Pith. sign in

REVIEW 2 major objections 5 minor 14 references

Memorization Inheritance in Sequence-Level Knowledge Distillation for Neural Machine Translation

T0 review · 2 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Sequence-level knowledge distillation transmits a teacher's memorized phrases and hallucinations to students, who memorize 57% more extractively than equally sized baselines despite never seeing the original corpus.

desk verdict SeqKD students do seem to memorize more than same-size baselines, but the 57% headline is confounded by the student only seeing a teacher-replicated 18.4% subset; the secondary-ExMem finding is the real news. read the letter →

arxiv 2502.01491 v2 pith:GFKEWOQ6 submitted 2025-02-03 cs.CL

classification cs.CL
keywords sequence-levelknowledgedistillationneuralmachinetranslationextractivememorizationhallucinationcounterfactualdenoisingmodelcompressionAdaptive-SeqKD
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Sequence-level knowledge distillation (SeqKD) is supposed to compress a large translation model into a smaller one by training the student on the teacher's translations rather than on the original parallel corpus. This paper sets out to show that what gets transmitted is not only the teacher's skill but also its faults: students who never see the original training data nevertheless memorize more of it than an equally sized baseline model trained on that data directly, with extractive memorization up 57% on average and oscillatory hallucinations up 31%. The paper also claims students pick up a second layer of memorization, reproducing text that the teacher itself hallucinated. If correct, this means standard quality metrics understate the risks of distillation, and practitioners need to monitor students for memorized fragments and hallucinations, not just translation quality scores.

What carries the argument

The load-bearing mechanism is the SeqKD pipeline itself: a large teacher $\theta_T$ trained on the parallel corpus $(S_C, T_C)$ generates synthetic targets $T_T$ for the same sources, and a smaller student $\theta_S$ trains on $(S_C, T_T)$, so the student's only window into the original data is through the teacher's outputs. The paper's quantitative lens is the ExMem metric of Raunak and Menezes, which flags targets the model emits after seeing at most 75% of the source, together with replication (exact match) rates and two hallucination rates (natural and oscillatory); counterfactual memorization (CM) scores, defined as the probability gap for a target between a model trained with and without that example, are used to subgroup the data. The argument that SeqKD amplifies memorization rests on comparing $\theta_S$ with a baseline $\theta_B$ of the same size trained directly on $(S_C, T_C)$, and the argument that denoising explains the effect rests on subgroup analyses showing students translate low-quality and low-confidence examples better than the teacher.

What would settle it

Train a same-size baseline model on exactly the teacher-generated corpus $(S_C, T_T)$ that the student saw, with no other changes; if that control baseline attains the same or higher ExMem and hallucination rates as the student, then the reported amplification is a property of the data subset, not of SeqKD. Alternatively, train students on shuffled or randomly subsetted versions of $T_T$ of the same size and check whether the ExMem increase persists.

Watch

Extended reading notes

Core claim

The central claim is that memorization is inherited through SeqKD and is amplified relative to a same-size baseline. The paper measures replication (exact match) and extractive memorization (ExMem) with respect to the original parallel corpus, and finds that students, despite seeing on average only 18.4% of the corpus through the teacher's outputs, exceed the baseline's ExMem rate by 57.0% (std. 15.4) and its replication rate by 3.4% (std. 0.9). Students also produce more oscillatory and natural hallucinations than the baseline, and they exhibit 'secondary' ExMem: they memorize outputs the teacher generated, including hallucinated fragments such as a website name that appears in the teacher's translation. The paper further argues that SeqKD acts as a denoising filter, and on low-quality corpus subgroups students show amplified denoising, translating more accurately than both teacher and baseline; this creates the paper's central paradox: students memorize more overall yet improve exactly where the teacher filtered noise. Finally, the paper proposes Adaptive-SeqKD, which finetunes the teacher briefly on a high-quality subset before generating student targets, and reports that this reduces student ExMem and oscillatory hallucinations without hurting translation quality.

Load-bearing premise

The paper's comparison assumes that the student's higher memorization and hallucination rates are caused by the SeqKD training procedure; but the student only ever sees the 18.4% of the original corpus that the teacher replicated, and no baseline was trained on that same teacher-generated subset, so the difference could partly reflect the memorizability of that particular selected data rather than an effect of distillation itself.

