{"id":"28f3d4c8-fb98-49e2-a008-f8c7b3e0017a","arxiv_id":"2602.19612","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Fact popularity and training stage jointly change how unlearning behaves: popular facts are harder to erase, and a model fine-tuned on the evaluation data forgets more smoothly and retains more than a pretrained model.","lead":"A new benchmark, DUET, grades how well large language models erase facts, splitting facts by popularity and by whether the model was further fine-tuned. It reports that popular facts resist erasure and that fine-tuned models forget more smoothly but retain more, though the comparison is partly built into the experimental design.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SFT variant was trained on the full DUET dataset, whose city forget/retain subsets are later used for evaluation; the pretrained-vs-SFT contrast conflates training stage with double exposure to the exact test items, so the central causal claim is not established.","rationale":"I focus on the exposure confound because it undermines the entire pretrained-vs-SFT comparison, not just one quantitative summary. Any difference in forgetting dynamics or retention between the two checkpoints can be explained by the number of times the model has seen the exact evaluation examples, rather than by the training paradigm. The paper frames the result as 'training type' shaping unlearning (Section 1), but the operationalization does not isolate training type. The abstract's own phrasing 'SFT step on the forget data' turns the confound into an intentional treatment, which makes the causal claim circular: the SFT model is expected to be more stable on data it has been fine-tuned on. The proposed held-out test is decisive and feasible with existing code. I note the Table 1 popular-row inconsistency as additional support that the unqualified '10-50% higher retention' claim is not robust, but it is secondary to the design problem. The reader's weakest_assumption identifies the same confound; I agree and would keep the CONDITIONAL verdict.","tokens_in":13158,"tokens_out":8639,"duration_ms":75699,"concrete_test":"Retrain the SFT variant on DUET with the city forget and retain evaluation sets removed (or on a disjoint random half of DUET), then rerun unlearning on the held-out city forget set at lr=2e-5. If the SFT advantage in retain ROUGE and stability disappears or reverses, the effect is due to direct exposure to the evaluation items rather than to supervised fine-tuning per se.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that SFT models respond differently to unlearning because of the training stage — is not identifiable in the current design. Section 3.1 states that the SFT variant is trained on the full DUET dataset (28.6k samples), and the unlearning experiments are run on city forget/retain subsets that are themselves drawn from DUET (Section 2, Figure 2). The SFT model has therefore seen the exact question–answer pairs used to measure forgetting and retention, while the pretrained checkpoint has not. The abstract's own wording, \"An SFT step on the forget data,\" makes this an explicit part of the treatment. As a result, the reported 10–50% higher retention and smoother forgetting may reflect the fact that the SFT model received additional gradient steps on the evaluation items, not a general property of supervised fine-tuning. The Limitations section lists scope, popularity proxies, and metrics, but does not mention this exposure confound. A secondary inconsistency: Table 1 shows for popular facts SFT retain ROUGE is lower than pretrained (e.g., GA 0.733 vs 0.864), which is difficult to reconcile with an unqualified \"10-50% higher retention\" claim; this underscores that the comparison is not controlled.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DUET, a 28.6k Wikidata-derived benchmark of question–answer pairs annotated with fact popularity via Wikipedia sitelink counts and LLM-based salience scores. Using this benchmark, the authors compare machine unlearning on a pretrained LLaMA-3.1-8B checkpoint and an SFT variant trained with LoRA on the full DUET set, applying GA, GD, and NPO across rare/popular forget splits. The central empirical claim is that pretrained and SFT models respond qualitatively differently to unlearning: SFT models exhibit smoother forgetting and 10–50% higher retention, while pretrained models are unstable and can even 'relearn' popular facts. The same qualitative pattern is reported for Gemma-7B, Qwen-2.5-7B, an additional algorithm (UNDIAL), and an LLM-as-a-Judge protocol.","tokens_in":13497,"tokens_out":3400,"duration_ms":33397,"significance":"If established, the claim that unlearning behavior depends jointly on fact popularity and training stage would be an important contribution to LLM unlearning evaluation. The DUET benchmark itself, with its popularity