acceptodds
Under review as a conference paper at ICLR 2027

Same Count, Different Memorisation: Spacing and Time Since the Last Copy

Abstract

Deduplication acts on one quantity: how often a sequence appears. How much leakage a count produces is not fixed by that count. Duplicates clustered into consecutive steps, which we call massed, leak less than spaced ones. We hold the count, pin the last copy and the learning-rate position, and vary only when earlier copies occur. Against massed placement, the leakiest of four spacings, selected in sample, adds +0.733 bits of sampled exposure, a rank-based measure, at 10.6M non-embedding parameters. That is a lower bound: at the top counts the exposure estimator censors at 9 bits. This massed penalty is positive in all ten arms, 3.2M to 1B parameters. We widen each interval for variation between canary panels. It still excludes zero in seven to ten arms on exposure, depending on the component, and six on depth, a likelihood gap measured in nats. Per count cell, within panel, the penalty excludes zero from count k >= 8 on exposure and k >= 16 on depth. The penalty appears on five separately generated canary panels, each excluding zero alone, and on a calibrated attack. So leakage at a fixed count is schedule-dependent in direction: within panel the fitted gain is 1.22 to 1.95x over the drop-free nine. We claim the direction, not the size: the two confounds that could manufacture it leave the penalty at 0.59 to 1.29 times its value. A second axis, retention, appears at one configuration. Time since the last copy flattens the count exponent on exposure, +0.375 to -0.002 as retention grows tenfold. It is present at 20 tokens per parameter once retention is matched, on exposure within panel. Depth does not follow on the first panel: matching moves it by +0.0006 nats, unmoved rather than weakened. Exposure resolves on three panels separately, not pooled; depth on one. Clustering is not a defence: it loses to conditional random placement, on the attack from k >= 8. So quote leakage with its schedule and retention rather than intervening.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
memorisation, deduplication, training data order, membership inference, language model privacy

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Q64qOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper discovers the spacing effect for memorization in language model training, i.e., the observation that training data repetitions that are spaced in time tend to produce stronger memorization than repetitions massed closer together in time. By pretraining transformer models with 3m-1b parameters on standard pretraining corpus combined with canary data with controlled repetition schedules, the paper validates this claim for small kk, measuring memorization by ranking-based exposure and likelihood-based depth metrics. However, they also observe a non-monotonic effect as kk continue to increase, which the paper leave as an unclaimed phenomenon. They also observe that although massed repetitions create memorization that are lower than spaced repetition, both schedules still produce stronger memorization than random placement of canary, indicating that massed repetition alone is insufficient to be used for reducing memorization.

Critical strengths

Overall, I find several new phenomena that the paper observed quite interesting and intuitively reasonable, especially

  1. spaced repetitions have higher memorization than massed repetitions
  2. both variants still produce memorization higher than conditional random repetitions.
  3. the canary repetition schedule appears to be carefully designed to rule out other confounding factors that affect memorization, though the presentation often lack the necessary clarity to fully understand and appreciate the design. See weaknesses - clarity.
Critical weaknesses
  • Clarity: The paper frequently uses vague or metaphorical language, ambiguous references, and phrases whose logical connections are difficult to follow, which substantially hinders understanding. For example, I had difficulty interpreting the seven quantities listed in the introduction as affecting memorization.
    • Item 3 states: "Decay-weighted copy count: with the last copy pinned, wider spacing still ages the earlier ones. The ageing is constitutive, not separable, so we cross this count with k instead." It is unclear what exactly is meant by "cross this count with k," or how the preceding explanation motivates this choice. The following statement, "Crossing has power only where the ageing is not multiplicative in k," is similarly difficult to interpret without further explanation.
    • Item 5 states: "Absolute position in training: a constant rate does not remove it, and it is large (Section 4.7)." It is unclear what "a constant rate" refers to, what exactly is meant by "it" or "large," or why absolute position in training should affect memorization.
    • As another example, the caption of Figure 1 describes the right panel as "what that buys on exposure, primary arm." The term "primary arm" is not explained, and neither the caption nor the panel title ("what it buys") clearly communicates what the figure is intended to illustrate.