Editorial extensions

If this is right

  • Practitioners using SeqKD should evaluate students for extractive memorization and hallucinations, since average metrics like BLEU and Comet-QE can mask systematically worse failure modes.
  • The teacher's hallucinated outputs can propagate into students as verbatim memorized text, so cleaning teacher outputs matters even when students do not literally see raw corpora.
  • Adaptive-SeqKD, which finetunes the teacher on a high-quality subset before distillation, offers a low-cost way to cut student ExMem and oscillatory hallucinations without degrading BLEU or Comet-QE.
  • Higher beam sizes during SeqKD reduce student hallucinations (especially oscillatory ones) while slightly changing quality, suggesting that the 'do not use greedy search' lesson is about failure modes as much as quality.
  • Students can simultaneously memorize more and translate low-quality examples better, so memorization and generalization improvements are not opposed in this distillation setting.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the underlying mechanism is reduced regularization from denoised targets, then the same memorization amplification should appear in other distillation settings with noisy or misaligned training data, including LLM distillation; this is a testable prediction the paper does not make.
  • The paper's Adaptive-SeqKD selects high-quality data by teacher confidence and near-memorization (chrF>90); one could replace the teacher's own confidence with an external reference-free quality estimator and apply the same finetuning step when the original training corpus is unavailable, which the paper only sketches.
  • A direct control baseline trained on the teacher-generated corpus—the missing experiment identified in the limitations—would tell whether the 57% ExMem increase is caused by SeqKD or by the selection of already-memorizable sentences; if the latter, the recommendation should be reframed as data curation rather than a property of distillation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper studies sequence-level knowledge distillation (SeqKD) for neural machine translation, asking whether students inherit instance-level memorization from teachers. Across five WMT20 language pairs, the authors train Transformer-large teachers, Transformer-base students on teacher-generated targets, and same-size baselines on original targets. They report that students replicate original targets at a slightly higher rate than baselines (+3.4% relative), have a 57% higher extractive-memorization (ExMem) rate, and hallucinate more, while also exhibiting 'secondary ExMem' of teacher-generated text. The paper further analyzes data subgroups and finds amplified denoising on low-quality subgroups, and proposes Adaptive-SeqKD, a finetuning intervention that reduces memorization and hallucinations.

Significance. If the central claim survives scrutiny, this is a valuable and timely empirical result: it challenges the default assumption that distillation only transfers quality, and it gives practitioners a concrete reason to monitor memorization-related failures in distilled NMT systems. The strengths are the breadth (five language pairs, multiple memorization and hallucination metrics, additional beam-size and model-size ablations in Appendix C), the open code/data pipeline, and the proposal of a simple mitigation with consistent effects. The main quantitative claim, however, currently rests on a comparison that does not control for the student's restricted exposure to original targets, and the subgroup analysis uses the same metric for stratification and evaluation; both are fixable with additional analysis.

major comments (2)
  1. [§2.2, Figure 3b, Table 4] The central claim that students extractively memorize more than same-size baselines is confounded by the fact that the student's training data includes original targets TC only on the subset A where the teacher replicated TC (18.4% on average), whereas the baseline sees TC for the full corpus. If A is a non-random subset (e.g., easier or noisier examples), the higher ExMem rate of θS could reflect selection into A rather than an effect of SeqKD. The text acknowledges 'within that smaller pool' but never reports θB's replication or ExMem rates restricted to A, nor a control baseline trained on A with original targets. The absolute counts in Table 4 (e.g., EN-DE θS primary ExMem 22k vs θB 12k) suggest the effect may survive such a restriction, since primary ExMem for θS can only arise on A; however, this needs to be verified explicitly. I request a restricted analysis of θB on A and, if feasible, a control model trained on A.
  2. [§3, Figure 6, Appendix F.2] The 'amplified denoising' conclusion is drawn from Comet-QE-22 improvements on subgroups that were constructed by bucketing examples on Comet-QE-22 scores of the corpus targets. This creates a circularity risk: the same reference-free metric is used both to stratify and to evaluate, so the gains may reflect regression to the mean or metric-specific bias rather than genuine denoising. The paper reports chrF and MSTTR in the appendix, but the headline subgroup claim is specifically framed in terms of Comet-QE-22. Please validate the amplified-denoising effect with an external measure (e.g., human ratings, chrF against a held-out reference, or a different QE model) or demonstrate that the pattern is robust when subgroups are stratified by an alternative quality signal.
minor comments (5)
  1. [Abstract and §2.2] The phrase 'despite not directly seeing the original training data' is inaccurate: the student directly sees TC for the 18.4% of examples that the teacher replicated. Consider rewording to 'despite seeing most original targets only indirectly through synthetic translations' or similar.
  2. [Abstract and Figure 3b] The headline 57% ExMem increase is a relative increase over a small base rate; please also report the absolute percentage-point differences so readers can calibrate practical significance.
  3. [Table 4] The column headings 'ExMem TC (#)' and 'ExMem TT (#)' do not specify the denominator; clarify whether the reported percentages are of replicated examples or of the full corpus, and whether the counts are in thousands.
  4. [Appendix F.1] The CM approximation for the four non-EN-DE pairs uses a single OUT model and only 10% of datapoints; while this is acknowledged, the subgroup analysis would benefit from a sensitivity check or from focusing the CM claims on EN-DE where the approximation is stronger.
  5. [Appendix E and D] There are minor typographical errors, e.g., 'sytems' in Appendix E and 'NatHat' in Appendix D; please proofread.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular reduction in the central memorization claim; minor self-referential subgroup evaluation and non-load-bearing self-citations, but no step is equivalent to its inputs by construction.

