acceptodds
Under review as a conference paper at ICLR 2027

How Does Chain-of-Thought Supervision Teach Models to Reason? A Theoretical Analysis

Abstract

Training with chain-of-thought (CoT) supervision is an effective approach for teaching models multi-step reasoning procedures. At inference time, however, models may need to continue reasoning beyond the finite teacher CoT sequences observed during training, and longer reasoning can instead degrade performance, a phenomenon known as overthinking. How the amount and structure of training-time CoT give rise to such failures remains poorly understood theoretically. To study this problem, we introduce an in-context weight prediction model trained on finite gradient descent trajectories for linear regression. Using high-dimensional asymptotic analysis, we show that the success and failure modes of long-horizon reasoning undergo phase transitions depending on the context length and the total amount of training CoT. We further distinguish two mechanisms of overthinking, error accumulation and error amplification, and show that when the training context is short, a model can learn to continue reasoning before it learns to correct errors. Experiments with fully trained linear and multi-head softmax attention further show that these qualitative behaviors persist. Our results provide a theoretical account linking properties of training-time CoT to inference-time failures and offer guidance for designing teacher CoT that supports stable reasoning.

Primary area:
learning theory
Keywords:
chain-of-thought, asymptotic theory, overthinking, phase transition

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

This paper studies training-time CoT supervision through gradient-descent trajectories on linear-regression tasks. A linear-attention model learns one-step transitions and is then applied recursively beyond the training horizon. The analysis characterizes regimes of GD recovery, bounded error, quadratic divergence, and exponential divergence, including non-monotonic effects of increasing the training CoT budget. Experiments with fully trained linear and softmax attention test whether these qualitative behaviors persist beyond the restricted theoretical model.

Critical strengths
  1. The paper addresses a concrete question beyond inference-time expressiveness: how finite trajectory supervision affects the stability of longer test-time computation. The non-monotonic dependence on supervision budget is an interesting finding.

  2. The experiments with fully trained attention provide useful support beyond the single-matrix theoretical parameterization. The distinction between error accumulation and amplification also helps organize the observed failure modes.

Critical weaknesses
  1. The theoretical setting builds substantial structure into both the teacher and the student. The teacher supplies GD state transitions, while the effective student fixes most attention blocks and learns only a matrix through a strongly convex objective. This gives a tractable account of recovering and iterating a prescribed update rule, but supports a narrower interpretation than a general explanation of how CoT supervision teaches reasoning.

  2. The broader explanatory contribution remains unclear. The paper extends prior in-context regression and iterative-dynamics analyses to finite teacher trajectories, but the empirical extensions remain within the same regression-trajectory setting. Moreover, the softmax experiment reproduces non-monotonicity without clearly separating the two divergence regimes. Thus, the evidence supports some qualitative predictions more strongly than the full four-regime account or the claimed practical guidance for CoT design.

Questions for authors
  1. Which qualitative conclusion is specific to finite teacher-trajectory supervision and cannot be obtained from the closest existing analyses of iterative in-context learning?

  2. In the fully trained experiments, how robust is the non-monotonic behavior to changes in teacher step size and trajectory length? Which parts of the four-regime picture should be expected to survive beyond linear attention?

3 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

Thank you for your thoughtful review. We address your questions and concerns point by point below. The manuscript has been revised accordingly, with changes highlighted in red.

W1: The theoretical setting builds substantial structure into both the teacher and the student. The teacher supplies GD state transitions, while the effective student fixes most attention blocks and learns only a matrix through a strongly convex objective. This gives a tractable account of recovering and iterating a prescribed update rule, but supports a narrower interpretation than a general explanation of how CoT supervision teaches reasoning.

Thank you for this comment. We agree that our analysis constrains the attention parameters and formulates the teacher CoT as GD, so it does not provide a universal explanation of CoT supervision in the complex mechanisms of real LLMs. At the same time, the constraints on the student are not arbitrary. As described in Section 3.3 and detailed in Appendix A, each fixed block is justified by one of three considerations: the rotational symmetry of the data, the fact that fixing it entails no loss of generality in our setting, or preventing the model from advancing its inference without consulting the context.

