acceptodds
Under review as a conference paper at ICLR 2027

The Clocks of Composition: Built by the Signal, Lost to Weight Decay, Kept by Rehearsal

Abstract

A network can learn a skill it will not keep. In a looped network, a signal on the intermediate state builds two-hop composition over facts seen only alone; when the signal stops, with the training data unchanged, the composition fades, and plain training does not build it (OOD 0.42). We measure the three clocks that govern its life. It is built on the signal's clock: at dense signal rates the time to compose barely depends on weight decay. It is lost on the weight-decay clock: the time to lose it scales as across a tenfold range of learning rates and a 33-fold range of weight decays, and without weight decay no seed loses it. It is kept only by rehearsal above a threshold; with a short second-moment memory, which removes the damage that rare rehearsal steps do under Adam's default, the threshold rises as . The law travels: registered in advance, it predicts the rehearsal threshold of a transformer's fact memory, a different architecture, task and quantity, with no parameter fitted (exponent 2.3). In a transformer with chain-of-thought, composition fails at high weight decay because no training row exercises the out-of-distribution facts where chain-of-thought reads them; presenting every fact once in that position, with no knowledge of the split, restores it from 0.38 to 0.95.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
weight decay, AdamW, grokking, compositional generalization, forgetting, rehearsal, continual learning, training dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_K2qPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how neural networks learn to combine known facts, why they can lose this ability during training, and whether occasional practice can preserve it. It measures how learning and forgetting depend on supervision of intermediate steps, weight decay, and optimizer settings. Experiments use small neural networks and transformers to study fact retention and two-step reasoning, including whether practicing individual facts in the format needed for the second reasoning step improves performance.

Critical strengths

I think the controlled forgetting experiments in the small looped network is the paper’s strongest contribution. By varying learning rate and weight decay separately, the authors find that the time to lose compositional ability is approximately inversely proportional to their product. This goes beyond showing that forgetting occurs and establishes a measurable relationship between training settings and the persistence of compositional ability.

Critical weaknesses
  1. The connection between the experiments is insufficiently established. The toy study maintains composition through an intermediate-state penalty, whereas the transformer studies examine fact rehearsal and additional examples that practice the second reasoning step. The paper does not convincingly show how the toy measurements explain the transformer composition results or guide these interventions. This limits the broader significance of the proposed common explanation and leaves its practical implications unclear.

  2. The evidence for transferring the rehearsal scaling relationship to transformers is fragile. The fact-retention experiment uses only two weight-decay values and two seeds per condition, with substantial training instability. Appendix A acknowledges that using the last rather than the first crossing of the retention threshold changes the reported exponent from 2.34 to 0.7. The main scaling plots also lack uncertainty estimates across seeds, and no confidence intervals are provided for the fitted exponents. These limitations make the claimed transfer difficult to assess and insufficiently supported.

Questions for authors

Clarification questions:

  1. What do the toy scaling results explain or predict about the transformer composition experiment?
  2. Is there a reason why the first 50% accuracy crossing used to define the rehearsal threshold, given that using the last crossing changes the transformer exponent from 2.34 to 0.7?
  3. Can you report per-seed rehearsal thresholds and uncertainty in the fitted exponents, particularly for the transformer fact-retention experiment?
3 comments
AuthorsOpenReview ↗

Response to Reviewer K2qP

Thank you for naming the controlled forgetting result, a time to lose composition proportional to 1/(lr × wd), as the paper's strongest contribution. The revision is built around it, now measured in transformers at 58M and 0.3B.

W1 / Q. The toy and transformer studies are not connected: what do the toy's scaling results predict about the transformer? Section 3 now states the toy's single role. In the looped network, facts-only training still trains everything composition needs, so there is no drift and only decay can act. It therefore tests the account's two predictions directly: (a) nothing is lost without weight decay (12 of 12 seeds keep composition at 1.00), and (b) the time to lose composition scales as 1/(ηλ), measured as (ηλ)^1.04 (CI [0.97, 1.09]). In the 58M transformer, (b) carries over to the part of the loss that weight decay adds, (ηλ)^1.24 (CI [1.04, 1.57]) (Section 5.1, Figure 4). (a) fails with learned positions, where the transformer loses composition without decay to drift, and largely holds with rotary positions, which remove most of the drift (0.87 and 1.00 after 60k facts-only steps). What the toy does not predict, the dose at which a short repair stops working, is measured in the transformer (Section 5). The account also guides the two interventions in Section 6: weight decay erases composition because it pulls every weight toward zero, including those the loss no longer uses, so decaying toward the starting weights instead keeps it; and the account says which rows must be replayed. The claim that the toy's rehearsal law predicts the transformer's is withdrawn.

