acceptodds
Under review as a conference paper at ICLR 2027

Is your LLM a sequence model on the training history? The origins and consequences of anticipation

Abstract

Recent work showed that, in cyclic fine-tuning settings, Transformers exhibit striking anticipatory recovery: the loss on an upcoming document decreases before that document is revisited, even though models hold no explicit memory of training history. Despite early efforts to explain this behavior, our understanding of anticipation remains limited. We first show that anticipation is broader than previously demonstrated: beyond individual-documents, Transformers anticipate upcoming data domains, such as English and French, and the behavior extends to pre-training, even to models 5000 times smaller. Second, we show that anticipation can be interpreted as sequence model-like behavior. Anticipating the next domain leads to lower training loss, and this loss advantage realizes a substantial fraction of that achievable by sequence models that explicitly learn class order. To a limited extent, anticipation transfers even to second-order structures, where the next domain can only be predicted from the previous two. Third, we offer a potential explanation of anticipation as gradient alignment developing through an implicit bias of SGD in cyclically ordered training. The resulting picture helps account for the surprising optimizer dependence of anticipation: at matched held-out loss, AdamW develops strong anticipation, while Muon does not anticipate. Finally, we show that structured training order can affect what the model learns, and can act as a source of representational bias during optimization.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLMs, implicit meta learning, training dynamics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CGiAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Previous work has identified an interesting behavior: when fine-tuning an LLM by cycling through a few documents one after the other, always in the same order, the loss on a single document will decrease before seeing the document, highlighting some anticipatory behavior. The paper reproduces this behavior in a much smaller setup, which enables more careful experimentations and studies, both empirically and theoretically, on which properties of optimization give raise to this behavior.

Critical strengths

The phenomenon studied here is intriguing and challenges our understanding of deep learning and stochastic optimization. The paper sheds some light into this behavior with well designed ablations; it feels like the paper is looking at the right quantities to understand this behavior, as varying them can yield significantly different dynamics.

Critical weaknesses

From my understanding, the paper falls short of some of its claims in different ways:

  • the theoretical analysis holds for stochastic gradient descent; yet Table 1, as far as I understand it, seems to suggest that the behavior only occurs for AdamW. It is therefore unclear how the analysis explains the desired behavior.
  • related to the previous point, there is no quantitative relationship linking T and Ring-related quantities to anticipation directly in the main text. It may be only a presentation problem, in which case I would encourage the authors to make this link clearer.
  • the core experiments hold for contexts of length 8 (!), which is very far from anything standard in large language modeling, and anticipation goes away when increasing the context length. This suggests that the current setup might not be suited to explain the phenomenon observed at scale.

I am happy to increase my scores if the authors answer my concerns or if my remarks are unfair.

Minor: it would be useful to have a description of the data in the main text.

Questions for authors

What does AdamW (no momentum) mean, and how is it different from SGD?

3 comments
AuthorsOpenReview ↗

Response to review

Thank you for your review! We are glad you found our paper’s topic intriguing, our analysis shedding light on this behaviour, and our ablations well-designed. We address your concerns below. All new experiments and results are added to our updated manuscript, and are marked in blue.

1. Theory holds for SGD but experiments find SGD does not anticipate: We break down our answer into multiple parts.

What our theory holds for: our theoretical results hold for SGD, and a simplified version of AdamW (fixed preconditioner), and we show in Figure 5 that this simplified approximation is consistent with AdamW’s anticipation. So the essence of the theory holds for both SGD and AdamW.

Clarifying our implicit bias argument: We derived implicit biases that can enable anticipation. When anticipation happens, our explicit regularization experiment shows that we can manipulate it, hence anticipation is linked to the implicit bias we derived. However, for anticipation to emerge, the implicit biases alone are not enough, we also need the right conditions for them to accumulate, such as long enough bursts, training horizon, etc.