full rationale

The paper's headline result—students trained with SeqKD have 3.4% higher replication and 57.0% higher extractive-memorization rates than baselines with respect to the original corpus—is an empirical comparison of model outputs, not a quantity defined in terms of itself. The rates are measured against the fixed WMT20 corpus and the teacher-generated corpus; no parameter is fitted to the reported numbers, and no equation forces the student rates to exceed baseline rates. The main caveat is a selection confound, not circularity: the student sees the teacher's replicated subset (on average 18.4% of TC), so its ExMem rate is computed over a smaller and possibly more memorable pool than the baseline's. The authors acknowledge this ('within that smaller pool') but do not control it by evaluating a baseline on the same subset; this weakens the causal interpretation but does not make the derivation circular. The subgroup analysis in §3 buckets training examples by Comet-QE-22 and then measures Comet-QE-22 gains on those same buckets; this is self-referential and would benefit from an independent reference metric, but the bucket definition (source-target quality) does not determine the model-output scores, so the finding is not true by construction. The self-citations (Raunak and Menezes 2022 for ExMem, Raunak et al. 2021 for hallucination metrics, Dankers et al. 2023 for the CM approximation) supply definitions and methodology rather than load-bearing uniqueness claims; they are disclosed and do not smuggle in the target result. No circular step meeting the quoted-reduction standard was found.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper's conclusions depend on several hand-chosen thresholds for what counts as memorization or hallucination, and on the assumption that proxy metrics capture real failure modes. The most fragile load-bearing premise is that the teacher-generated subset is not inherently more memorizable; the paper does not run the control that would prove the effect is caused by SeqKD rather than by selection. No new theoretical entities are introduced.