W2 / Q. Why the first 50% crossing, given the instability? It was the rule we used, and on a non-monotone curve it is not a safe one: the last crossing gives 0.7, and Appendix C reports both. The instability is real in the follow-up too (101 of 120 runs have a transient spike, so the registered rule that excludes every spiking run is inconclusive). The registered follow-up removes the choice instead of defending it: each run is read in an endpoint window that excludes transient spikes (fixed before any rehearsed run existed), and the registered cycle-average endpoint gives the same exponent.

Per-seed thresholds and uncertainty in the fitted exponents. All are in Appendix D (Table 10). Per-seed p50, four seeds: λ = 0.3: 5.42, 5.57, 5.63, 5.54 × 10^-4; λ = 0.5: 5.53, 5.61, 5.56, 5.59 × 10^-4; λ = 1: 1.60, 1.49, 1.56, 1.69 × 10^-3; at λ = 2 and 3 the facts were kept at every rate tested. The exponent is 0.91 (95% CI [0.86, 0.96]; per-seed slopes 0.85-0.97), outside the predicted [1.4, 2.8]. The interval covers seed noise only, and we report 0.91 as the registered fit that rejected the transfer, not as a law. The other two exponents are 1.04 [0.97, 1.09] (looped network) and 1.24 [1.04, 1.57] (transformer decay part; the loosest, as seven of its eleven cells have two bases). For the new 0.3B results we give three repair seeds per state and a never-composed control per base; one of the two pairs is clean, and we report it as one pair (Sections 5 and 8).

The submission's exponent of 2.34 is withdrawn, and nothing in the revision rests on it. If the revision gives what the submission lacked, one clear result with a transformer at its centre and uncertainty stated where it matters, we would be grateful if you would reconsider your score.

I appreciate the authors' response. The changes address some of my concerns.

The authors withdrew the toy model’s rehearsal law prediction for the transformer. But this leaves a disconnection in the paper regarding the purpose of the toy model and the practical implications of its lessons. The remaining connection is mainly about the additional forgetting with the added weight-decay. The toy model looks like a specific example where weight decay can destroy composition, while the explanations for the transformer results (e.g., the positional encoding explanation, repair without weight decay) remain separate from the toy. This weakens the coherence of the paper.

I would expect the toy model to first capture the practical phenomenon, including the forgetting observed without weight decay, and to derive lessons using the simplicity of the toy model. Then the lessons learned in the simplified setting should lead to predictions about how the practical setting should behave, which the authors can test. Otherwise, the toy model’s explanatory value remains unclear. I suggest the authors develop the toy model further for this purpose.

Given these remaining concerns, I’m keeping my score unchanged.

Thank you; we agree the toy should carry the argument, and the revision now makes it do so (Section 3). The toy captures the phenomenon, including forgetting without weight decay. In the looped network the account's forces can be seen alone. Under AdamW (decay alone) it keeps composition without weight decay and loses it at a rate proportional to (ηλ)^1.04. Under Lion, the sign step alone more than halves composition within 100 steps without weight decay and removes it within about 800, its facts at 1.00. In a new registered test with a never-composed control trained the same way: while it is partly lost, 100 steps after the signal stops, a short repair adds 0.45–0.55 more composition than it builds in the control; once lost, from 2,000 steps on, at most 0.07 more (4 of 4 seeds; Appendix C, OPT-R). Lessons, and their tests in the transformer. Decay acts on ηλ: in the 58M transformer the part of the loss that weight decay adds follows (ηλ)^1.24 (Section 5.1). Read as a dose, this places erasure before a run: for a learning rate and weight decay never used, we registered the 58M crossing at a dose of 1 within a factor of two (0.80 in the integrated dose, window 0.30–1.80); it landed at 0.72 and 0.84 (Section 5). The toy suggests why the position encoding matters. Under AdamW without weight decay, composition survived facts-only training in our runs where the facts loss exercises the computation it needs. In the toy the second hop is the atomic step applied twice, so it is kept, even with the two-hop rows dropped (Appendix C, U1). In the 58M transformer the second hop reads facts at positions whose learned embeddings no single-fact row trains: there composition is lost without weight decay yet a short repair restores it (0.91–0.99), and with rotary positions, which carry no position-specific weights, it is kept (two bases; Section 5.1). In our from-scratch runs, every run that erased moved every weight (decay, or a sign step); drift under AdamW at most hid. We believe this resolves the coherence concern you raised; if it does, we would be grateful if you would reconsider your score.