Clarifying SGD: Since our SGD experiment did not show anticipation, during the rebuttal, we tuned the learning rate (lr=0.3) and found that SGD anticipates as well (successor advantage, SA ≈0.16\approx 0.16). We also added new analysis (Figure 15) about why the original, lower learning rate failed: the realized effect of successor-gradient alignment (SGA) was much smaller due to low learning rate, so anticipation could not accumulate on the tested time horizon.

Improved narrative: Our new analysis makes our explanation of optimizer-dependence much cleaner: instead of having to compare T2T_2 and T3T_3, now effective SGA (SGA measured along the optimizer's update) is predictive of anticipation, in accordance with our Theorem C.3. We have also updated Figure 6 (right) to include an SA vs SGA plot, which confirms this relationship empirically for SGD, AdamW and Muon. This ties our theory and experimental results closer.

2. No quantitative relationship linking T and Ring-related quantities to anticipation directly in the main text: Thank you for your observation. We have edited section 4.1 to include our quantitative formula in Result 2: T2T_2 increases SGA, and SGA in turn increases SA (Result 3). See also our response to 1: the new Figure 6 on page 7 (right) now quantitatively links SA (anticipation) to SGA. In addition, our explicit regularization analysis links T2T_2 (what our regularizer estimates) to control on SA.

3. Context length 8

We appreciate the point that context length 8 is non-standard in language modeling. We have picked this setting due to computational constraints, as a simple setting that allows us to cheaply manipulate anticipation.

For the rebuttal, we have rerun our main experiment (see Figure 8) for context lengths of up to 1024, and found that anticipation is preserved, but requires longer bursts (e.g. 128).

This confirms that context length 8 is not the only setting where anticipation emerges (our theory also does not depend on context length), rather, it’s a cheap setting where anticipation can be reproduced on a short training horizon.

What does AdamW (no momentum) mean, and how is it different from SGD?

It is AdamW with β1=0\beta_1 = 0. AdamW modifies SGD in multiple ways apart from adding momentum. It also scales the learning rate per-parameter, using an exponential moving average (EMA) of the squared past gradients (controlled by the so-called second moment β2\beta_2). This normalizes step sizes across parameters that may have different gradient magnitudes. AdamW also adds weight decay, which is not automatically present in SGD.

We amended our manuscript to include a description of the data on page 3.

We also added third- and fourth-order experiments: models anticipate the successor class even when it depends on the last three or four bursts together (Figures 4, 13).

Thank you again! We are happy to address any remaining questions.

I thank the authors for the detailed rebuttal, it addresses most of my concerns; I believe it is an interesting and solid contribution that deserves being accepted to the conference. I have updated my score accordingly.

The fact that the behavior only appears for SGD for sufficiently large learning rates is interesting and it is not directly clear to me why this is the case from the theory. I would encourage the authors to emphasize this point more clearly in the main text. I am also wondering how much of Figure 8 can be explained by this insight.

AuthorsOpenReview ↗

Response to follow-up questions

Thank you for your continued interest in our work and for increasing your score! Here are our answers to your follow-up questions. We updated the manuscript to emphasise these points.

SGD learning rate vs theory. Our theory does predict dependence on the learning rate. The learning rate η appears in our theoretical results twice:

  • Theorem 2: the alignment pressure of Ring with respect to Bursty is proportional to η
  • Theorem 3: the successor advantage (SA) gained per update is η times the (effective) SGA

The second point is shown in Fig. 15 (SGD ablation). With the low learning rate, SGA still increases, but the realized effect (step size × SGA) is much smaller, so anticipation does not develop within the training horizon (step size is learning rate adjusted with the effect of momentum and grad clipping).

Is the context length result related? The same reasoning can be applied to why longer contexts need longer bursts (Fig 8). Both effects above act once per update, so their total effect within one burst grows with the learning rate times the number of updates in the burst. If this product is small, the model barely changes during a burst and hence the order of the bursts does not matter, and no anticipation happens.

We hypothesise that longer context means slower for class-specific learning. For long contexts, the model can identify the class from the prefix, and so each update changes the class information stored in the weights less. Hence more updates per burst are needed to get to anticipation. Consistent with this, Fig. 22 shows that at context length 128 and late in training, the per-token SA is largest at the early token positions.