stratification and multi-domain subsets, is a potentially useful resource, and the paper works with multiple models, multiple unlearning algorithms, and several evaluation lenses. The main evidence, however, does not currently support the headline causal claim. The SFT model was trained on the same DUET data that is later split into the forget and retain evaluation sets, so the pretrained-vs-SFT contrast conflates training stage with direct exposure to the exact evaluation examples. The paper also lacks error bars, reports a single configuration as the main result, and its own Table 1 contradicts the unqualified '10–50% higher retention' statement. The benchmark and the research question are valuable, but the central comparison needs a cleaner experimental design before the conclusions can be accepted.","major_comments":[{"comment":"The main confound is that the SFT variant is trained on the full DUET dataset (28.6k samples) while the forget and retain evaluation sets are subsets of DUET. The SFT model therefore receives additional gradient steps on the exact question–answer pairs used to measure forgetting and retention, whereas the pretrained checkpoint has not seen any of them. The abstract's own phrasing — 'An SFT step on the forget data' — makes this exposure part of the treatment. The observed 10–50% higher retention and smoother forgetting may simply reflect double exposure to the evaluation items, not a general property of supervised fine-tuning. This confound is not mentioned in the Limitations section. The central causal claim is not identifiable from the current design. To fix it, the SFT training data must exclude the forget and retain evaluation splits (or the SFT model must be compared against a pretra","section":"§3.1, Figure 2"},{"comment":"The abstract and Section 3.2 claim '10–50% higher retention' for SFT models, but Table 1 shows the opposite on popular facts. For popular forget at lr=2e-5, SFT retain ROUGE-L is lower than pretrained retain ROUGE-L for all three algorithms: GA 0.733 vs 0.864, GD 0.760 vs 0.884, NPO 0.849 vs 0.888. The claim is only supported on rare facts, and even there the size varies by algorithm (e.g., NPO rare: 0.944 vs 0.790). The unqualified '10–50%' statement is therefore not supported by the paper's main table. This is a load-bearing numerical claim and must be corrected, qualified by popularity, or accompanied by the full breakdown.","section":"Table 1"},{"comment":"No uncertainty quantification is reported. The main results appear to come from a single run at a single learning rate, with no seeds, no error bars, and no statistical tests. The hyperparameters (two epochs, lr=2e-5) are selected post hoc from the same curves shown in Figure 3, so the reported differences may be partly overfit to the chosen configuration. Since the paper makes quantitative claims (e.g., '10–50% higher retention', 'roughly half the risk of catastrophic forgetting'), it should report multiple random seeds/initiations and show variability, or at minimum explicitly state that these are single-run observations.","section":"§3.1, Figure 3"},{"comment":"The abstract says 'An SFT step on the forget data yields smoother forgetting...' but Section 3.1 describes the SFT variant as trained on the full DUET dataset, not specifically on the forget set. These are different treatments. If the intended comparison is SFT-on-forget-data, the experiments do not implement it. If the intended comparison is SFT-on-all-data, the abstract should be reworded, and the exposure confound in the first major comment still applies. The paper should clarify exactly which SFT training set is used and align the treatment description throughout.","section":"Abstract vs §3.1"}],"minor_comments":[{"comment":"The figure is a 2x2 grid with small text; consider enlarging or using a table format to improve readability.","section":"Figure 1 caption"},{"comment":"The captions say 'Left: forget; right: retain' but do not specify axes or the exact split names in a consistent way. Please state the x-axis (learning rate) and y-axis (ROUGE-L) explicitly.","section":"Figure 3/4/5 captions"},{"comment":"The Limitations section lists scope, popularity proxies, and metrics, but it does not acknowledge the training/evaluation overlap that is the main threat to validity. Please address it directly.","section":"Limitations"},{"comment":"The UNDIAL results table lacks a column for the algorithm name in each block; it appears only in the section text. Consider adding an 'Algorithm' column for self-contained readability.","section":"Section C.6, Table 9"},{"comment":"The table reports token-level rank changes but not standard deviations. Since ranks can be highly variable across examples, adding dispersion or a short note would help assess reliability.","section":"Section C.4, Table 6"}],"recommendation":"major_revision","confidential_remarks":"The