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

The paper studies how two-hop composition is learned and retained in a looped network. After installing the skill through intermediate supervision, the authors measure its decay and the rehearsal needed to maintain it. They also test rehearsal scaling in a transformer's fact memory and improve OOD composition by training facts in the layout used by chain-of-thought.

Critical strengths
  1. Removing intermediate supervision while keeping the original task data gives a useful test of skill retention. The learning-rate and weight-decay sweeps, including the no-decay control, make the toy results fairly clear (§4, Table 1).

  2. The layout experiment separates fact recall from composition: ordinary repetition improves recall, but position-matched rows raise second-hop OOD accuracy from 0.38 to 0.95 (§7, Table 5). The failed predictions and unstable runs are also reported openly in Appendices A–C.

Critical weaknesses
  1. The added contribution of the three-clock account is unclear. The loss scale is already given by the AdamW EMA analysis; an exponent of −1.04 is consistent with it. Measuring this at the behavioral level is useful, but it does not by itself establish a new composition-specific law. The layout intervention makes a different point about accessing facts in the right context. As §7 acknowledges, it is close to identity-bridge supervision and upper-layer representation training, although the methods differ. It does not directly test the proposed acquisition, loss, or rehearsal scaling in that Transformer. I would like a clearer statement of what the clock account predicts beyond these existing explanations.

  2. The claimed transfer of the rehearsal law depends heavily on how the threshold is read. Table 4 has two weight-decay levels and two seeds per configuration. At λ=0.3, accuracy crosses 0.5 more than once as rehearsal increases. Appendix A notes that using the last crossing changes the exponent from 2.34 to about 0.7. This makes the reported agreement difficult to interpret. Figure 4 normalizes each system by its own threshold at λ=1, so it supports a comparison of relative scaling rather than an absolute-rate prediction. The one-coordinate explanation also fails its held-out test against a simpler null (Appendix B). These results leave the generality of the keep clock unresolved.

Questions for authors
  1. What does the three-clock account predict beyond the known AdamW timescale and bridge/context alignment? Which part of that prediction does the layout experiment test?

  2. How is p50 chosen for a non-monotone rehearsal curve? Please explain the 2.34 versus 0.7 sensitivity and the uncertainty supported by the existing seeds and threshold interpolation. Does the transfer claim concern relative scaling or absolute rehearsal rates?

3 comments
AuthorsOpenReview ↗

Response to Reviewer GsbE (1/2): what the account adds

Thank you for the close reading, and for crediting the retention test, the controls and the open reporting of failures. Your questions shaped the revision.

W1 / Q1. What does the account add beyond the AdamW timescale (the exponent of 1.04 is consistent with it) and bridge/context alignment? We agree the timescale is not ours. Section 3 builds on it explicitly (Wang & Aitchison, 2024) and treats the account as a baseline, not a contribution. The timescale predicts a rate of loss; it does not say whether what is lost can be brought back by a short repair, or until when. The revision measures what it does not give:

  • Recall and composition separate. The facts stay recalled at 1.00 while composition falls to zero (Section 4).
  • Hidden, then erased, as the dose grows (58M). The short repair fails by a dose of 1.5 in every setting where the facts were still recalled. For an unseen setting we registered the crossing at a dose of 0.80 (window 0.30-1.80) before running it; it crossed at 0.72 and 0.84 (Section 5, Figure 3).
  • A bound at a second size, against a history-matched control (0.3B). After a pilot on the same base, we registered that the short repair (last 6 of 24 blocks) would be erased by a dose of 6, judged against a never-composed model with the same training history. In the one clean pair it was erased at λ = 1 by a dose of 6 without gradient clipping and 4.5 with it, and at λ = 2 by 2.86 (Table 3). The registered "in both pairs" was met only with the pair we then set apart after a collapse of its facts. What the account does not deliver at 0.3B: equal doses do not give equal repair (gaps up to 0.27), so the dose bounds erasure but is not shown to time it, and a hidden state before erasure is not established (Sections 5 and 8). The 58M crossing remains our only held-out prediction of a dose.
  • Without decay. In our 300k-step bases nothing is erased without weight decay at either size: at 58M composition is hidden (learned positions; repair 0.91-0.99) or kept (rotary; 0.87, 1.00); at 0.3B it is never lost (0.94-1.00 through 61k steps). In pretrained Qwen2.5-1.5B the loss without decay is hidden in both fact graphs (repair 0.83-0.88 against 0.41-0.54 for a never-composed control).

On bridge/context alignment: we read the loss without decay that way too. Retraining only the last block repairs it, in line with forgetting as a misaligned readout over features that survive (Section 5). What weight decay adds is that, past a dose, the same repair no longer works.

AuthorsOpenReview ↗

Response to Reviewer GsbE (2/2): the layout experiment, thresholds and uncertainty

W1 / Q1. Which part does the layout experiment test? It is close to identity-bridge supervision and upper-layer training, and does not test acquisition, loss or rehearsal scaling in that transformer. We agree. It tests prevention, not the account, so the main text keeps one pointer sentence (Section 6) and the experiment is in Appendix B (Table 4), with both neighbours stated. One extra row per fact, placed where chain-of-thought reads the second hop, lifts composition over facts that no two-hop row uses from 0.38 to 0.95 (5 seeds); repeating the same facts 13 or 26 times reaches only 0.56-0.66, and an identity objective on bridge entities (Lin et al., 2025) gives 0.34-0.37. Training upper layers on held-out atomic facts (Zhang et al., 2026, Section 5.1) is close in spirit; what we add is why repetition does not substitute. The revision no longer claims that this experiment tests acquisition, loss or rehearsal scaling.

W2 / Q2. The rehearsal-law transfer depends on the threshold. We withdraw it, in the relative reading (each system normalised at λ = 1, as in the submission's Figure 4) and in the absolute one. The registered fork design (five weight decays, four fresh seeds, 120 runs) gives 0.91 (95% CI [0.86, 0.96]), outside the predicted [1.4, 2.8] (Appendices A and C).

  • How p50 is chosen on a non-monotone curve. The fork design reads each run in an endpoint window that excludes transient spikes, fixed by amendment after the runs without rehearsal were read and before any rehearsed run existed; the registered cycle-average endpoint gives 0.91 too (Appendix C).
  • Uncertainty from seeds and interpolation. Bootstrap 95% intervals for every fitted exponent and all per-seed thresholds are in Appendix D (per-seed slopes 0.85-0.97). The interval covers seed noise only: each threshold sits between adjacent grid rates, and the fit has three points with a flat-then-rising shape. We report 0.91 as the registered fit that rejected the transfer prediction, not as a law.
  • The one-coordinate explanation fails its held-out test against a simpler null. It still does, and the revision reports it as a rejected prediction (10 of 19 conditions within 2×; the null is closer). No claim rests on it (Appendices C and H).
  • Uncertainty in the new results. At 0.3B the unit of replication is a pair (a composing base and a never-composed control with the same history), with three repair seeds per state. The spread over repair seeds of the 2k-step repair is at most 0.07, so the uncertainty that matters is between pairs, and only one pair is clean; we say so in Sections 5 and 8 rather than pool the two.

Beyond our synthetic setting. We have registered and are running two tests (general response, part 3): the paper's setting on real data in two pretrained families, and whether erasure at 1.5B follows the dose or the number of steps. We will post their registered readings before the discussion closes, whatever they show.

If the revision answers these questions, we would be grateful if you would reconsider your score.