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

The paper studies anticipation under structured training orders: a Transformer becomes preferentially adapted to an upcoming class before revisiting it. The authors extend anticipatory recovery from repeated documents to recurring domains, introduce Successor Advantage, and examine optimizer dependence, second-order transitions, and changes in inter-class gradient geometry. They explain the behavior through a local gradient-alignment bias induced by sequential updates.

Critical strengths
  1. The controls are well chosen. Ring and Bursty separate predictable successor identity from burst structure, while the paired Shuffled condition preserves the training-example multiset. Evaluation on held-out examples supports anticipation beyond recovery on repeated samples (§2; Appendix B.5).

  2. The optimizer comparison is interesting. AdamW shows strong anticipation while Muon does not, despite similar aggregate held-out losses. Resetting AdamW’s state at burst transitions also shows that persistent optimizer-state memory across bursts is unnecessary in this setting (Table 6; Figure 12).

  3. The gradient-alignment analysis offers a plausible explanation. The suppression of anticipation through negative regularization provides experimental support for this account (§4.3). The order-dependent gradient relationships are also a useful empirical observation (§6).

Critical weaknesses
  1. My main concern is the incremental contribution over the anticipatory-recovery phenomenon discussed in §7. Extending recovery from repeated documents to held-out examples from recurring domains is meaningful, but I find the conceptual advance modest. It broadens the unit of anticipation while retaining the central observation of preferential recovery under predictable training order.

  2. The optimizer dependence and alignment analysis add more than this extension alone. However, I find their combined scientific significance insufficient for acceptance. They provide a useful explanation and further characterization of anticipation, but do not, in my assessment, establish a sufficiently substantial advance over the starting phenomenon.

  3. The second-order experiments broaden the findings, although their scope is limited. Both tested configurations show their largest positive SA at 32 steps per burst; the effects at 16 and 64 steps are much weaker or negative (Table 4). I regard this as evidence for a higher-order effect under particular conditions, with its broader range of occurrence still unclear.

Questions for authors
  1. Which result do you regard as the main conceptual advance beyond cyclic-document anticipatory recovery? Please distinguish the domain-level extension from the insight added by optimizer dependence and the alignment mechanism.

  2. The sign of T3T_3 in Theorem C.1 is opposite to the measurement formula in Appendix D.5 after removing the positive learning-rate factor. Which convention was implemented? This leaves norm ratios unchanged but affects the cosine interpretation in Figure 11.

  3. Appendix D.1 supplies the true class to the oracle. With geometrically distributed bursts, history alone does not reveal whether the next batch will switch classes. How does this oracle reference relate to the loss achievable by a predictor conditioned only on training history, and how should the reported 30%/40% figures be interpreted?

4 comments
AuthorsOpenReview ↗

Response to review

Thank you for your review! We are glad you found our optimizer comparison interesting, our gradient-alignment analysis plausible as an explanation and empirically well-supported, and our controls well-chosen. All new experiments and results are added to our updated manuscript, and are marked in blue.

Q1: We believe our paper makes multiple parallel contributions that (1) broaden anticipation to novel settings, (2) explain it theoretically, (3) reframe it as curious sequence model-like behavior, and (4) provide novel analysis on its consequences on what models learn. The main advance is (3): networks act, to a degree, like sequence models of their training history. Two results show this.

  • Lower training loss. Averaged over all steps, the training loss is lower with a predictable order than with shuffled data (Figure 2). So this is not just a side effect: ordinary training gives models an incentive to track their own training history.
  • Anticipation up to fourth order. As expected of a sequence model, networks pick up longer patterns in their training history: they anticipate the successor class even when it depends on the last two, three or four bursts together (Table 4, Figures 4, 13). Note that third and fourth order anticipation are new experiments.