That said, we emphasize that the problem is far from trivial even in this setting. For example, although the student's hypothesis class contains the parameters that reproduce GD, there is no guarantee that training finds them, because the amount of data is finite. Our main results, the phase transitions and the non-monotonicity, are all caused by this finiteness of data and are not built into the model structure by hand.

In addition, reaching an answer by repeatedly applying the same update rule is a structure shared by many algorithms, as discussed in the new Section 3.6, "Connection to CoT in Practice," and whether reasoning generalizes beyond the length of the teacher CoT is one of the central questions about reasoning generalization. Our reasoning setting is therefore not far removed from real tasks.

Furthermore, Section 6 confirms that similar qualitative trends appear when linear and softmax attention are trained end-to-end on a non-convex objective, and the new Appendix G shows that the non-monotonicity in these models is robust to the teacher step size. More importantly, qualitatively similar phenomena have already been observed in prior work based on large LLMs. For example, [1] reports a non-monotonic behavior in reasoning SFT with long CoT, where generalization performance first drops during training and then recovers. We discuss the relation of this observation to our theory at the end of Section 5.3, and we believe our work is one of the first to explain this behavior. A theory that can predict the reasoning of LLMs will be built through a series of studies, and we believe this paper is an important step.

[1] Q. Ren, P. Wang, R. Cai, S. Shao, D. Guo, Y. Xie, Y. Li, Q. Zhang, X. Hu, J. Shao, D. Liu, "Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability," ICML 2026.

AuthorsOpenReview ↗

Rebuttal (2/3)

W2: The broader explanatory contribution remains unclear. The paper extends prior in-context regression and iterative-dynamics analyses to finite teacher trajectories, but the empirical extensions remain within the same regression-trajectory setting. Moreover, the softmax experiment reproduces non-monotonicity without clearly separating the two divergence regimes. Thus, the evidence supports some qualitative predictions more strongly than the full four-regime account or the claimed practical guidance for CoT design.

Thank you for the comment. For the contribution beyond prior work, please see the reply to Q1 below.

We agree that our experimental results mainly provide strong support for some of the qualitative predictions. Although we do not reproduce all four regimes in the fully trained experiments, we believe that our results still lead to the following practical guidance for CoT design.

  • Prepare more teacher CoT for tasks with shorter contexts, or lengthen the contexts to reduce the teacher CoT required. Since the budget required for recovery depends on the context length (αφ>1\alpha\varphi>1), the amount of teacher CoT should not be chosen independently of the context length (Section 5.1, paragraph "Context length determines the CoT budget required").
  • With short contexts, be careful when increasing teacher CoT gradually. Because an intermediate budget can worsen long-horizon performance, teacher CoT should be added until the budget exceeds the unstable range, and long test-time reasoning should be avoided if this is not possible. We have added this guidance to Section 5.3. This non-monotonicity appears in both linear and softmax attention and persists across teacher step sizes (Section 6 and the new Appendix G).
  • Collect more diverse tasks rather than lengthening the CoT of the same tasks. The value of teacher CoT is determined by the diversity of states rather than the number of steps, and if successive teacher states are redundant, shorter teacher CoT from more diverse tasks is more informative (Section 5.1, paragraph "CoT budget matters more than task diversity or CoT length alone"). In end-to-end trained models, increasing the CoT length at a fixed number of tasks did not reproduce the effect of adding tasks (new Appendix G).

Q1: Which qualitative conclusion is specific to finite teacher-trajectory supervision and cannot be obtained from the closest existing analyses of iterative in-context learning?

Thank you for the question. As summarized in the Related Work (paragraph "Theoretical studies of CoT"), prior analyses of iterative in-context learning mainly study the test-time behavior of trained models and do not characterize the effect of providing finite teacher CoT as training data. We have added Appendix I, "Further Related Work," which compares our work with the closest analyses in detail.

In short, unlike [1], we show that overthinking can occur even under in-distribution generalization and consider the regime where the number of samples and the dimension are of the same order. Unlike [5], where the model is trained only on the final answer, we train the model on finite teacher trajectories. This yields the following conclusions specific to finite teacher-trajectory supervision. First, the learned algorithm can succeed beyond the number of steps seen in the teacher CoT (Result 4.1 and Section 5.1). Second, the number of tasks and the CoT length enter only through the training CoT budget φ=τ(K−1)\varphi=\tau(K-1), and the budget required for recovery depends on the context length through the boundary αφ=1\alpha\varphi=1 (Section 5.1). Third, a failure to reproduce the algorithm takes two forms, accumulation and amplification (Section 5.2). Fourth, long-horizon performance is non-monotonic in the training CoT budget (Section 5.3).