free parameters (7)
  • ExMem source prefix threshold = 75% of source
    An example is extractively memorized if the model emits the target after seeing at most 75% of the source. Hand-chosen; the memorization-inheritance rates depend on it.
  • NatHal repetition threshold = 5 identical translations
    A source is a natural hallucination if its translation is repeated at least five times in the model's outputs. Hand-chosen.
  • OscHal bigram threshold = bigram repeated at least 10 times and at least 4x more often than in source
    A translation is an oscillatory hallucination if its most frequent bigram appears at least 10 times and at least 4 times more often than in the source. Authors report rankings are stable when the max count is varied.
  • ExMem length-ratio exclusion = target/source length ratio > 1.3 excluded
    Examples with extreme length ratios are excluded to increase precision; hand-chosen and affects the ExMem denominator.
  • ExMem min source length = source shorter than 4 words excluded
    Examples with sources shorter than 4 words are excluded from ExMem; hand-chosen.
  • NatHal Comet-QE exclusion = Comet-QE-22 > 0.85 excluded
    Examples where source and translation score above 0.85 are excluded from NatHal to avoid paraphrase false positives; hand-chosen.
  • Adaptive-SeqKD data selection thresholds = chrF > 90, confidence > 0.9, source length > 5 tokens
    Used to select the 500k sequences for fine-tuning the teacher; hand-chosen and specific to the proposed intervention.
assumptions (4)
  • domain assumption The memorization metrics (ExMem, replication, NatHal, OscHal) are valid proxies for memorization and failure modes.
    The central claim is quantified through these metrics from prior work (Raunak and Menezes, 2022; Guerreiro et al., 2023). If these metrics mis-measure, the quantitative claims shift, though the qualitative direction would likely hold.
  • domain assumption Student and baseline differ only in training targets, isolating the effect of SeqKD.
    The paper states θS and θB differ only in targets, but θS only sees the teacher-generated subset (18.4% of the corpus), which confounds the comparison (see §2.2).
  • domain assumption Comet-QE-22 is a valid reference-free quality metric for defining and evaluating subgroups.
    Used to bucket quality subgroups and to measure 'amplified denoising'; the same metric appears on both sides of the subgroup analysis (§3).
  • domain assumption The five WMT20 language pairs and the Transformer-large-to-base setup are representative of SeqKD usage.
    The paper acknowledges in Limitations that findings may not transfer to LLM-based distillation or other settings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Memorization Inheritance in Sequence-Level Knowledge Distillation for Neural Machine Translation." pith.science (2026). https://pith.science/paper/GFKEWOQ6

@misc{pith2026250201491,
  author       = {Pith},
  title        = {Pith review of: Memorization Inheritance in Sequence-Level Knowledge Distillation for Neural Machine Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GFKEWOQ6}},
  note         = {Machine review of arXiv:2502.01491}
}
read the original abstract

In this work, we explore how instance-level memorization in the teacher Neural Machine Translation (NMT) model gets inherited by the student model in sequence-level knowledge distillation (SeqKD). We find that despite not directly seeing the original training data, students memorize more than baseline models (models of the same size, trained on the original data) -- 3.4% for exact matches and 57% for extractive memorization -- and show increased hallucination rates. Further, under this SeqKD setting, we also characterize how students behave on specific training data subgroups, such as subgroups with low quality and specific counterfactual memorization (CM) scores, and find that students exhibit amplified denoising on low-quality subgroups. Finally, we propose a modification to SeqKD named Adaptive-SeqKD, which intervenes in SeqKD to reduce memorization and hallucinations. Overall, we recommend caution when applying SeqKD: students inherit both their teachers' superior performance and their fault modes, thereby requiring active monitoring.

Figures

Figures reproduced from arXiv: 2502.01491 by the authors.

Figure 1
Figure 1. An illustration of our findings. Sources [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Performance of teacher, student and base [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Hallucination metrics for θT , θS and θB and the percentual increase comparing θS to θB. SeqKD facilitates memorization [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (10 more)
Figure 3
Figure 3. Figure 3: Memorization metrics for θT , θS and θB and the percentual increase comparing θS to θB. 75% of the source, e.g., see Example (1). The ExMem rate is the percentage of extractively mem￾orized examples out of the replicated examples. We also quantify the hallucination rat…
Figure 5
Figure 5. Figure 5: Illustration of how subgroups (indicated [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Comet-QE-22 increases compared to the teacher per subgroup, averaged over language pairs. ing and θT partly filters that noise through SeqKD, training θS with reduced regularization could lead to increased memorization compared to θB. Re￾duced regularization is traditi…
Figure 7
Figure 7. Figure 7: Illustration of how model quality and memorization metrics change as a result of changing the [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Performance changes observed for the different language pairs when applying Adaptive-SeqKD. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Evaluation metrics applied to the random subgroup, for all five language pairs. The square [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Evaluation metrics applied to the quality subgroups, aggregated over language pairs with error [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Evaluation metrics applied to the CM subgroups, aggregated over language pairs with error [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Evaluation metrics applied to the confidence subgroups, aggregated over language pairs with [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Relative increases comparing students and baselines to the teacher models, for the Comet-22- [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 9 canonical work pages

  1. [1]

    To the best of our knowledge, our use is in line with CC’s terms of use

    1M CommonCrawl examples, sampled from the first 100M monolingual CommonCrawl datapoints provided by WMT20. To the best of our knowledge, our use is in line with CC’s terms of use

  2. [2]

    Thank you for your visit at our website

    Up to 1M Pulpo examples, from De la Rosa et al. (2023)’s multilingual Prolific Unannotated Literary Poetry Corpus containing verses and stanzas. Pulpo contains monolingual sequences for all language pairs, apart from monolingual Polish data. We selected the data because it is expected to be out-of- distribution compared to the WMT20 training corpora. Pulp...