As for the distinction you asked about: moving from repeated documents to whole domains (languages) shows that anticipation goes beyond memorization, since models anticipate held-out text. The optimizer and gradient-alignment results help explain why it happens.

Strength of contribution. We would like to argue that our contribution goes far beyond extending anticipatory recovery from repeated documents (previous work) to held-out examples from recurring domains. Besides the two results above:

  • Conceptually, reframing anticipation as sequence model-like behaviour lets us measure empirically to what extent a Transformer can be a sequence model of the training history. Surprisingly, we find that Transformers are ~40% there (40% of the gain of an oracle that is told the class) despite no explicit memory of training history.
  • Theoretically, building on previous work on SGD implicit biases, we derive new implicit biases for Ring training and link them directly to anticipation. This provides the first explanation of anticipation that’s grounded in deep learning theory (Yang et al provided an intuitive toy computational model). We believe that explaining an existing phenomenon is in itself a valuable contribution.
  • Empirically, we support our theory by explaining optimizer dependence and demonstrating control on anticipation via explicit regularization. New in this revision: if we regularize class c towards c+5, five classes ahead in the Ring (and not the next class, c+1), some of anticipation transfers to class c+5 (Figure 18).
  • Finally, we explicitly test how much a specific order alters models’ gradient similarities, and surprisingly find that order itself can make Spanish’s gradients more aligned to Finnish’s than Portuguese’s.

Second-order experiments. Second-order anticipation is a much more difficult task, so it is expected that the effect is smaller. However, the effects are significant based on the variance estimated from random seeds. To broaden the conditions of anticipation, we added new third and fourth order experiments that also demonstrate anticipation. At third order, the effect holds over a range of burst lengths (Figure 14).

Q2: Theorem C.1’s formula was correct, but we implemented the one in the old D.5, so the signs indeed flip on the old appendix Figure 11, which plotted the cosine similarity between T2T_2 and T3T_3. Please note that due to Reviewer CGiA’s comment, we have expanded our understanding of SGD’s anticipation, and found that effective successor-gradient alignment (SGA: how well updates align with the next class's gradient) predicts anticipation for SGD, AdamW and Muon (Figure 6), so comparing T2T_2 and T3T_3 is no longer necessary. Hence we decided to remove this figure and the appendix section “Details on measuring T3T_3 interference” with the wrong sign, and updated Figure 6. Our theorems and other results do not depend on it.

Q3: You are right: with random burst lengths, the history tells which class comes next but not when, so the oracle knows slightly more than a history-based predictor could. To make our argument cleaner, we reran this experiment with fixed burst lengths. Now the history determines the class of every batch, so the oracle accurately simulates a sequence model that inferred the next class from its training history. Our conclusions are unchanged (Figure 2): Ring closes about 30% of the gap between Shuffled and the theoretical limit, and 40% of the gap between Shuffled and the oracle.

Thank you again! We are happy to address any remaining questions.

Thank you for running the additional experiments. I find the third- and fourth-order results convincing, particularly the third-order effect across several burst lengths. The fixed-burst experiment also addresses my concern about the oracle comparison, since the training history now determines both when a switch occurs and which class follows. I appreciate the clarification about the sign convention in the original Figure 11.

One question remains for me. To what extent can the higher-order results distinguish a model that tracks the recent sequence of classes from one whose parameters respond differently to different sequences of recent updates? Both accounts seem consistent with positive Successor Advantage. The new experiments strengthen the evidence for higher-order anticipation, but I still see the sequence-model interpretation as an open question. On balance, I will keep my original rating of weak reject.

AuthorsOpenReview ↗

Response to follow-up question

Thank you for your response. We are glad the additional experiments addressed your concerns.

On your question: we see the two accounts as compatible. Responding differently to recent updates could be how the model tracks its training history. What matters is that this response favours the class that comes next, and not simply reflects which classes were recently trained. Our results establish this:

  • Anticipation goes against ordinary forgetting. Models usually remember recently trained classes best. However, in Ring, the successor is the class trained longest ago, yet its held-out loss improves before the class returns.