Thank you again for your review. On your question of what the account adds beyond the AdamW timescale: the timescale says how fast an unused weight shrinks; the account says which of three outcomes follows, and when. A dose at which composition is erased, placed before a run: for a learning rate and weight decay never used, we registered the 58M crossing at a dose of 1 within a factor of two; it landed at 0.72 and 0.84 (Section 5). What is lost without weight decay is hidden, not erased: a short repair restores it far above a never-composed control, at 58M (0.91–0.99) and now in a pretrained Llama-3.2-1B taught facts in natural language (0.40–0.54 above the control, 2 of 2 fact worlds; Section 6). A fix that follows from it: decaying toward the starting weights instead of toward zero keeps composition in models trained from scratch and leaves it repairable in a pretrained 0.5B model (Section 7). On the threshold questions: the registered fork design reads each seed's threshold in a window that excludes transient spikes (Appendix C), and every fitted exponent has a bootstrap 95% interval with per-seed thresholds (Appendix D). If these answer your questions, we would be grateful if you would reconsider your score.

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

The paper studies how auxiliary supervision, weight decay, and rehearsal affect the learning and retention of two-hop compositional reasoning, primarily in synthetic networks. It reports empirical relationships governing how quickly this ability is acquired, forgotten, and maintained. However, I found the presentation very difficult to follow: the underlying problem and intended contribution are obscured by vague terminology and repeated metaphors, making the significance of the results difficult to assess.

Critical strengths

please see above.

Critical weaknesses

See above in addition to these points:

Poor exposition. The abstract and introduction repeatedly invoke “clocks,” “signals,” and “exercising” facts before clearly explaining the task, supervision, and evaluation. The reader has to reconstruct the basic setup from later sections.

Unclear contribution. The introduction motivates a training algorithm that learns composition early, whereas Section 9 states that the paper does not propose a method. The distinction between the broader motivation and the actual empirical contribution needs to be much clearer.

Insufficient support for a central claim. The claimed transfer of the rehearsal scaling law to transformers rests on only two weight-decay values and two seeds. Appendix A acknowledges that choosing a different threshold crossing changes the reported exponent from 2.34 to 0.7, substantially weakening this claim.

Limited scope. Much of the evidence comes from a small synthetic setting with access to ground-truth intermediate supervision. The implications for more general compositional reasoning remain unclear.

Questions for authors

see my comments above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer dPbc

Thank you. Your review named the two problems that mattered most, and both changed the paper.

1. "The underlying problem and intended contribution are obscured by vague terminology and repeated metaphors"; "clocks", "signals" and "exercising" appear before the task, supervision and evaluation. You were right. The introduction now introduces the task in its first paragraph: a transformer trained from scratch to answer two-hop questions over a knowledge graph, with chain-of-thought, then trained further on the single facts alone, with its two measurements, recall (accuracy on the facts) and composition (accuracy on held-out two-hop questions). The second paragraph defines the short repair and the three outcomes (kept, hidden, erased), once each. The metaphors are gone from the main text; working labels left in the appendices are defined in Appendix A (Sections 1-2, Table 1).

2. The introduction motivates a training algorithm, but the paper proposes no method. The contribution is now one stated result: which of kept, hidden and erased a facts-only update leaves, by weight decay, and the weight-decay dose that bounds when lost composition is erased (Section 1, Table 1). Two ways to prevent the loss follow from it (Section 6). Decaying toward the starting weights instead of toward zero keeps composition through 60k facts-only steps at λ = 1 (58M: 1.00 and 1.00 with rotary positions, 1.00 and 0.97 with learned positions; 0.3B: 1.00 in both bases through 61k steps), where ordinary decay loses it within 5k-7k steps. Keeping a little two-hop data in the update also prevents the loss (58M, pretrained 0.5B). We state the limits: at 0.3B composition is also kept without any decay, and decay toward the start does not help the pretrained 0.5B model (Section 6).

3. The transfer of the rehearsal law rests on two weight decays and two seeds; a different threshold crossing changes the exponent from 2.34 to 0.7. We agree, and we withdraw the claim and everything built on it. A registered follow-up (saved models forked into five weight decays, four fresh seeds, 120 runs) gives 0.91 (95% CI [0.86, 0.96]), outside the predicted [1.4, 2.8] (Appendices A, C and D). Nothing in the revision rests on the transfer.