These are not merely stylistic issues: the ambiguous language makes it difficult to identify and evaluate the paper's claims. I encourage the authors to use more precise, self-contained descriptions, particularly when introducing key concepts and explaining figures.

  • Scope: the paper restricts itself to canary memorization, which largely made the canary's memorization independent of the remaining training data, and made the impact to practical training data repetition less clear. It would be nice if the authors could at least add a discussion on whether such observations are likely to appear for practical repetition beyond canary repetition. The only place that looks related to practical canary is the "real prose" discussed in line 396. However, it consists of only a couple lines of vague references, and I find it challenging to make sense of the result and configuration.

"Real prose. On WikiText-103 test passages disjoint from train and validation the massed penalty holds on depth in all 12 matched seeds (p= 4.9 ×10−4†). The interior contrast does not resolve on exposure (p= 0.15)."

Questions for authors
  • Could you clarify what exactly is the configuration and result for the "real prose" experiment in Line 396? This is for understanding whether the observed phenomenon hold for practical data repetition beyond random canary.
2 comments
AuthorsOpenReview ↗

Response to Reviewer Q64q

Thank you for reading the paper so closely, and for the specific examples. You are right that these passages were unclear, and we are sorry for the extra effort they caused. The design itself is unchanged, and we rewrote the text that explains it. All seven items in the introduction's list (lines 49 to 79) are now in plain words, and the sentence after the list says how the design handles each one.

Item 3 (lines 57 to 64). Before: "The ageing is constitutive, not separable, so we cross this count with k instead." After: "With the last copy pinned, a wider spacing has to put the earlier copies earlier, so they have longer to fade. That is part of what spacing means, so no design can hold it fixed. Its direction is what matters: fading can only make wide spacing leak less, so it can hide the effect but not create it. Every count also runs at every spacing, which we call crossing count with spacing. That compares spacings within each count but does not remove the fading. Crossing separates fading from a spacing effect only if fading hits some counts harder than others; if it removed the same share at every count, crossing could not tell them apart." We hope this makes clear what the earlier "not multiplicative in k" sentence meant.

Item 5 (lines 67 to 71). Before: "Absolute position in training: a constant rate does not remove it, and it is large (Section 4.7)." After, the item is headed "How far into training a copy arrives, or absolute position" and reads: "Even at a constant learning rate, a model at step 500 is less trained than one at step 4,000. In a separate test with one copy per canary set and the copy's distance to the end of training held fixed, later copies leak more (Table 6). Wider spacing moves copies earlier, so where we measured it, this pointed against the effect." In the old sentence, "it" was how far into training a copy arrives, and "a constant rate" meant the constant learning rate of item 4. Holding the copy's distance to the end fixed keeps this separate from recency, item 2. We also removed "it is large".

Figure 1 (lines 108 to 125). The right panel shows the main result: at the same count and the same last copy, spaced copies give higher exposure than massed ones. Each point is how much higher exposure is at that spacing than at s=1, paired within seed. We claim the rise from s=1, not the peak or the fall. The panel titles are now "the design: four schedules" and "the result", and the caption defines the primary arm as our 10.6M reference configuration.

Real prose (your question, lines 409 to 419). The original passage was too brief, and the paragraph now gives the full configuration. It is the primary configuration, with the same counts, 2 to 64, and the same pinned last copy, but every injected sequence is a real 256-token passage from the WikiText-103 test split, disjoint from training and validation. Tokens 0 to 127 are the prompt and 128 to 191 are scored. Exposure ranks the true span against spans resampled from all of the run's passages, injected or not. With 12 seeds at each of four spacings, the massed penalty is +0.152 nats [+0.136, +0.168] on depth, positive in all 12 seeds, and +0.240 bits [+0.168, +0.312] on exposure. Both intervals are within one canary panel. The p=0.15 in the original sentence was a different test, the peak against the widest spacing, which we do not claim.