  • Models use higher-order information. In the third-order setup, under any two of the three preceding classes, the successor is 50–50 (unpredictable). Hence positive SA shows that the model exploits information available only from their combination.

  • Anticipation is not tied to a specific order. We swap the two successor languages in half the seeds, so the same preceding class sequence predicts opposite successors. This rules out fixed language similarities as an explanation for the effect. Randomising the successor removes SA.

We stress that our sequence model claim is behavioural. The models' behaviour reflects not just what it has seen, but what its training history predicts will come next. We do not claim that models explicitly represent recent class identities. Understanding how it stores and uses this information is an interesting area of future work, related to [1].

Thank you again for your engagement with our work. We hope our additional results and clarifications address your remaining concerns, and would greatly appreciate your reconsideration of the overall assessment.

[1] Krasheninnikov, D., Turner, R. E., and Krueger, D. Fresh in memory: Training-order recency is linearly encoded in language model activations. ICLR 2026.

AuthorsOpenReview ↗

Response to follow-up question, part 2

Quick comment on a question similar to yours, in case we misunderstood you above -- whether the model is (a) able to track the order of the previous 3 or 4 classes, or (b) whether it just tracks which 3 or 4 classes came previously but without tracking their order (like a bag of words model disregarding word order).

Our 3rd/4th order experiments are designed s.t. tracking the order is needed for anticipation, thanks to the clockwise / anticlockwise component -- e.g. French -> German -> Arabic implies a different 4th class compared to German -> French -> Arabic).

Reviewer_SGZMOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates anticipatory adaptation under structured training streams. It extends earlier document-level anticipatory recovery to recurring data classes, introduces successor advantage on held-out class examples, and studies first- and second-order transitions. It also derives a local decomposition of sequential updates into variance, gradient-distance, and order-sensitive curvature terms, and uses optimizer comparisons and explicit regularization to investigate a possible mechanism. Finally, it shows that cyclic training order changes relative inter-class gradient alignment.

Critical strengths
  1. The extension from repeatedly revisited documents to held-out examples from recurring distributions is meaningful. Ring, Bursty, and Shuffled provide useful controls, and the main results include substantial replication across seeds.

  2. The optimizer dependence is particularly interesting. AdamW and Muon attain similar held-out mixture losses while exhibiting markedly different successor advantages (Table 6), providing a useful distinction beyond differences in optimization progress.

  3. The second-order experiments strengthen the behavioral finding beyond a simple current-class-to-successor association. The paper also reports informative limitations, including weaker effects at longer context lengths and worse mixture evaluation loss despite improved online training loss.

Critical weaknesses
  1. The local analysis does not establish long-term accumulation. Could signed projections onto ∇SGA or measured local changes better support the proposed mechanism?

  2. Theorem C.1 and Appendix D.5 use opposite signs for T3. Which convention is used in Figure 11?

  3. The negative-λ intervention changes both gradient distance and within-class variance. A variance-reduced estimator or non-successor control would better isolate T2’s role.

  4. Is class identifiability justified for 8-byte samples? Please distinguish the true-class oracle from a history-conditioned predictor facing random burst transitions when interpreting Figure 2.

  5. Effective SGA uses a momentum-free transformation. Does the same conclusion hold using Muon’s full update direction?

  6. Token-position losses or sequence-level advantages would help distinguish dilution by sequence length from a weaker anticipatory mechanism in Figure 14.

  7. Section 6 measures local gradient relationships. What supports interpreting these changes as persistent representational bias rather than transient optimization geometry?

Questions for authors

see weakneses

4 comments
AuthorsOpenReview ↗

Response to review

Thank you for your review! We are glad you found our problem setting meaningful and our analysis on optimizer dependence interesting, and that you liked our second-order experiments. All new experiments and results are added to our updated manuscript, and are marked in blue.