4. Small synthetic setting with ground-truth intermediate supervision. The toy's intermediate-state supervision is no longer part of any claim. The main results are in transformers trained from scratch on the standard two-hop task of Wang et al. (2024) and updated on facts alone, at 58M and now at 0.3B (Sections 2 and 5). We added pretrained Qwen2.5-0.5B and 1.5B models taught invented facts, where facts-only training also removes the trained two-hop format while recall stays high; at 1.5B the loss is hidden without decay in both fact graphs (Sections 4 and 5).

A non-synthetic test, and what it shows. We registered and ran a test on a real task pair (Appendix C): Llama-3.2-1B fine-tuned on grade-school maths (MetaMathQA; 0.38 on 400 GSM8K test items, against 0.01 for a control trained for the same steps on maths-free instructions), then on code alone (Magicoder-OSS-Instruct, 6,000 steps), with the paper's short repair (last 3 of 16 layers, 250 steps, three seeds). Without weight decay the maths is lost (0.015-0.06 from 1,500 code steps on), and the repair restores only 0.08-0.35 of the advantage over the control, below the 0.5 we registered for "hidden". The repair brings back the trained answer format, not the accuracy. We checked the instrument with tests registered before they ran:

  • A longer repair: four times the repair steps gives 0.29, against 0.20 for the registered repair on the same rerun. A repair of all 16 layers is not a usable probe at our learning rate, because it lowers the unforgotten model itself from 0.38 to 0.10.
  • Replay: with a tenth of each batch from the original model, as in our 1.5B runs, the maths is lost just the same and the repair restores 0.07-0.21.

We report this as a limit of scope: the paper's hidden results hold for composition over facts the model still recalls (Section 8). It also shows that the probe does not report "hidden" for any loss: here, against a never-trained control, it reports "not hidden". Open: one model, one seed, and no full-depth repair at a lower learning rate. Weight decay toward zero at the rate we used destroyed both models before the first read, so erasure on this task is not measured.

The paper's own setting on real data. We have registered, and are running, new facts in natural language composed into two-hop answers in Llama-3.2-1B and Qwen2.5-1.5B, followed by a code-instruction stage without decay, with decay and with decay toward the starting weights (general response, part 3). We will post its registered readings before the discussion closes, whatever they show.

If the revision resolves your concerns, we would be grateful if you would reconsider your score.

Thank you again for your review. Since our response, the revision adds the test that answers your scope concern directly (point 4: a small synthetic setting). In a pretrained language model, Llama-3.2-1B, taught new facts in natural language about invented people and then trained on code instructions, composition is lost while the facts stay recalled, and a short repair restores it 0.40–0.54 above a never-composed control given the same steps (2 of 2 fact worlds; registered before its runs; Section 6, Figure 5). Together with the pretrained Qwen2.5-0.5B and 1.5B records, the loss of composition with recall kept now appears from scratch and in pretrained models taught facts in natural language, where it is hidden without weight decay. On your points 1–3: the paper is rewritten around this one result, defines the task and its two measurements before anything else, uses no metaphors, and withdraws the rehearsal-law transfer claim. If this answers your concerns, we would be grateful if you would reconsider your score.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Summary of the revision (1/3): what changed, and where

We thank all three reviewers. The reviews raise three points, and we think all three are right: the submission was hard to follow (dPbc, K2qP); its central transfer claim (the rehearsal exponent of 2.34) was not supported (all three); and the looped toy network did not clearly explain the transformer results (K2qP).

The revision answers each point. It withdraws the transfer claim, removes the metaphors, and is rebuilt around the result the reviews singled out as strong, now measured in transformers: when a transformer is trained further on facts alone, it keeps the facts but can lose the ability to compose them, and weight decay decides whether that composition is kept, hidden (a short repair restores it) or erased (it does not); under weight decay, the dose of the update bounds when it is erased. The main text is 10 pages; Appendix A lists every change since submission.