[1] J. Huang, Z. Wang, J. D. Lee, "Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought," ICLR 2025.

[2] J. Kim, T. Suzuki, "Transformers Provably Solve Parity Efficiently with Chain of Thought," ICLR 2025.

[3] E. Abbe, S. Bengio, A. Lotfi, C. Sandon, O. Saremi, "How Far Can Transformers Reason? The Globality Barrier and Inductive Scratchpad," NeurIPS 2024.

[4] A. Javanmard, B. Mirzasoleiman, V. Mirrokni, "Understanding the Role of Training Data in Test-Time Scaling," ICLR 2026.

[5] K. Takanami, C. Pehlevan, "An Asymptotic Theory of Chain-of-Thought in In-Context Learning," NeurIPS 2026.

AuthorsOpenReview ↗

Rebuttal (3/3)

Q2: In the fully trained experiments, how robust is the non-monotonic behavior to changes in teacher step size and trajectory length?

Thank you for the question. We conducted additional experiments with end-to-end trained linear and softmax attention (Appendix G, Figure 4).

To compare teacher step sizes η∈{0.05,0.1,0.2}\eta\in\{0.05,0.1,0.2\} at the same amount of GD progress, we evaluated the error at the depth satisfying ηt=3.2\eta t=3.2. For every step size, both models show the non-monotonic pattern. The step size shifts the budget at which the error drops and changes the height of the peak, but it does not remove the non-monotonicity.

When the budget was increased through the teacher CoT length at a fixed number of tasks, the abrupt transition to stable inference observed when increasing the number of tasks was not reproduced. For linear attention, the error decreases only gradually with the CoT length, and for softmax attention it remains almost unchanged. This suggests that in end-to-end training the model must also learn how to use the context, and since this information is supplied once per task, a longer CoT cannot compensate for it. In the effective model, this part is fixed by the architecture, so the learned solution depends only on φ\varphi.

These results suggest that the qualitative effect of increasing task diversity holds fairly generally in experiments, whereas the result that performance depends only on the teacher CoT budget may be limited to the effective model. We have included these results in Appendix G.

Q2(continued): Which parts of the four-regime picture should be expected to survive beyond linear attention?

Thank you for the question. Even for softmax attention, a sufficient teacher CoT budget yields stable inference beyond the training horizon, and the non-monotonicity in the budget (stable, then unstable, then stable) persists. However, we cannot confirm whether the unstable region clearly separates into the accumulation and amplification regimes. We attribute this to the fact that the Jacobian of softmax attention depends on the state and changes along the inference trajectory (Section 6). Even so, the transition with respect to the budget is phase-transition-like, which is consistent with the prediction of the effective model.

Reviewer_PyDGOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3(+1)
Presentation
3
Soundness
3
Read review
Paper summary

It is a theoretical paper studying CoT for supervised learning where an in-context weight prediction model is trained on finite gradient descent trajectories for linear regression. It shows the success and failure modes of long-horizon reasoning undergo phase transitions depending on the context length and the total amount of training CoT. The experiments with fully trained linear attention and 2-head softmax attention reproduce length generalization and the non-monotonic relationship between the amount of CoT supervision and long-horizon performance.

Critical strengths

S1. The paper provides theoretical conditions on how finite-length CoT supervision generalizes reasoning beyond the training horizon. I find the theory for CoT length generalization novel and valuable.

S2. Fig1 clearly shows how different learning regimes affect long-horizon generalization.

S3. This paper reveals a counterintuitive conclusion on non-monotonic effects of the training CoT budget.

Critical weaknesses

W1. The CoT formulation is both a strength and a limitation. It introduces the teacher supervision for studying the supervised learning procedure, while the prompt for predicting sμ,k+1s_{\mu,k+1} contains only the task context QμQ_\mu and the current state sμ,ks_{\mu,k}, but no earlier states. In real world reasoning models, they usually extend the length of cot by non-linear reasoning behaviours like backtracking or self-correction.

