acceptodds
Under review as a conference paper at ICLR 2027

Donor-Derived Token Organization Changes What a Neural Network Can Reach

Abstract

Learned neural representations can alter what a new learner subsequently reaches, but identifying which properties of a transferred representation carry that effect requires interventions that preserve some properties while replacing others. We study this problem by transplanting a rank-4 carrier learned on modular addition into independently initialized recipients. In a prospectively specified experiment, donor carriers reach a fixed generalization criterion in 45 of 64 donor–recipient cells, whereas matched substitutes preserving the donor spectrum, right basis, norm, rank, and recipient coupling reach it in 0 of 64. The effect replicates on a fresh recipient cohort (47/64 versus 0/64). Severe projection onto only the donor’s dominant Fourier modes retains substantial reachability (35/64). Moreover, synthetic carriers matched to these projections in rank, Fourier-support complexity, dominant-mode concentration, and singular spectrum are themselves active (36/64 versus 0/64 for norm-matched random carriers), although the experiment does not identify which matched property carries the effect. Conversely, randomly permuting a donor carrier across token identities reduces reachability from 47/64 to 0/64 while leaving it in exactly the same row-permutation orbit. Thus, descriptors invariant to row-to-token assignment cannot determine learning reachability at this operating point. Together, these interventions show that compact representational organization can causally alter finite-horizon learning reachability, while progressively narrowing—but not yet identifying—the structural properties sufficient to carry the effect.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, transfer learning, representation transfer, grokking, learning dynamics, mechanistic interpretability, neural network representations, modular arithmetic

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_w1UaOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies which properties of transferred token embeddings affect subsequent learning on modular addition modulo 113. It compares donor-derived rank-4 carriers against controls preserving selected spectral and geometric properties, Fourier projections, synthetic Fourier carriers, and row-permuted carriers. Success means attaining at least 95% validation accuracy for three consecutive evaluations within 50,000 training steps. Donor carriers succeed in 45/64 donor–recipient cells, versus 0/64 for the principal controls, and achieve 47/64 on fresh recipient initializations. Fourier projections retain substantial success, while a matched synthetic construction achieves 36/64 versus 35/64 for the projections in a further cohort. The contribution is a controlled empirical study of initialization effects, with particularly careful matching and explicit limitations.

Critical strengths
  1. Preserving the donor spectrum, right-side structure, and recipient coupling while replacing the token frame rules out several simple explanations simultaneously. Together with cloned recipient parameters, this makes the local intervention result credible.

  2. Its success prevents attributing the projection’s advantage over the original synthetic baseline to donor origin alone. This is a useful scientific correction, and the authors appropriately avoid interpreting a nonsignificant difference as equivalence.

Critical weaknesses
  1. The paper’s own related-work discussion already establishes Fourier representations and embedding transfer as relevant ingredients. Its strongest formal conclusion—that row-permutation-invariant descriptions cannot determine subsequent success—does not, by itself, provide much additional explanation: arbitrary row permutations destroy the correspondence between token identities and modular structure. More importantly, the matched synthetic carrier succeeds in 36/64 cells, versus 35/64 for the donor projection. These counts do not establish equivalence, but they leave the added value of the donor-derived realization unestablished. The synthetic construction still imports donor-derived structural targets, while a simple donor-independent, concentrated Fourier baseline is missing. Consequently, the experiments neither identify the operative property nor establish that learning a donor supplies something beyond an appropriate Fourier initialization. The manuscript carefully acknowledges these limitations, but acknowledging them does not supply the missing contribution.

  2. Of the 45 donor runs counted as successful, 43 subsequently fall below the threshold; only 11/64 satisfy the three-evaluation criterion at training’s end, and only two remain continuously above threshold from 300 steps after attainment onward. Finite-horizon attainment is a legitimate endpoint, but it does not establish stable generalization or separate altered eventual attainability from altered timing beyond the chosen horizon. This weakens the paper’s proposed distinction from acceleration work. The evidence also comes from one task, one architecture family, one partition, and the same eight donors across the recipient replications. The main operating point was explicitly selected partly because an unstructured carrier failed. At the only other tested operating point, donor success falls to 8/64 while the matched substitution reaches 25/64. That contrast is not statistically resolved, but it provides no support for robustness. The demonstrated phenomenon is therefore considerably narrower than the conceptual framing suggests.

Questions for authors
  1. What are the exact donor validity gate and checkpoint-harvesting rule, and how many candidate donor seeds were excluded? This determines how strongly the conclusions depend on selecting a particular class of successful donors.

  2. What are the terminal success rates and post-attainment occupancy statistics for FP and FSM on their shared cohort? Do their advantages over the random reference persist under these outcomes?

  3. Can a donor-independent one- or two-frequency Fourier carrier, with a fixed spectrum and matched norm, reproduce the observed benefit? Establishing this comparison, particularly on independently sampled partitions.

Reviewer_1DDiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In this paper, the authors investigate which properties of a learned weight matrix need to be preserved in order for it to keep its effect when transplanted into a freshly initialized network. They study this in a modular addition task (modulo113), learned by a three-layer MLP whose input is a 4d embedding of the 113 tokens (integers mod 113), defined by the 113x4 matrix AA. They pretrain this embedding matrix AA, then transfer it or one of several controls into a new MLP; they quantify the performance of the weight transfer by testing whether validation accuracy with the new MLP ever stays at or above 0.95 for three consecutive evaluations, over a fixed number of training steps. Their main findings are that random matrices do not perform well, the pretrained matrix as well as its approximation by top Fourier modes do; and finally, the pretrained matrix with its rows randomly permuted across tokens does not.

Critical strengths
  • The core result is clean: the paper shows three things about the pretrained embedding’s effect after transfer:

    • It survives compression to one or two Fourier modes.
    • It survives replacement by a synthetic Fourier embedding at random frequencies.
    • It does not survive scrambling which token gets which row, even though scrambling preserves every statistic that ignores row assignment.
  • The experimental protocol is done quite meticulously, and many things are pre-registered

Critical weaknesses
  • The scope is the one well-controlled experiment. Furthermore,. because they do not show learning curves (for example), it seems to me that they do not clearly distinguish reachability from acceleration, which is, according to them, the main distinction between the current results and the closest antecedent (GrokTransfer). Other forms of robustness, for example to hyperparameter changes, could also be investigated further.

  • Many significant aspects of the paradigm do not seem to be discussed in the main text, for example, the pretraining paradigm for the transferred weight matrix, how the input tokens are combined in the modular addition task, or whether the transferred matrix itself continues to train after the transfer.

Questions for authors
  • Is AA trained after transplantation? Please give the architecture, the learning rates and weight decay for AA and BB, and how/when the rank-4 matrix is extracted from the donor.

  • Why does the donor matrix do worse than a random one at weight decay 3.0?

  • In runs that end with training accuracy below 0.02, has the embedding collapsed to zero, and when does this happen relative to reaching the criterion?

Reviewer_UKh5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper explores the intersection of weight re-use and grokking, by pre-training a token carrier and then injecting it into another network. Properties of the donor and the donor-recipient relationship change how often the resulting network learns a modular addition task.

Critical strengths

None

Critical weaknesses
  1. The paper seems heavily AI-written. Although the authors do disclose AI assistance in the declarations section, I find this paper really hard to understand, in large part due to characteristic AI writing patterns.
  2. The paper is extremely hard to follow. Regardless of AI writing patterns, I struggle to understand the motivation and methodological setup of this paper. It's not even clear what architectures are being studied.
  3. The evaluation is limited to a single task and model architecture. This is an extremely niche setting that says little about problems of interest in the field.
Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.