Scope. The real-prose arm answers part of your question about canaries. Its passages are real text of the same kind the model trains on, not digit strings, and the penalty holds on them, although they are still injected copies. In practice, shuffling, curriculum, epoch structure and data ordering all set a duplicate's schedule at a fixed count. Section 4.3 reports two more experiments on real held-out passages with heavy-tailed counts from 2 to 55. In the first, a comparison of three schedules, random placement of the earlier copies approximates a shuffled corpus, and its leakage is the lowest: 0.627 nats, against 0.721 clustered and 0.847 at s=20. In the second, a calibrated membership attack (Table 5, now labelled as real prose), copies at s=20 leak more than massed ones by +0.078 in true-positive rate [+0.060, +0.098]. The gap is larger at k of 5 or less, the low counts real corpora mostly contain: +0.118 [+0.080, +0.160]. So one count can carry different leakage depending on its schedule, and a new paragraph (lines 420 to 425) makes this point. Organic copies usually sit far apart, beyond the spacings where we claim the rise, so we do not extend the rise to them.

You wrote that the unclear language made the claims hard to identify and evaluate. The abstract, the introduction and Figure 1 now state the claims in plain words. We would be happy to clarify anything that is still unclear, or to address any other concern. If the revision resolves your concerns, we would be grateful if you would consider updating your assessment.

AuthorsOpenReview ↗

Clarity revisions and results on larger models (85.1M and 302.3M)

Thank you again for your detailed comments on clarity. As the discussion closes soon, we wanted to check whether the revised text addresses your concerns. Every passage you quoted is now written in plain words, and our reply above shows each one before and after. The main changes are the abstract (lines 11 to 37), the list of confounds and the new Terms paragraph (lines 49 to 92), and Figure 1 (lines 108 to 125).

On your question about real data, the penalty holds on real prose: +0.152 nats [+0.136, +0.168] on depth, positive in all 12 seeds (lines 409 to 419). We have also since run the main retention test at two larger model sizes, 85.1M and 302.3M, and the penalty is present and larger at both (Section 4.8).

If any passage is still unclear, we would be glad to rewrite it before the discussion closes.

Reviewer_mFGqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how the spacing of duplicated training sequences affects LLM memorization. By controlling duplicate counts and the timing of the last occurrence, the authors show that spaced repetitions generally cause more leakage than clustered repetitions. The paper also investigates retention time and evaluates the effect across model sizes, datasets, and membership inference attacks.

Critical strengths
  1. I liked the experimental design. Controlling duplicate counts, recency, and learning-rate effects helps isolate the relationship between training schedules and memorization.

  2. The results are interesting and supported by extensive experiments. I particularly appreciated the calibrated membership inference experiments and the authors' careful reporting of negative results and limitations.

Critical weaknesses
  1. My main concern is generalization to realistic LLM training. As the authors acknowledge in Section 5, the largest model is substantially undertrained, while the retention experiments cover only one model size and one corpus. It remains unclear how well the findings transfer to fully trained models and multi-epoch training.

  2. Some statistical conclusions rely on uncertainty estimates transferred across canary panels, as discussed in Section 5. The exposure metric is also capped at 9 bits, which limits the interpretation of larger effects. Although the authors discuss these limitations, they still make the exact effect sizes difficult to assess. The absence of released experimental artifacts further limits independent verification.

Questions for authors
  1. How do the authors expect the spacing effect to change in fully trained models or multi-epoch settings?

  2. How sensitive are the main conclusions to the between-panel uncertainty estimates, particularly when using the larger uncertainty component?

  3. Given the limited retention experiments, what evidence supports transferring the retention findings to other model sizes and datasets?

2 comments
AuthorsOpenReview ↗

Response to Reviewer mFGq

Thank you for the careful review and the precise questions. They helped us sharpen several parts of the paper. We answer your three questions in order, then the cap and effect sizes.