W2. There still a gap between the experimental setting and the real world chain of thoughts in LLMs. Controlled experiments with GPT-2 on a synthetic CoT reasoning task would make the conclusions more convincing.

Questions for authors

Q1. Can your CoT formulation be extended to capture reasoning behaviors such as backtracking or explicit self-correction? I suggest adding a subsection in sec3 to discuss the connection between your theoretical modeling and real world CoT, and explain why training CoT consists of GD trajectories. Is it because it is easy to generalize in length? I think it comes from [1], and they have discussed about the generalization in sec4.3. What is the difference of your paper? Also, there are a lot of different CoT formulations should be compared with.

Q2. Could you provide a concrete CoT example in real world, e.g., math problems, to explain the conclusions derived from your theory? I believe it would help to build an intuition beyond the linear regression setting.

It will raise my assessment to this paper if two questions above are fully addressed.

Q3. How would you construct a CoT training dataset for a real world language model to test your theoretical predictions? Specifically, if the non-monotonic effects in sec5.3 can be reproduced or be measured in real world llm training, it would substantially improve my assessment.

[1] Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought. Huang et al. (2025)

5 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

Thank you for your thoughtful review. We address your questions and concerns point by point below. The manuscript has been revised accordingly, with changes highlighted in red.

W1: The CoT formulation is both a strength and a limitation. It introduces the teacher supervision for studying the supervised learning procedure, while the prompt for predicting sμ,k+1 contains only the task context Qμ and the current state sμ,k, but no earlier states. In real world reasoning models, they usually extend the length of cot by non-linear reasoning behaviours like backtracking or self-correction.

Thank you for this insightful comment. We agree that the Markovian structure is a restriction needed to make the model solvable. However, our test-time CoT repeatedly applies the same model to a state, which corresponds to latent and recurrent-depth reasoning. Moreover, CoT formats in which each step depends only on the input and the previous state have been proposed for algorithmic tasks and improve generalization to longer reasoning [1]. We therefore do not regard the Markovian structure itself as a weakness of this work. We have added this discussion to Section 3.6.

We also believe that extensions including backtracking or self-correction are possible starting from our model. For example, if GD with momentum is used as the teacher CoT, the learned attention becomes a minimal model that adjusts its update by referring to past states. The test-time dynamics are then described by a higher-order recursion. We have added this direction to the future work in the conclusion (Section 7).

In this sense, our analysis corresponds to practical reasoning to some extent and provides a mathematical model that can serve as a starting point for analyzing more advanced CoT models.

[1] E. Abbe, S. Bengio, A. Lotfi, C. Sandon, O. Saremi, "How Far Can Transformers Reason? The Globality Barrier and Inductive Scratchpad," NeurIPS 2024.

Also, see the reply to Q1 below.

W2: There still a gap between the experimental setting and the real world chain of thoughts in LLMs. Controlled experiments with GPT-2 on a synthetic CoT reasoning task would make the conclusions more convincing.

Thank you for the suggestion. We have not conducted experiments with GPT-2 in this revision. As controlled experiments beyond the linear effective model, Section 6 trains linear and softmax attention end-to-end on the same task and shows that the non-monotonic effect of the training CoT budget persists. We have also added Appendix G (Figure 4), which further varies the teacher step size and the training CoT length in these models. The non-monotonicity persists across teacher step sizes, whereas increasing the CoT length at a fixed number of tasks does not reproduce the effect of increasing the number of tasks.

See also the reply to Q3 below for how our predictions could be tested with language models.

AuthorsOpenReview ↗

Rebuttal (2/3)

Q1: Can your CoT formulation be extended to capture reasoning behaviors such as backtracking or explicit self-correction? I suggest adding a subsection in sec3 to discuss the connection between your theoretical modeling and real world CoT, and explain why training CoT consists of GD trajectories. Is it because it is easy to generalize in length? I think it comes from [1], and they have discussed about the generalization in sec4.3. What is the difference of your paper? Also, there are a lot of different CoT formulations should be compared with.

Thank you for the helpful suggestion. We have added Section 3.6, which discusses how our model relates to CoT in practice and why the training CoT consists of GD trajectories.

