are evaluation questions that the pretrained model answered correctly. The results below are for Llama-3.2-3B after sequential unlearning.
All methods seem to forget in this test, since forget NLL rises considerably. However, we know from Tab. 3 that GradDiff, NPO, PDU, and RMU leak on at least 29% of the facts, so we created another test where we compare the ground truth answer with plausible wrong answers.
For each fact the pretrained model knows, we used Claude Opus 5.5 to generate 5 wrong answers that follow the same format, and manually verified them, e.g., for “How many seconds after liftoff did the Challenger break apart?" (GT: 73 seconds), the answers are 47, 58, 64, 81, and 92 seconds. We then compared how likely the GT answer is relative to the wrong answers, using the average over the same 10 paraphrases as before (5 direct and 5 indirect), and ranking the 6 answers per fact. We show the total / per-token log-likelihood. AUC is the probability that the GT answer is more likely than a wrong answer (similar to the truth ratio in TOFU). If the knowledge is removed, the model cannot distinguish the GT answer from the wrong answers, so the best under this is random (Top-1=Bottom-1=16.7% and AUC is 0.5), since a model that often ranks the GT answer last still recognises it.