Fully trained and multi-epoch settings. We expect the direction to hold in fully trained models, with its size set by how long training runs after the last copy rather than by the budget. Across our retention experiments, the size on exposure tracks that time. Two results support this. First, a 25.2M model trained to 20 tokens per parameter, the Chinchilla ratio, shows the penalty on exposure once that time is matched: +0.257 bits [+0.042, +0.472], p = 0.021. The same matched test also resolves at 5 and 10 tokens per parameter, at +0.347 and +0.395 bits. With ten times longer retention at 20 tokens per parameter, the contrast is -0.035 bits [-0.180, +0.107]. So on exposure it is the time after the last copy, not the budget, that removes the penalty. On the original panel, depth does not resolve in these matched cells. Second, across the ten main configurations, the penalty shows no resolved trend with convergence. Our 1B model is the least trained, so we use it only to show the effect exists at that size. The 85.1M run now in progress tests the first result at a larger width.

In multi-epoch training with one copy per epoch, copies sit an epoch apart, beyond the spacings our grid covers, so we do not claim the rise there. Our closest test runs about 25 epochs over a 2M-token subset. The penalty is positive at every count on per-token top-1. On the calibrated attack it is +0.153 in true-positive rate [+0.083, +0.219], and it still excludes zero when the reference models are resampled. That regime memorises by overfitting, so we read it as a check on the sign only.

Sensitivity to the larger between-panel component. At the primary configuration, the direction does not depend on it. All five canary panels were run there, and each excludes zero on its own. Across the ten configurations, the exposure tally does depend on it: 10 of 10 with 0.1071 bits, and 7 of 10 with 0.196 bits, the larger value, measured at the matched-retention cell. The three that lose are the three smallest of the ten exposure penalties, C4, natural prose and the 1B arm, and all three stay positive. On depth the two components are close, and the tally is 6 of 10 under both. At the matched-retention cell, adding either component makes the interval cover zero: [-0.042, +0.557] with 0.1071 bits and [-0.183, +0.697] with 0.196 bits. This is why the retention claim rests on three canary panels, each excluding zero alone on exposure (+0.257, +0.715 and +0.596 bits), rather than on a widened or pooled interval.

Transferring the retention findings. We agree this is a real limitation. The retention result is measured at one width and one corpus, and the paper claims it only there. The evidence for transfer is indirect. At that configuration, cells at the same retention and three budgets spanning a factor of 10 (2, 10 and 20 tokens per parameter) give indistinguishable count exponents on both measures. The matched penalty, tested in separate cells, resolves on exposure at 5, 10 and 20 tokens per parameter, and on three canary panels at 20. The 85.1M run is the first test of the matched penalty at a second width. It trains on 1.70B tokens of C4 in a single pass. We have not tested a second corpus at matched retention, and Section 5 states this limit.

The 9-bit cap and effect sizes. The cap works against our claim rather than for it. At k=64 in the primary arm, it censors 29 to 37% of canaries at s=20, against 8 to 10% when massed. So the leakier cell is truncated more, and the 9-bit penalty is measured smaller than it is. We also tested this directly. We retrained the first 24 seeds of the primary grid with 2,048 alternatives instead of 256, which raises the cap to 12 bits. On the same seeds, the penalty moves from +0.691 bits at 9 bits to +0.847 bits [+0.681, +1.018] at 12 bits, positive in all 24 seeds. The paired increase, +0.156 [-0.079, +0.390], does not resolve. Since the penalty is positive in every seed at 12 bits, the cap does not create it. This is new Appendix E.3. The check is post hoc, so the headline stays the pre-specified +0.733 bits on all 36 seeds. On size more broadly, we claim the direction rather than the size. The two confounds that could manufacture the effect leave it at 0.59 to 1.29 times its value.