In short, Section 3.6 explains that our test-time CoT, which repeatedly applies the same model to a state, corresponds to latent and recurrent-depth reasoning, which is often trained with explicit CoT data, and that our training setting abstracts this procedure by supervising the intermediate states with a finite teacher trajectory. It also explains that CoT formats in which each step depends only on the input and the previous state have been proposed for algorithmic tasks, and that in-context linear regression solved by GD is the simplest such setting that can be analyzed exactly. For the first question on backtracking and self-correction, please see our reply to W1.

We adopted GD as an abstraction of CoT because prior work [1,4,5] has shown it to be a useful model for revealing several properties of CoT.

We have also added Appendix I, "Further Related Work," which compares our work with prior studies in detail. In short, compared with Section 4.3 of [1], we show that overthinking can occur even under in-distribution generalization, and we identify its conditions and two distinct mechanisms. Compared with [5], where the model is trained only on the final answer, we train the model on finite teacher trajectories, which reveals generalization beyond the training horizon, the two failure mechanisms, and the non-monotonicity in the training CoT budget. Furthermore, formulations of CoT in which each step solves a separate subproblem [2,3] are complementary to ours, where each step refines a prediction for the same problem.

[1] J. Huang, Z. Wang, J. D. Lee, "Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought," ICLR 2025.

[2] J. Kim, T. Suzuki, "Transformers Provably Solve Parity Efficiently with Chain of Thought," ICLR 2025.

[3] E. Abbe, S. Bengio, A. Lotfi, C. Sandon, O. Saremi, "How Far Can Transformers Reason? The Globality Barrier and Inductive Scratchpad," NeurIPS 2024.

[4] A. Javanmard, B. Mirzasoleiman, V. Mirrokni, "Understanding the Role of Training Data in Test-Time Scaling," ICLR 2026.

[5] K. Takanami, C. Pehlevan, "An Asymptotic Theory of Chain-of-Thought in In-Context Learning," NeurIPS 2026.

Q2: Could you provide a concrete CoT example in real world, e.g., math problems, to explain the conclusions derived from your theory? I believe it would help to build an intuition beyond the linear regression setting.

Thank you for the question. Section 5.1 (paragraph "Context length determines the CoT budget required") gives a concrete example of our conclusion that a short context requires a larger training CoT budget. Suppose the task is sorting, but the context contains only lists without duplicates, such as [2,1,3]. Then sorting and removing duplicates before sorting produce the same output, so the examples do not identify the intended rule, whereas an example with duplicates, such as [1,2,2,3], distinguishes the two rules. Similarly, in our model a short context constrains the update rule only on a subspace, so more teacher CoT across contexts is needed to identify it.

See also the reply to Q3 for tasks that could be used to test our predictions with language models.

AuthorsOpenReview ↗

Rebuttal (3/3)

Q3: How would you construct a CoT training dataset for a real world language model to test your theoretical predictions? Specifically, if the non-monotonic effects in sec5.3 can be reproduced or be measured in real world llm training, it would substantially improve my assessment.

Thank you for the constructive suggestion. As a first step toward testing our predictions with language models, we think multi-step algorithmic tasks in which every reasoning step applies the same state-transition rule are appropriate. Examples include the following [1].

(i) Parity checking of bit strings, where the teacher CoT gives the parity of the first kk bits.

(ii) Multi-digit addition, where the teacher CoT gives the partial sum of the first kk digits, including carries.

(iii) Graph reachability, where the teacher CoT gives the sequence of vertices reached by following kk edges from the starting vertex.

When the context for these tasks is insufficient, one can study how increasing the number of hint sequences determines the final long-horizon reasoning performance. This would test the qualitative non-monotonicity predicted by our paper, namely that with short contexts, adding teacher CoT can temporarily worsen long-horizon performance. We view our work as establishing a theory that can be tested with such tasks, and we have added this direction to the future work in the conclusion (Section 7).

We agree that the suggested synthetic experiments with GPT-2 would be valuable. More importantly, however, qualitatively similar phenomena have already been observed in prior work based on large LLMs. For example, [2] reports a non-monotonic behavior in reasoning SFT with long CoT, where generalization performance first drops during training and then recovers. We discuss the relation of this observation to our theory at the end of Section 5.3, and we believe our work is one of the first to explain this behavior. A theory that can predict the reasoning of LLMs will be built through a series of studies, and we believe this paper is an important step.