DUET benchmark and popularity stratification are potentially useful assets, but the central pretrained-vs-SFT comparison cannot be accepted as evidence for the causal claim while the SFT model has seen the exact evaluation examples. A redesign with held-out SFT data, plus uncertainty quantification, is needed before the paper's main conclusion can be published. The contradiction in Table 1 regarding retention on popular facts should also be resolved in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe DUET benchmark is a real contribution, but the paper's central claim about pretrained vs SFT unlearning behavior is not established by the current design. The SFT variant was trained on the full DUET set, and the forget/retain evaluation pairs are subsets of DUET. So the SFT model has seen the exact test items, while the pretrained model hasn't. The reported 10-50% higher retention could just be double exposure. That's a load-bearing confound, and the Limitations section doesn't mention it.\n\nWhat's genuinely new: a 28.6k Wikidata-derived QA benchmark with explicit popularity splits, and the observation that popularity interacts with unlearning in a non-trivial way. They ran three unlearning algorithms across three base models (LLaMA, Gemma, Qwen), plus a distillation method, and consistently see that rare and popular facts behave differently. The benchmark itself is reusable and the framing—evaluate both pretrained and SFT checkpoints—is worth taking seriously.\n\nThe soft spots are more than cosmetic. First, the exposure confound makes the pretrained-vs-SFT contrast uninterpretable. If they want to claim that supervised fine-tuning per se changes unlearning, they need to train SFT on data that is disjoint from the evaluation sets, or at least show the effect holds when the SFT data is unrelated. Second, Table 1 actually contradicts the unqualified '10-50% higher retention' claim: for popular facts, the SFT model's retain ROUGE is lower than the pretrained model's (e.g., GA 0.733 vs 0.864). The headline claim should be conditional on popularity. Third, no error bars and post-hoc hyperparameter selection make it hard to tell how robust the differences are.\n\nNone of this kills the benchmark. The popularity annotation and the separate analysis of pretrained vs SFT checkpoints are a useful starting point, and the qualitative finding that popular facts are more resistant to unlearning in pretrained models (even if the SFT comparison is confounded) is plausible. But the paper would need a major revision before I'd trust the causal story.\n\nWorth sending to referees—the benchmark deserves scrutiny and the question is important. But my verdict would be 'major revision' with a request for a controlled SFT setup and a reconciled headline claim.\n\nBest,\n\n[Your name]","headline":"The DUET benchmark is a genuine resource, but the paper's central claim about pretrained vs SFT unlearning is compromised because the SFT model was trained on the exact evaluation pairs.","tokens_in":13964,"tokens_out":2842,"would_cite":true,"duration_ms":23668,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that machine unlearning is jointly shaped by fact popularity and the model's training stage, with supervised-fine-tuned models forgetting smoothly and retaining 10-50% more knowledge than pretrained models.","keywords":["machine unlearning","fact popularity","supervised fine-tuning","pretrained models","knowledge retention","LLM benchmarking","Wikidata","catastrophic forgetting"],"falsifier":"Train an SFT model on DUET with the forget and retain evaluation sets held out (or replaced by a disjoint set of comparable facts), re-run the unlearning comparison, and see whether the retention advantage and smoother forgetting persist. If they disappear, the paper's headline claim is confounded.","tokens_in":13052,"feed_emoji":"🧠","tokens_out":5551,"duration_ms":42902,"temperature":0.7,"pith_summary":"The paper argues that the standard assumption that all facts are equally forgettable is wrong. It introduces DUET, a benchmark of 28.6k Wikidata-derived question-answer pairs annotated with fact popularity, and shows that the outcome of unlearning depends on both how popular a fact is and whether the model is a pretrained base model or one that has gone through supervised fine-tuning (SFT). The central experimental finding is that SFT models forget more smoothly, tune more stably, and retain 10-50% more of the retain set, while pretrained models are unstable, prone to catastrophic forgetting, and can even improve on popular facts when unlearning is applied. If this result holds, unlearning pipelines and benchmarks must stratify by fact popularity and account for the training stage, and a preliminary SFT step becomes a practical way to stabilize forgetting.","feed_headline":"Fine-tuning first boosts unlearning retention by 10-50%","feed_subtitle":"A 