Artifacts. We agree this matters for verification, and we have extended the pseudocode in Appendix A.3 (pages 19 to 25) in the revision. Its pseudocode now covers the full setup, from data preparation, the model and the training loop to the configuration of every arm in the injection ledger and the main statistical procedures, at the level a reimplementation needs.

Thank you again. If these answers resolve your concerns, we would be grateful if you would consider raising your score.

AuthorsOpenReview ↗

Results on large model

Thank you again for pointing out that the retention results covered one model size. We are glad to report that the matched test has now run at a second width. At 85.1M and 20 tokens per parameter, with the time after the last copy matched at 1,539 steps, the penalty is +0.899 bits [+0.673, +1.136] on exposure, positive in 40 of 48 seeds, against +0.257 at 25.2M. On depth, which did not resolve in the 25.2M matched cell, it now excludes zero: +0.047 nats [+0.006, +0.093]. The full result is in our update under the general comment, and we would be glad to answer any further questions.

Reviewer_b7XdOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how the timing of repeated training examples affects memorization. It holds duplicate count and the last occurrence fixed while varying the spacing of earlier copies. Within the tested schedules, spaced repetitions generally produce stronger memorization than consecutive repetitions. The paper also examines the time between the last occurrence and evaluation, finding that longer retention intervals can weaken the relationship between duplicate count and memorization. These findings suggest that memorization should be evaluated together with the presentation schedule, rather than duplicate count alone.

Critical strengths

The paper provides a useful insight: duplicate count alone is not enough to describe memorization risk. The timing of repeated examples and the amount of training afterward also matter. This gives a more complete picture of memorization and highlights factors that should be considered when comparing results.

The experimental controls are a strong part of the work. Fixing the last occurrence helps separate spacing from recency. Additional experiments address differences in canary difficulty, batch composition, and the natural text replaced by injected examples. Tests across multiple model sizes, independent canary sets, and natural passages provide further support for the central observation. The membership inference evaluation also connects the measured differences to privacy risk.

The paper clearly discusses the limits of its findings. In particular, random placement can produce less leakage than consecutive repetition, so clustering duplicates is not recommended as a defense. Separating supported findings from tentative explanations makes the conclusions more convincing.

Critical weaknesses

Although the experiments extend to around 1B parameters and include additional learning-rate schedules and training budgets, the larger models remain undertrained. It is therefore unclear how well the findings transfer to larger, well-trained language models.

The evidence is stronger for the existence of the effect than for its underlying mechanism. Some retention results are clearer on exposure than on depth, and the proposed explanation does not reliably predict results across settings. The lack of code and experimental artifacts also makes the extensive experiments and statistical analyses difficult to independently verify.

Questions for authors

Do the main findings still hold for larger language models trained with sufficient data and standard learning-rate schedules? Evidence in this setting would help clarify how broadly the conclusions apply.

2 comments
AuthorsOpenReview ↗

Response to Reviewer b7Xd

Thank you for the thoughtful review, and for the kind words about the controls. We answer your question first, then the other points.

Larger models with enough data and standard schedules. We tested data and schedule separately. On data, a 25.2M model trained to 20 tokens per parameter, the Chinchilla ratio, shows the penalty on exposure when the last copy comes 1,539 steps before the end of training. Two further canary panels repeat this on their own. When the last copy comes ten times further from the end, the penalty is gone. So at this configuration, on exposure, what matters is the time since the last copy, not the amount of data. On schedules, at 10.6M the penalty excludes zero on both measures under cosine decay and under warmup-stable-decay. Under warmup-stable-decay it is 1.68 times its constant-rate size on exposure. Across the ten main configurations, from 3.2M to 1B parameters, the penalty is positive in every one, with no resolved trend as models converge.

To test a second size at the same tokens per parameter, we are training an 85.1M model to 20 tokens per parameter, with 48 seeds at each spacing. Its design and decision rule were fixed before the first run, and the general comment states the rule. We will post the result here before the discussion closes, whichever way it comes out. Unfortunately, training a 1B model to that budget is beyond our compute, so the paper uses its 1B run only to show that the effect exists at that size.