ConcernWhat we changedWhere
Hard to follow: "clocks", "signals" and "exercising" before the task, supervision and evaluation (dPbc; presentation 1 from dPbc and K2qP)The introduction defines the task and its two measurements (recall of the facts, composition on held-out two-hop questions) first, then kept, hidden and erased, once each. The metaphors are gone from the main textAbstract, Sections 1-2, Table 1
Unclear contribution; what beyond the AdamW timescale? (dPbc, GsbE)One stated contribution: which of kept, hidden and erased a facts-only update leaves, by weight decay, at 58M and 0.3B, and the weight-decay dose that bounds erasure. The timescale gives a rate of loss; it does not say whether a short repair can bring the skill back, or until whenSection 1, Table 1, Figure 1, Section 5
The paper proposes no method (dPbc)Two ways to prevent the loss: keep a little two-hop data in the update, or decay toward the starting weights instead of toward zeroSection 6
Rehearsal-law transfer rested on two weight decays, two seeds and the first of several crossings; the last crossing gives 0.7 (all three reviewers)Withdrawn. A registered fork design (five weight decays, four fresh seeds, 120 runs) gives 0.91 (95% CI [0.86, 0.96]), outside the predicted [1.4, 2.8]Appendices A and C
No uncertainty on exponents; per-seed thresholds (GsbE, K2qP)Bootstrap 95% intervals for every fitted exponent, and all per-seed thresholdsAppendix D, Table 10
Toy and transformer not connected (K2qP)The toy keeps one stated role: it has no drift, so it tests the decay account directly ((ηλ)^1.04, CI [0.97, 1.09]); the transformer's decay part follows the same variable ((ηλ)^1.24, CI [1.04, 1.57])Sections 3 and 5.1
Layout experiment is close to identity-bridge and upper-layer training and does not test the account (GsbE)Agreed: it tests prevention. Moved to the appendix with both comparisons; one pointer sentence remainsSection 6, Appendix B (Table 4)
One-coordinate model fails its held-out test against a simpler null (GsbE)Reported as a rejected prediction; no claim rests on itAppendices C and H
Small synthetic setting (dPbc)Added a 0.3B transformer trained from scratch, pretrained Qwen2.5-0.5B and 1.5B models, and a registered real-task test (Llama-3.2-1B, maths then code) that bounds the scopeSections 4, 5 and 8; Appendix C
AuthorsOpenReview ↗

Summary of the revision (2/3): continuity with the submission, and the result

Continuity with the submission. The question and the instrument are the same: what does a network keep of a computation its training loss stops using, and on what timescale? The submission answered it mainly in a looped network and claimed that its rehearsal law carries over to a transformer; that claim failed its registered test, and we withdraw it. Asking the same question where models are actually updated, continued training on facts, gave a narrower and better-supported answer. The title now names that answer ("Kept, Hidden, Erased: Weight Decay Bounds How Long Lost Composition Can Be Repaired"), because the submitted title's three clocks no longer describe the paper. The looped network's other results are in Appendices E and F. Because the framing changed so much, we would be grateful if you read the revision as a new paper rather than as a diff.

The result (Table 1). "Erased" means "not repaired within the stated repair budget" throughout.

  • 58M, from scratch. Without weight decay, lost composition stays hidden (learned positions; repair 0.91-0.99) or is kept (rotary positions; 0.87, 1.00). With weight decay it is erased by a dose of 1.5 in every setting where the facts were still recalled. For a setting never used before, we registered the crossing at a dose of 0.80 (window 0.30-1.80); it landed at 0.72 and 0.84 (Figure 3).
  • 0.3B, from scratch, judged against a never-composed control with the same history. Without decay, composition is kept (0.94-1.00 through 61k steps). With decay, in the one clean pair, it is erased by a dose of 4.5 with gradient clipping (6 without) and 2.86 at twice the decay rate. At this size the dose bounds erasure but is not shown to time it, and a hidden state before erasure is not established (Section 5, Table 3). The two sizes also differ in repair depth, erasure criterion and, at λ = 0, position encoding, so we do not attribute their difference to size.
  • Pretrained Qwen2.5-0.5B and 1.5B, taught invented facts and their composition. Facts-only training removes the trained two-hop format while recall stays high (0.5B: recall 0.96-0.99, composition 0.98 to 0.01-0.02). At 1.5B, registered before any continuation, the loss is hidden without decay in both fact graphs (repair 0.83-0.88 against 0.41-0.54 for a never-composed control). With decay, erasure was registered for both graphs and met in one. A registered rerun of the erased graph on the other graph's GPU type and code path again loses its advantage over the control from a dose of 3, but its own repair partly recovers (0.45-0.49), so by the rule fixed in advance the split is attributed to neither the graph nor the path (Section 5, Table 5, Appendix C). We report the pretrained results as records on one model each.

