- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies joint learning of attention and an MLP logic module on permutation-based CoT state tracking. A mean-field loss and empirical attention closure yield approximate dynamics for three order parameters. These describe an initial mixed heuristic, followed by attention specialization and stronger target alignment. Predictions quantitatively match constrained-model trajectories and qualitatively capture the rollout-accuracy transition. Architectural relaxations and standard one-block transformers show similar stages.
Critical strengths
S1. The model gives an interpretable quantitative account of the learning stages. Eq. 15 links attention learning to the logic margin , while improved retrieval strengthens target alignment. Fig. 2 supports this mutual dependence.
S2. The staged behavior persists beyond the constrained architecture. In standard one-block transformers, holding attention uniform or freezing the MLP at initialization leaves rollout accuracy near chance, supporting the importance of learning in both subsystems.
Critical weaknesses
W1. The qualitative advance is limited. Yang et al. (2024) already analyze MLP-first learning followed by joint attention–MLP evolution, and Cabannes et al. (2024) identify positional retrieval followed by state updates. The quantitative model is useful, but the central staged-learning picture offers limited conceptual novelty.
W2. The closures need stronger validation. Eq. 15a is empirical; taking does not remove its approximation error, so Appendix B.3 should qualify its exactness claim. More importantly, marginal prediction symmetry does not justify Eq. 96's recovery approximation: conditioning on an incorrect state can change the states and learned-map entries encountered during rollout. The relevant conditional probabilities are not directly checked.
Questions for authors
Q1. How closely does match the measured attention-gradient factor through the learning transition? Does Appendix B.3's exactness claim as concern only discretization error, retaining the mean-field and empirical-closure assumptions?
Q2. Around the rollout-accuracy transition, can you measure in Eq. 93 and compare with ? Separately, how well does Eq. 17's logit-based match measured ? This would separate recovery-closure errors from Gaussian-logit errors.
We thank the reviewer for the constructive comments and the precise questions. Q1 and Q2 led to the two main changes of the revision.
W1: Qualitative advance. We agree that the order, MLP first and attention later, is not new. The revised Section 4.2 relates it to both papers and states what our analysis adds. In Yang et al. (2024), a linear MLP classifies all training samples correctly by the end of the first phase, with attention near its initial values. Attention and MLP then enlarge the margin, and the attention probabilities change only by the order of the inverse squared sequence length (their Corollary F.17). In our task the first stage does not solve the problem: with attention held uniform, the rollout accuracy stays at chance (Table 2), and the task is solved only when attention moves from 1/L to nearly 1. Our equations describe this transition. Cabannes et al. (2024) identify the iteration head, which retrieves the current input token by its position so that the MLP can apply the state update, and study its emergence and transfer between tasks; they leave a detailed study of the training dynamics outside their scope. What we add: (i) the first stage is a mixed heuristic, in which on- and off-target alignment both grow; (ii) the margin R−S between them gates attention learning and is opposed by logit noise that grows as attention sharpens; (iii) closed differential equations for three order parameters predict the timing of both stages in the baseline model (general comment); (iv) the rollout transition is a separate, later event: attention saturates at α≈3, and the rollout accuracy rises at α≈5, sharply because the next-token accuracy enters to the L-th power.
W2/Q1: Gradient factor and exactness. LcωA(1−A) does not match the measured factor through the transition. It is correct at A=1/L, where its coefficient was derived, but too small by a factor of 1.9 at A=0.2 and of about 5.5 near A=0.7; the general comment explains why Fig. 2 agreed nonetheless. We now derive c(A) instead of fitting its form. The new Fig. 6 (Appendix B.3 of the revised manuscript) compares it with the measured factor throughout training; the two agree to within about 10%. On exactness the reviewer is right: η→0 removes only the discretization error, while the mean-field evaluation of the loss and the assumptions behind c(A) remain. Appendix B.3 now says so and lists the remaining approximations.
W2/Q2: Recovery approximation. We agree that symmetry alone does not establish Eq. (96). The equation assumes that, from an incorrect state, the model errs as often as from a correct one, and that its errors land on each of the other N−1 states with equal probability. We now measure At and Bt of Eq. (93) throughout training, including the rollout transition. The measured Bt show no dependence on the position t and follow (1−At)/(N−1) within the resolution of the new Fig. 7 (Appendix B.6).
Q2: ρ versus At. Here, At is the next-token accuracy, which can be measured empirically, whereas ρ of Eq. (17) is the next-token accuracy under the assumption of an explicit logit distribution. In the course of the more detailed mean-field investigation, we updated this distribution to include the effects of the uniform initialization of the logic module's matrices (see general comment). A comparison is shown in Fig. 18 (middle panel, Appendix C.5), with the empirical At, averaged over the positions since there is no observable position dependence, and ρ using the empirical variances. Together, the two comparisons separate the errors of the recovery approximation from those of the logit distribution.