1. Accumulation. We added new empirical analysis (Figure 19 and Appendix Section D.8). We measure how much a single burst changes the loss of each class. Adding up these small changes over a cycle reproduces the successor advantage (SA, our measure of anticipation) observed late in training, also on held-out seeds. This empirically supports our accumulation hypothesis.

2. Sign of T3T_3. Theorem C.1’s formula was correct, but we implemented the one in the old D.5, so the signs indeed flip on the old appendix Figure 11, which plotted the cosine similarity between T2T_2 and T3T_3. Please note that due to Reviewer CGiA’s comment, we have expanded our understanding of SGD’s anticipation, and found that effective successor-gradient alignment (SGA: how well updates align with the next class) predicts anticipation for SGD, AdamW and Muon (Figure 6), so comparing T2T_2 and T3T_3 is no longer necessary. Hence we decided to remove this figure and the appendix section “Details on measuring T3T_3 interference” with the wrong sign, and updated Figure 6. Our theorems and other results do not depend on it.

3. Regularization. As suggested, we have added a new, variance-reduced EMA regularizer for negative λ regularization, and find that anticipation goes away (Figure 7). So we can now confirm that the disappearance of anticipation after negative-λ regularization was not due to noise. We also demonstrate stronger control on anticipation: with positive λ, if we regularize class c towards c+5 (and not c+1), some of the anticipation transfers to class c+5 (Figure 18).

4.1 Class identifiability: we believe this is a reasonable assumption for most samples. Examples from our evaluation set include: "that is" (English), "il fait" (French), "Es ist" (German), "de agua" (Spanish).

4.2 True-class oracle vs a history-conditioned predictor: To make our argument cleaner, we reran the Fig 2 experiment with fixed burst lengths, so the oracle now accurately simulates a sequence model that inferred the next class from its training history. Our conclusions are unchanged.

5. Muon. Muon’s full update direction is nonlinear in the gradient due to the Newton-Schulz steps. We therefore do not claim that our theory holds for Muon. However, our experiments find that effective SGA remains predictive of anticipation for Muon too (Figure 6).

6. Token positions. We reran the context-length 8 and 128 settings of Figure 21 (old 14) with held-out loss logged at every token position, and report SA separately for the first 8 tokens and for tokens 9–128 of the context length 128 model (Figure 22). As expected, in the Bursty control, SA is close to zero in every token group throughout training.

The results indicate that both effects are present:

  • Dilution: At the end of training, tokens 9–128 of the length-128 model carry little anticipation per token.
  • Weaker anticipation at matched positions: On the first 8 tokens, the length-128 model ends at about half the SA of the length-8 model (0.15 against 0.31 nats per token).

At the sequence level (when you add up token-level SA instead of averaging as in Figure 21), anticipation is larger for length 128 than for length 8 throughout training (7.3 against 2.5 nats per sequence over the last 20k steps). This means that the lower per-token SA in Figure 21 reflects how the advantage is spread over token positions, together with a decline at late positions as training progresses. This is consistent with our explanation in Section 5: once a long prefix identifies the class, the model no longer needs to rely on the training order, so anticipation is most prominent on the early tokens.

We also note that per-token anticipation is possible for long context lengths. We have also rerun our main experiment (see Figure 8) for context lengths of up to 1024, and found that anticipation is preserved, but requires longer bursts (e.g. 128).

7. Interpreting Section 6. Our results support both interpretations (persistent representational bias and optimization geometry), and they are not in conflict with each other. For the former, Fig 30 shows that different Ring orders of the same data result in different final train and held-out losses, so the order affects the finished model. For the latter, Fig 11 shows that the weights cycle with the training order.

We also added third- and fourth-order experiments: models anticipate the successor class even when it depends on the last three or four bursts together (Figures 4, 13).

Thank you again! We are happy to address any remaining questions.

Thanks for the response. My main concern is that the paper focuses on LLMs, but the largest model tested is only 52M parameters. It remains unclear whether the same effects occur in larger models and what these results tell us about training LLMs. Do the same effects occur in larger models, and what can these results tell us about training LLMs?