A real task pair, as a limit of scope. In Llama-3.2-1B fine-tuned on grade-school maths and then on code alone (registered before any run), forgetting without weight decay was not hidden: the short repair restored only 0.08-0.35 of the maths advantage over a never-maths control, below the registered 0.5, and neither a four times longer repair nor replay changed the verdict. The paper's hidden results are therefore stated for composition over facts the model still recalls (Section 8, Appendix C). The test also shows that our repair probe does not return "hidden" by construction: it can return "not hidden", and here it does. Details are in our reply to Reviewer dPbc.

AuthorsOpenReview ↗

Summary of the revision (3/3): registered runs in progress, and the camera-ready

Registered runs in progress; readings posted before the discussion closes. To test the paper's own setting on real data, we have registered two experiments, with designs and readings fixed before any run, and they are running now:

  1. New-fact composition in pretrained models, with a realistic second stage. Pretrained Llama-3.2-1B and Qwen2.5-1.5B are taught new facts in natural language about invented people (employer, birth city, university, and relations to other invented people) and how to compose them into two-hop chain-of-thought answers. A second stage then trains them on code-instruction data, with the facts still in the training stream and no two-hop data, in three arms: no weight decay, decay toward zero, and decay toward the starting weights. Each family has several independent fact sets (up to six for Llama, two or three for Qwen, as many as land). Each state is read with the paper's short repair (last 3 layers, 250 steps, here on two-hop chains neither model saw in the first stage) against a never-composed control given the same steps.
  2. Dose or steps, at 1.5B. At a fixed learning rate, the decay rate is set to λ = 1.5 and to λ = 6 (four times higher), with states saved at matched doses, in two fact graphs. The test asks whether the repair stops working at the same dose at both rates, or whether it follows the number of steps instead.

Before the discussion closes (Nov 18, AoE), we will post the registered readings on every run that has landed by then, whatever they show, labelled partial where runs are still going.

If accepted, the camera-ready will include the complete registered real-data results (all fact sets, both model families, and the decay-toward-start arm), the dose-or-steps test on both fact graphs, and the code, the data generators and the repair probe, released as a short script that others can run on their own fine-tunes.

Supplementary material. Refreshed with the revision: code, run commands and results, including the 1.5B rerun and the real-task test.

A short reading guide. Page 1 (abstract, Figure 1) and Table 1 carry the result. Section 2: the setting and the short repair. Section 3: the baseline account and the toy's one role. Section 4: the facts stay recalled while composition falls, also in pretrained 0.5B. Section 5: the dose bound at 58M (with the registered prediction), at 0.3B, and in pretrained 0.5B and 1.5B. Section 6: two ways to prevent the loss. Section 8: limitations, including the real-task result.

We know discussion time is short, and we are happy to answer any question.

AuthorsOpenReview ↗

Revision: a registered real-language test (new Section 6).

In Llama-3.2-1B taught new natural-language facts about invented people and their two-hop composition, then trained on code instructions without weight decay with the facts kept in the stream, composition was lost while the facts stayed recalled, and a short repair of the last three layers restored it 0.40–0.54 above a never-composed control given the same steps (2/2 fact worlds; we registered 0.2 for "hidden"). We first ran the stage at twice this learning rate: there too the composed model repaired above the control at every lost state (0.10–0.39), but the code stage degraded the whole model and the margin fell below 0.2 at some states, so we registered this stage, read only while both models stay intact. With weight decay the facts themselves were lost too often for our erasure rule to be read, so in pretrained models the paper shows that lost composition is hidden, not that it is erased. Because the revision adds Section 6 ("In pretrained language models"), the numbers cited in our earlier responses have shifted: v2 Sections 6–9 are now 7–10 (Limitations, v2 Section 8, is now Section 9), v2 Table 3 (0.3B repair) is now Table 5 in Appendix C, v2 Table 4 (where the two-hop data must enter) is now Table 3, v2 Table 5 (1.5B) is now Table 4, and the real-task (maths-then-code) result moved from Limitations to Section 6.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.