28.6k-fact benchmark shows fact popularity and training stage jointly decide what models actually forget.","key_machinery":"The central object is the DUET benchmark: 28.6k Wikidata-derived question-answer triplets, each annotated with a popularity score based on Wikipedia sitelink counts and validated against LLM-based salience judgments. The benchmark stratifies facts into rare and popular forget sets at 1%, 5%, and 10% scales, with a retain intersection and compact retain subsets for efficient evaluation. The paper then compares Pretrained and SFT versions of LLaMA-3.1-8B (and Gemma-7B, Qwen-2.5 7B) under three gradient-based unlearning algorithms. The popularity annotation is what carries the argument: it turns unlearning from a single scalar outcome into a function of where the fact sits in the model's memory","core_discovery":"On the paper's own terms, the discovery is that pretrained and supervised-fine-tuned variants of the same architecture respond qualitatively differently to the same unlearning algorithms. On popular facts, the pretrained model's ROUGE-L on the forget set increases during gradient ascent, gradient difference, and NPO, as if the unlearning signal were additional fine-tuning, whereas the SFT model forgets as intended. On the retain set, SFT models show 10-50% higher retention and roughly half the risk of catastrophic collapse, and token-level and hidden-state diagnostics show that SFT produces localized representation changes while pretrained models barely change internally. The paper concludes","pith_inferences":["The paper attributes the SFT advantage to the training stage, but the SFT model was fine-tuned on the same evaluation examples; a holdout-SFT control would test whether the 10-50% retention gain is a training-regime effect or a double-exposure artifact.","Wikipedia link popularity is a plausible proxy for pretraining corpus frequency; a direct test would measure how unlearning difficulty tracks token counts in the base model's pretraining data.","The rare/popular asymmetry suggests adaptive unlearning recipes: stronger signals for rare facts, gentler ones for popular facts, with an optional preliminary SFT pass.","If the double-exposure confound is real, the paper's main quantitative claim would shrink, but the qualitative popularity asymmetry is independently supported by the multi-domain and intrinsic-analysis results in the appendix."],"forward_implications":["Unlearning evaluations will need to report results stratified by fact popularity; aggregate curves hide the fact that popular and rare facts erase at different rates.","A preliminary supervised fine-tuning step on the forget data becomes a practical recipe for more stable, controllable forgetting with higher retention.","Pretrained models can 'relearn' popular facts during unlearning, so practitioners who skip fine-tuning risk the exact opposite of the intended effect.","SFT models' retention risk is concentrated in popular-fact removal, while pretrained models degrade abruptly regardless of fact type, pointing to different failure modes."],"fun_headline_variants":["SFT models forget 10-50% more stably than pretrained","Fact salience and SFT shape unlearning, benchmark shows","SFT boosts unlearning retention by 10-50% over pretrained","Forgetting depends on training stage: SFT wins by 10-50%","28.6k-fact study: SFT forgets smoother than pretrained"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that supervised fine-tuning stabilizes unlearning rests on comparing an SFT model that was trained on the exact forget/retain evaluation examples against a pretrained model that never saw them, so part or all of the observed advantage may be double exposure to the test items rather than a property of the training stage.","fun_headline_variants_meta":{"raw":{"variants":["SFT models forget 10-50% more stably than pretrained","Fact salience and SFT shape unlearning, benchmark shows","SFT boosts unlearning retention by 10-50% over pretrained","Forgetting depends on training stage: SFT wins by 10-50%","28.6k-fact study: SFT forgets smoother than pretrained"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2712,"prompt_tokens":656,"completion_tokens":2056,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":400,"completion_tokens_details":{"reasoning_tokens":1956}},"tokens_in":400,"tokens_out":2056,"duration_ms":13690,"temperature":1.0,"reasoning_tokens":1956,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:34:04.519912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train an SFT model on DUET with the forget and retain evaluation sets held out (or replaced by a disjoint set of comparable facts), re-run the unlearning comparison, and see whether the retention advantage and smoother forgetting persist. If they disappear, the paper's headline claim is confounded.","supporting_citations":[],"review_version":2}