Mechanism. We agree. As you noted, the paper keeps supported findings apart from tentative explanations, and the revision keeps that separation. We tested several candidate explanations. Three are excluded by our own measurements as what sets where the peak sits, and a two-state decay model predicts held-out sets in only 1 of 5. That is why the paper claims the effect and its direction, but not its cause.

Exposure versus depth. Retention gives two results. First, the count exponent falls as retention grows, on both measures. Second, the matched spacing penalty is claimed on exposure, our primary measure for canaries, fixed before these runs. Every canary shares a template. Exposure ranks the secret against alternatives that share that template, so it cancels template learning. Depth does not cancel it. On exposure the matched penalty resolves on three panels separately, and on depth on one. Why depth does not follow is still open, and the abstract and Section 4.8 report the split. The main spacing result holds on both measures. Within each configuration's canary panel, the penalty excludes zero in 10 of 10 configurations on exposure and 9 of 10 on depth.

Artifacts. We agree this matters for verification, and we have extended the pseudocode in Appendix A.3 (pages 19 to 25) in the revision. Its pseudocode now covers the full setup, from data preparation, the model and the training loop to the configuration of every arm in the injection ledger and the main statistical procedures, at the level a reimplementation needs.

Thank you again for your time. If these answers address your concerns, we would be grateful if you would consider raising your score.

AuthorsOpenReview ↗

Results on large model (85.1M and 302.3M)

Thank you again for your question about larger models trained with enough data. The matched test has now finished at two larger sizes. Both are trained to 20 tokens per parameter, with the time after the last copy matched. By the rule we fixed before the first run, the penalty is present at both.

ModelExposure gap (bits)Seeds positive
25.2M+0.257 [+0.042, +0.472]32 of 48
85.1M+0.899 [+0.673, +1.136]40 of 48
302.3M+1.865 [+1.081, +2.678]7 of 8

So the penalty gets larger at each size. At 85.1M, depth also excludes zero, at +0.047 nats [+0.006, +0.093]. The runs keep the constant learning rate of the main design, so for schedules the evidence is still the cosine and warmup-stable-decay runs at 10.6M. The full result is in our update under the general comment. If these results answer your question, we would be grateful if you would take them into account in your final assessment.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Summary of revisions

We sincerely thank all three reviewers for their careful and constructive reviews. We were glad that all three found the results interesting or useful, and that each commented on the care in the design. The design holds the count fixed, pins the last copy and keeps the learning rate constant, so what varies is the timing of the earlier copies.

Your comments led to one experiment added to the paper, one experiment now in progress, and rewritten text in the places listed below. The revised PDF uses the extra page for these changes. To make them easy to find, line numbers below refer to the revised PDF. The main changes are the abstract (lines 11 to 37), the list of confounds and the Terms paragraph (lines 49 to 92), Figure 1 and its caption (lines 108 to 125), the 12-bit check (lines 322 to 347), the Table 5 caption (lines 369 to 372), three paragraphs in Section 4.4 (lines 409 to 425 and 450 to 459), the pseudocode in Appendix A.3 (pages 19 to 25) and Appendix E.3 on pages 48 and 49. We also replaced shorthand in Tables 1, 2 and 6, Figure 2, and Sections 3 to 5.

While revising, we also made three small corrections. The upper end of one depth interval was +0.182 in two places and +0.179 in one. It is now [+0.081, +0.178] in all three. In Section 4.4 (line 454), the batch-256 arm's ratios to the primary arm were printed in the wrong order. They now read 0.86 on depth and 0.51 on exposure, and the matching appendix sentence now names the measures. The run totals now include the 96 added runs, 5,512 in all.