[1] E. Abbe, S. Bengio, A. Lotfi, C. Sandon, O. Saremi, "How Far Can Transformers Reason? The Globality Barrier and Inductive Scratchpad," NeurIPS 2024.

[2] Q. Ren, P. Wang, R. Cai, S. Shao, D. Guo, Y. Xie, Y. Li, Q. Zhang, X. Hu, J. Shao, D. Liu, "Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability," ICML 2026.

Reviewer_6JyhOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates how finite CoT supervision affects the long-horizon reasoning behavior of attention models. Using gradient descent trajectories for linear regression as teacher CoT, the authors analyze when a learned update rule can recover gradient descent and when recurrent inference leads to bounded error, quadratic divergence, or exponential divergence.

Critical strengths
  • The theoretical framework connects the amount of CoT supervision and context length to long-horizon inference stability.
  • The distinction between error accumulation and error amplification provides a clear explanation of two different mechanisms underlying long-horizon reasoning failures.
  • The non-monotonic trends observed in fully trained attention models provide additional support for the theoretical findings.
Critical weaknesses
  • When multiple matrices can fit the training CoT, the analysis considers only the minimum-norm solution. It does not discuss how other valid solutions might affect the predicted error regimes.
  • The phase transition analysis uses a simplified random matrix, but it is not shown whether this approximation preserves the spectral properties required for the theoretical conclusions.
  • For 0<α<10<\alpha<1 and 0<ϕ<1/α0<\phi<1/\alpha, the phase transition analysis focuses on the long-time limit t→∞t\to\infty and does not characterize the generalization error at finite test-time depths in these regimes.
Questions for authors
  • How does the correlation between successive teacher CoT states affect the tradeoff between CoT length and task diversity?
  • Can overthinking occur due to overfitting noise even when GD is perfectly recovered? How would this relate to the mechanisms discussed in the paper?
3 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Thank you for your thoughtful review. We address your questions and concerns point by point below. The manuscript has been revised accordingly, with changes highlighted in red.

W1: When multiple matrices can fit the training CoT, the analysis considers only the minimum-norm solution. It does not discuss how other valid solutions might affect the predicted error regimes.

Thank you for raising this point. We have added a brief discussion in Appendix H and a pointer to it in Section 3.3.

In short, Appendix H makes the following points. The GD parameter itself is one of the interpolating solutions, so choosing a different interpolating solution can change the error regime. Whether the training CoT determines the update uniquely, however, does not depend on the selection rule. The boundary φ=max⁡{1,1/α}\varphi=\max\{1,1/\alpha\} does not depend on the choice of interpolating solution. Also, the solution we analyze is the one selected by standard training, in the sense that GD on the effective model from zero initialization selects the same minimum-Frobenius-norm interpolating solution as the limit λ→0\lambda\to0 of ridge regression.

W2: The phase transition analysis uses a simplified random matrix, but it is not shown whether this approximation preserves the spectral properties required for the theoretical conclusions.

Thank you for this comment. We agree that, instead of analyzing the spectrum of the learned matrix A^\hat{A} directly, we replace it with a moment-matched effective matrix A^eff\hat{A}_{\mathrm{eff}}, and that this replacement is not rigorously justified.

We have clarified this point in Appendix E.2. In short, the replacement is motivated by the fact that, conditional on the state projectors, the centered part of A^\hat{A} is a sum of M=O(D)M=O(D) independent task contributions. It follows the same idea as the Gaussian equivalence principle [1], which replaces a nonlinear model by a Gaussian model with matched moments and was later established rigorously in related settings [2,3]. A similar replacement of a random matrix by an effective matrix with matched second moments was used in [4], with results in close agreement with numerical experiments. Although we do not establish the replacement rigorously, the resulting phase boundaries agree with finite-dimensional experiments.

For the last point, please see Fig. 1C (D=96D=96, α=0.625\alpha=0.625), where the numerically measured error changes its behavior at the theoretical phase boundaries (vertical dotted lines), and Fig. 1D, where the long-time behavior in each phase matches the theoretical predictions.