  3. [3]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 4500–4511

    Distilling multiple domains for neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 4500–4511. Raj Dabre and Atsushi Fujita. 2020. Combining se- quence distillation and transfer learning for efficient low-resource neural machine translation models. In Proceedings of the Fif...

  4. [4]

    Explaining Sequence-Level Knowledge Distillation as Data-Augmentation for Neural Machine Translation

    Memorisation cartography: Mapping out the memorisation-generalisation continuum in neural ma- chine translation. In Proceedings of the 2023 Con- ference on Empirical Methods in Natural Language Processing, pages 8323–8343. Ona De Gibert, Mikko Aulamo, Yves Scherrer, and Jörg Tiedemann. 2024. Hybrid distillation from RBMT and NMT: Helsinki-NLP’s submission...

  5. [5]

    In Proceedings of the 17th Conference of the European Chapter of the Association for Compu- tational Linguistics, pages 1059–1075

    Looking for a needle in a haystack: A com- prehensive study of hallucinations in neural machine translation. In Proceedings of the 17th Conference of the European Chapter of the Association for Compu- tational Linguistics, pages 1059–1075. Varun Gumma, Raj Dabre, and Pratyush Kumar. 2023. An empirical study of leveraging knowledge distil- lation for compr...

  6. [8]

    In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 578–585

    COMET-22: Unbabel-IST 2022 submission for the metrics shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 578–585. Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for mt evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (E...

  7. [9]

    arXiv preprint arXiv:2407.10456

    Don’t throw away data: Better se- quence knowledge distillation. arXiv preprint arXiv:2407.10456. Shushu Wang, Jing Wu, Kai Fan, Wei Luo, Jun Xiao, and Zhongqiang Huang. 2023. Better simultaneous translation with monotonic knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pap...

  8. [10]

    In Proceedings of the 61st Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 8062–8079

    Towards understanding and improving knowl- edge distillation for neural machine translation. In Proceedings of the 61st Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 8062–8079. Yang Zhao, Junnan Zhu, Lu Xiang, Jiajun Zhang, Yu Zhou, Feifei Zhai, and Chengqing Zong. 2022. Life-long learning for multilingua...

Show all 14 references
  1. [11]

    In International Conference on Learning Representations

    Understanding knowledge distillation in non- autoregressive machine translation. In International Conference on Learning Representations. Yuhang Zhou, Jing Zhu, Paiheng Xu, Xiaoyu Liu, Xiyao Wang, Danai Koutra, Wei Ai, and Furong Huang

  2. [12]

    In Findings of the Association for Computa- tional Linguistics: EMNLP 2024, pages 3315–3333

    Multi-stage balanced distillation: Addressing long-tail challenges in sequence-level knowledge dis- tillation. In Findings of the Association for Computa- tional Linguistics: EMNLP 2024, pages 3315–3333. A Data and experimental setup WMT data We download the parallel corpora f...

  3. [2020]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2914–2924

    Lifelong language knowledge distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2914–2924. Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice ...

  4. [2022]

    In Proceed- ings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1816–1826

    An empirical analysis of memorization in fine- tuned autoregressive language models. In Proceed- ings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1816–1826. Milad Nasr, Javier Rando, Nicholas Carlini, Jonathan Hayase, Matthew Jagielski, A ...

  5. [2023]

    In The Eleventh International Confer- ence on Learning Representations

    Quantifying memorization across neural lan- guage models. In The Eleventh International Confer- ence on Learning Representations. Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-V oss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Er...

  6. [2024]

    Nick McKenna, Tianyi Li, Liang Cheng, Moham- mad Javad Hosseini, Mark Johnson, and Mark Steed- man

    What do larger image classifiers memorise? Transactions on Machine Learning Research. Nick McKenna, Tianyi Li, Liang Cheng, Moham- mad Javad Hosseini, Mark Johnson, and Mark Steed- man. 2023. Sources of hallucination by large lan- guage models on inference tasks. In Conference...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.