AuthorsOpenReview ↗

Response to follow-up questions

Thank you for your continued interest in our work. Here are our responses.

Do the effects occur in larger models? Yes. Anticipatory recovery was first shown by [1] when fine-tuning pre-trained Pythia models from 160M to 2.8B parameters, and they found that larger models show a stronger effect. Our contribution goes in the other direction. We show that anticipation is already present in pre-training from scratch, in models as small as 0.19M parameters, and on held-out samples of a class rather than repeated documents. In our own model-size ablation (Figure 24), anticipation also grows with model size.

We used small models for two reasons: (1) [1] described anticipatory recovery as emerging only at scale, and we wanted to test whether it needs scale (it does not), (2) our experiments need many runs (12 seeds per setting across optimizers, training orders and ablations, with gradient statistics logged during training), which was not feasible at larger scale with our compute. Our theory does not depend on model size.

What does this tell us about training LLMs? We do not propose a new training recipe. Our paper explains a surprising, previously unexplained phenomenon in the literature, both theoretically and empirically. There are two areas where our findings are particularly relevant.

  • AI safety. Models can learn information that is in no individual training example. The next class is only predictable from the order of training, yet, surprisingly, the model learns to use it. This adds to earlier findings that language models pick up information about their own training process [2, 3].
  • Continual learning and curriculum learning. In these areas, data (or domains) are not perfectly shuffled. We showed that different orderings of the same data lead to different final training and held-out losses (Figure 30).

Thank you again for your engagement with our work. We hope our additional results and clarifications address your remaining concerns, and would greatly appreciate your reconsideration of the overall assessment.

[1] Yang, Y., Jones, M., Mozer, M. C., and Ren, M. Reawakening knowledge: Anticipatory recovery from catastrophic interference via structured training. NeurIPS 2024.

[2] Krasheninnikov, D., Krasheninnikov, E., Mlodozeniec, B., Maharaj, T., and Krueger, D. Implicit meta-learning may lead language models to trust more reliable sources. ICML 2024.

[3] Krasheninnikov, D., Turner, R. E., and Krueger, D. Fresh in memory: Training-order recency is linearly encoded in language model activations. ICLR 2026.

Thank you for your detailed responses and clarifications. I appreciate the authors' efforts in addressing my concerns, and I have raised my rating to 3 with Borderline.

However, I will keep my confidence at 1, as I remain uncertain about the implications of these findings for LLM training. In particular, the largest model evaluated in this paper has only 52M parameters, which may be insufficient to establish whether the findings generalize to larger-scale LLMs. While the cited prior work provides evidence at larger scales, the proposed analysis and findings have not been directly validated on larger models in this paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to all reviewers

We thank all reviewers for their thoughtful comments. Our revision adds new experiments that address your concerns:

  1. SGD. With a tuned learning rate, SGD anticipates too (Table 1, Figure 15). This addresses reviewer CGiA's concern that our theory is for SGD, but SGD did not anticipate.
  2. Third and fourth order. Models anticipate the successor class even when it depends on the last three or four bursts together (Figures 4, 13). This addresses reviewer HTsg's concern that our second-order results are limited in scope.
  3. Longer contexts. We observe anticipation on samples of up to 1,024 tokens, given longer bursts (Figure 8). This addresses the concern of reviewers CGiA and SGZM that anticipation fades with context length.
  4. Accumulation. In a new experiment, we measure the local effect of ring training (which we explain with our implicit bias analysis): how much each burst changes the loss of every class. Late in training, these small changes add up to the anticipation we observe (Figure 19). This addresses reviewer SGZM's question on whether the local bias accumulates.
  5. Controlling anticipation. In two new experiments, a less noisy version of our regularizer can both remove anticipation (Figure 7) and redirect part of it to a different class (Figure 18). This addresses reviewer SGZM's concern that noise in the regularizer drives the effect.

Smaller additions are described in the individual replies. Changes to the PDF are in blue, and figure numbers refer to the revised PDF. One appendix figure (the old Figure 11) is removed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.