[1] S. Goldt, B. Loureiro, G. Reeves, F. Krzakala, M. Mézard, L. Zdeborová, "The Gaussian equivalence of generative models for learning with shallow neural networks," MSML 2021

[2] H. Hu, Y. M. Lu, "Universality Laws for High-Dimensional Learning with Random Features," IEEE Transactions on Information Theory, 69(3), 2023.

[3] A. Montanari, B. N. Saeed, "Universality of empirical risk minimization," COLT 2022

[4] K. Takanami, T. Takahashi, Y. Kabashima, "Learning Linear Regression with Low-Rank Tasks In-Context," AISTATS 2026

W3: For 0 < α < 1 and 0 < ϕ < 1/α, the phase transition analysis focuses on the long-time limit t → ∞ and does not characterize the generalization error at finite test-time depths in these regimes.

Thank you for pointing this out. As explained in Section 4, the finite-depth error requires two-point functions of the learned update, and in our setting the corresponding cavity equations do not close because the training CoT couples successive transitions. In the long-time limit, by contrast, the type of behavior is determined by one-point spectral quantities alone, which is why Result 4.2 is stated for t→∞t\to\infty in this regime.

We note three points. First, the finite-depth error is characterized exactly in the GD and projected-GD regimes (Result 4.1, with derivations in Appendices C and D). Second, the finite-depth behavior in the regime 0<α<10<\alpha<1 and 0<φ<1/α0<\varphi<1/\alpha is examined numerically in Fig. 1C, which shows the error at several finite test-time depths as a function of φ\varphi at α=0.625\alpha=0.625, and in Fig. 1D, which shows the full test-time error dynamics E(t)\mathcal{E}(t) in each phase. Third, the goal of this work is to identify which reasoning failures underlie the long-time behavior, and this can be answered without resolving the intermediate dynamics.

AuthorsOpenReview ↗

Rebuttal (2/2)

Q1: How does the correlation between successive teacher CoT states affect the tradeoff between CoT length and task diversity?

Thank you for the question. Within the effective model, the correlation between successive teacher CoT states does not affect the tradeoff, as long as the training CoT length KK is O(1)O(1). What matters for identifying the update is the span of the training states, not their correlation. In Appendix C (around Eq. (C.15)), we show that the K−1K-1 nonzero states of each task are linearly independent almost surely even though they are correlated, and that the states of different tasks are in general position. The dimension of the training-state span is therefore min⁡{D,M(K−1)}\min\{D,M(K-1)\}, so the recovery condition depends on MM and KK only through the budget φ=τ(K−1)\varphi=\tau(K-1) (Section 5.1, paragraph "CoT budget matters more than task diversity or CoT length alone").

Correlation becomes relevant when successive states are redundant, for example for much longer teacher CoT whose states become nearly collinear as GD converges. In that case, additional steps add little new information, and more diverse tasks are more informative. Section 5.1 discusses this point. Intuitively, a longer CoT can substitute for more diverse tasks when each CoT step applies the same procedure to a new situation, but not when the teacher CoT keeps repeating essentially the same situation.

Beyond the effective model, Appendix G (Figure 4C,D) shows that, for end-to-end trained linear and softmax attention, increasing the CoT length at a fixed number of tasks does not reproduce the effect of increasing the number of tasks. This suggests that, when the model must also learn how to use the context, longer CoT cannot fully substitute for task diversity.

Q2: Can overthinking occur due to overfitting noise even when GD is perfectly recovered? How would this relate to the mechanisms discussed in the paper?

Thank you for the question. Yes. Even when GD is perfectly recovered, the generalization error can be non-monotonic, because GD on the test context eventually fits the observation noise. This effect is already contained in EGD(t)\mathcal{E}_{\mathrm{GD}}(t) itself and is a property of the algorithm, not of learning it. The divergences we analyze are different. They arise because the model fails to learn GD from finite teacher CoT, and we characterize overthinking through the deviation of E(t)\mathcal{E}(t) from EGD(t)\mathcal{E}_{\mathrm{GD}}(t).

We have added this discussion as a footnote in Section 4, right after Result 4.1.

The rebuttal addresses most of the concerns. I would maintain the original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.