  1. Clarity (Q64q). We rewrote every passage Q64q quoted in plain words, and our reply to Q64q shows how each one now reads. A new Terms paragraph in the introduction defines the paper's own terms. The abstract no longer uses the coined terms it used before (arm, widened, drop-free, fitted gain).
  2. Larger and better-trained models (b7Xd, mFGq). A 25.2M model trained to 20 tokens per parameter, the Chinchilla ratio, shows the penalty on exposure once the time since the last copy is matched. It holds on two further canary panels too, and when that time is ten times longer, the penalty is gone. We use the Chinchilla ratio only as a reference point. Our parameter counts exclude embeddings, so at 25.2M it is about 10 tokens per total parameter. At 10.6M, the penalty grows under warmup-stable-decay and holds under cosine (Appendix F.1). A new paragraph in Section 4.4 collects the evidence on better-trained models.
  3. Experiment in progress (b7Xd, mFGq). We are training an 85.1M model, 3.4 times the size of the 25.2M model, to the same 20 tokens per parameter, with 48 seeds at each spacing. It tests the matched result at a second size and the same tokens per parameter. We fixed the rule before the first run. The time after the last copy is matched at 1,539 steps. The contrast is s=150 minus s=1 on exposure, paired by seed. The penalty counts as present if its 95% seed-bootstrap interval excludes zero, and depth is reported the same way. If its noise matches the 25.2M cell, 48 seeds detect an exposure penalty of about 0.3 bits four times in five, and the 25.2M effect of 0.26 bits about two times in three. So a result that does not resolve would not show that the effect is absent. We will post the result here before the discussion closes, whichever way it comes out.
  4. The 9-bit exposure cap (mFGq). An experiment added to the paper raises the cap to 12 bits. On the first 24 seeds the penalty is +0.847 bits, against +0.691 at 9 bits on the same seeds, and it is positive in all 24.
  5. Between-panel uncertainty (mFGq). At the primary configuration, the direction does not depend on it. Five independently generated canary panels were run there, and each excludes zero on its own. Across the ten configurations, exposure excludes zero in 10 of 10 with the component carried over from the primary arm and 7 of 10 with the larger one, and all ten stay positive.
  6. Real data (Q64q). The real-prose experiment is now fully specified in the main text, and a new paragraph discusses what the results mean for duplicates in a real, shuffled corpus.

Verification. We have extended the pseudocode in Appendix A.3 so that it covers the full setup. It now gives data preparation, the model and optimiser, the training loop with the injection schedule, and, in twenty-one numbered procedures, the configuration of every arm in the injection ledger, the four measures and the main statistical procedures, at the level a reimplementation needs. It also gives the run-time check that delivered injections match requested ones, which covers 2,371,842 injections over 5,512 runs with none lost. The resampling seed is fixed in Appendix C.1, and both corpora are public.

Thank you again for your time. We would be glad to answer any further questions.

AuthorsOpenReview ↗

Results on large model (85.1M and 302.3M)

Thank you again to all three reviewers. In our first comment we described an 85.1M model trained to 20 tokens per parameter, with the test and its decision rule fixed before the first run. That experiment has finished. We then ran the same test at a third, larger size, 302.3M. We are glad to share both results.

The penalty holds at both larger sizes, and it grows with size. Each model is trained to 20 tokens per parameter, with the time after the last copy matched at 1,539 steps. Each seed ran at s=1 and at s=150. The contrast is s=150 minus s=1, paired by seed. The penalty counts as present if its 95% seed-bootstrap interval excludes zero.

ModelSeedsExposure gap (bits)Seeds positive
25.2M48+0.257 [+0.042, +0.472]32 of 48
85.1M48+0.899 [+0.673, +1.136]40 of 48
302.3M8+1.865 [+1.081, +2.678]7 of 8

On exposure, our primary measure for canaries, the interval excludes zero at all three sizes, and the gap gets larger at each size. On depth, the secondary measure, the gap is positive at all three sizes, and at 85.1M its interval also excludes zero: +0.047 nats [+0.006, +0.093].

The 302.3M model is 12 times the size of the 25.2M model and trains on about 6B tokens of C4 in a single pass. Each of its runs takes about two days on one GPU, so it has 8 seeds. Validation loss is almost the same at the two spacings at both new sizes: 4.335 and 4.333 at 85.1M, and 4.137 and 4.139 at 302.3M. So the gap does not come from a difference in how well the models trained. Every run delivered all 504 of its injections, and the run ledger in the revised PDF now totals 5,624 runs and 2,428,290 injections.

We hope this helps with the question that reviewers b7Xd and mFGq raised about larger models trained with enough data. The revised PDF now includes both sizes in Section 4.8 (lines 521 to 523), in the appendix next to the 25.2M matched cell, and in the run ledger. The added text also moved some passages our first comment cites: the 12-bit check is now at lines 356 to 360, the Table 5 caption at 379 to 381, the Section 4.4 paragraphs at 417 to 431 and 455 to 464, and the batch-256 ratios at line 460. Eq. (1) and the primary configuration, which we expanded in reply to Reviewer Q64q, are at lines 240 to 252 and 266 to 277. Thank you for suggesting this test, and we would be happy to answer any questions about it.

AuthorsOpenReview ↗

More Results on large model (85.1M and 302.3M)

Thank you for the kind follow-up and for supporting acceptance. We have one more result that bears on the gap to larger, well-trained models. After the 85.1M cell, we ran the same matched test at 302.3M, again trained to 20 tokens per parameter, with the time after the last copy matched at 1,539 steps. The test and its rule were fixed before the first run.

ModelExposure gap (bits)Seeds positive
25.2M+0.257 [+0.042, +0.472]32 of 48
85.1M+0.899 [+0.673, +1.136]40 of 48
302.3M+2.019 [+1.465, +2.539]11 of 12

On exposure the penalty is present at all three sizes, and it is larger at each. On depth, which did not resolve in the 25.2M matched cell, it now resolves at 85.1M, +0.047 nats [+0.006, +0.093], and at 302.3M, +0.239 nats [+0.087, +0.406]. All of these numbers, including the 302.3M results with all 12 seed pairs, are in the revised PDF: in Section 4.8 (lines 521 to 524), with the full details in the appendix next to the 25.2M matched cell.

If this helps with the remaining gap you mentioned, we would be grateful if you would take it into account in your final assessment. Thank you again for your careful review and for the questions that led to these tests.

AuthorsOpenReview ↗

Short summary of the discussion for the Area Chair

Dear Area Chair,

Thank you for handling our paper. We would like to give a short summary of the discussion.

  • Reviewer mFGq wrote that the 85.1M replication and the 12-bit check largely address their concerns, and that they support acceptance.
  • Reviewer b7Xd wrote that the new experiments show the exposure effect persists, and may become stronger, up to 302M parameters, and kept their positive rating.
  • Reviewer Q64q's concerns were about clarity. We made all four changes they asked for in the revised PDF. Every symbol of Eq. (1) is now defined at the equation. The primary configuration is in one paragraph. The Table 2 counts are explicit. Tables 3 and 4 have self-contained captions.

New results. We ran the matched test at two larger sizes, each trained to 20 tokens per parameter, with the design and decision rule fixed before the first run.

ModelExposure gap (bits)Seeds positive
25.2M+0.257 [+0.042, +0.472]32 of 48
85.1M+0.899 [+0.673, +1.136]40 of 48
302.3M+2.019 [+1.465, +2.539]11 of 12

On depth, the gap now excludes zero at 85.1M, +0.047 nats [+0.006, +0.093], and at 302.3M, +0.239 nats [+0.087, +0.406]. The 302.3M row uses all 12 seed pairs of the design. All of these numbers are in the revised PDF, in Section 4.8 (lines 521 to 524) and in the appendix next to the 25.2M matched cell.

Our earlier replies first reported the 302.3M result on 8 seed pairs, the ones complete at the time. When the last 4 pairs finished, we updated the 302.3M numbers in those replies to all 12 pairs instead of posting new comments. The earlier versions remain in each comment's revision history.

Thank you again for your time.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.