acceptodds
Under review as a conference paper at ICLR 2027

Beyond Stable Convergence: Noise Covariance Design for Efficient Test-Time Scaling in Looped Models

Abstract

Looped models scale test-time computation through recurrent refinement, but stable convergence alone does not ensure that additional iterations improve predictions. Even near an attracting fixed point, the fixed-point residual can decrease while the predictive loss increases. We formalize this discrepancy as misalignment between the attractor and task landscapes and develop a noise covariance learning method for efficient test-time depth scaling. Through a local analysis, we establish a noise covariance design framework that accounts for task-geometry adaptation under noisy training and optimizes a directional task-loss descent bound. Building on this framework, we propose Noise Designed Looped Models (NDLM), which employs an equivalent matrix-residual objective to train a lightweight network to generate structured perturbations efficiently. Experiments across multiple reasoning tasks demonstrate NDLM's competitive performance. Remarkably, NDLM enables test-time depth scaling with no observed overthinking and matches the isotropic-noise baseline's reference accuracy with an estimated reduction in recurrent depth on Maze-Unique.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Looped Models, Reasoning, Implicit Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_TPAUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper aim to establish the relation between fixed point convergence of a implcicit deep learning model and its actual performance over tasks. To achieve so, a continuous residual dynamics is formulated and analyzed under local linearization. In my opinion the established mismatch is true and well supported. Following this the paper tries to establish a new training algorithm that retrains the fixed point to prediction part of the model in the hope to address the mismatch, they do so through noisy perturbation near the fixed point, in my understanding essentially readjusting the sensitivity of the loss so that it decreases along the converging dynamics. The method is evaluated on several different datasets and compared against some baselines, one thing it does seem to establish quite well is the monotonicity of the task performance as the number of loop increases. The rest of the results is a bit hard to parse for me.

Critical strengths

I think the paper identified a real issue that is rather underdiscussed in looped models or DEQ style models in general, that is whether the contraction of the dynamics alone means anything to the actual prediction of the model, and I think the paper did a good job establishing their claim, both in theory and numerically, and this is where I do think the linearization model is acceptable.

Under the papers own assumptions and setup the theory claims and derivations seem correct.

Critical weaknesses

I am in general very skeptical of the linearization assumptions made throughout the paper, where this is a common technique used in dynamical systems, but mostly on low dimensional problem, here the same assumptions are made willing to large neural network models with millions of parameters with high dimensional hidden state sizes, I do not consider it to be legal until some evidence can be shown, whether theoretical or empirically. this is particularly troubling when you are making a Lyapunov argument solely based on these assumptions, to which I remain skeptical.

The numerical results section reads weak to me. First, the main table is poor in presentation, with models of different origins, sizes, and numbers mixed together. Details on these models and how they are trained are largely missing from the draft. I also think the datasets used in the work are on the smaller and simpler side, where it is very easy to saturate. I find the claims of the paper numerically unconvincing when multiple rows in the table are hitting extremely high accuracy at the same time as it will be hard to actually separate the reason behind these results. Table 2 is also a bit hard to get, for the Maze-Hard example B=1 essentially is all you need, doesn't that go against the idea of the paper? looks like the learned perturbation is just degenerate.

Lastly, there is some language that I’m not sure is well substantiated. For instance, the paper uses a lot of wording such as “reasoning” or “overthinking,” but I’m not sure there is evidence to suggest anything meaningful along those lines. Similarly, the depth-scaling results are also mainly based on one of the maze examples. The paper mainly targets a DEQ problem and aims to determine the fixed point convergence and corresponding prediction loss behavior. The consistent upselling in my opinion dilutes the actual contribution and focus of the work and makes it very hard to read.

Questions for authors

The theories are derived in continuous time, while the actual model uses a discrete recurrence that the paper interprets as an Euler discretization of the dynamics, this is fine but not without details taken care of, contraction properties and Lyapunov conditions also can break if time stepping is not chosen carefully, do you consider anything other than dt = 1? should i trust dt = 1? this would be a very large step size for most dynamical systems.

The second half of the theory is mostly on the output part of the full model with the dynamics fixed and not updated, i don't think you are assuming a new fixed point when deriving it, but in the actual implementation and algorithm the dynamics are also updated? why? seems like a pretty major discrepancy? does the method actually not work if you do it the way the theory suggest? I am overall skeptical in the assumptions made regarding the method and this major difference is certainly not helping.

The selection of K is not well discussed in my opinion? since everything in the paper is suppose to be "local" getting K wrong can be catastrophic under the setup right? How do you make sure K is properly determined? and what happens if I set it too small?

5 comments

We thank the reviewer for the detailed feedback and for recognizing the attractor–task mismatch as a meaningful issue, as well as the correctness of our derivations under the stated assumptions. We address the concerns regarding the local approximation, continuous-time formulation, training procedure, choice of K, experimental evidence, and scope of the claims below.

Q1. Is the local linearization meaningful for large neural network models?

Understanding why additional recurrent computation remains beneficial near convergence is still limited. Following prior analyses of looped and equilibrium models through local dynamics and Jacobian-based characterizations [1–3], we study the neighborhood of an attracting fixed point using a local linearization. Although this setting simplifies the full nonlinear dynamics, it reveals an important mechanism: residual contraction and task-loss descent are governed by different local geometries, and appropriately designed noise covariance can improve their alignment. This analysis directly motivates NDLM and leads to a practical covariance-learning algorithm that improves test-time depth efficiency in our experiments.

For practical deployment and fair comparison, NDLM uses the standard backbone and training protocol shared by TRM, EqR, and GRAM, with no additional contraction regularization imposed by the theory. To address the gap between the theoretical regime and the actual trained model, we directly examine whether the assumptions underlying the local analysis are reflected in practice. Specifically, we evaluate the residual magnitude, the spectrum of the symmetric residual Jacobian, and the local linearization error on Maze-Unique at late recurrent states. We measure the relative linearization error as

$ E_{\mathrm{lin}} = \frac{|R(z+\delta)-R(z)-J(z)\delta|_2} {|J(z)\delta|_2+\varepsilon}. $

using random perturbations δ\delta.

Table R4.1: Empirical validation of the local theoretical regime on Maze-Unique.

StepMean Relative Residual (%)Mean Negative Eigenvalue Ratio (%)Mean Near-Zero Eigenvalue Ratio (%)Mean Relative Linearization Error (%)
322.192499.2080.005080.4977
642.220399.1990.001720.5218
1282.312799.2080.002020.5201

Across the evaluated late iterations, the relative residual remains around 2.2%, indicating that the recurrent states operate close to the fixed-point regime considered in our analysis. Approximately 99.2% of the eigenvalues of the symmetric residual Jacobian are negative, indicating predominantly contractive local dynamics. The relative linearization error is also only about 0.5%, showing that the first-order approximation accurately describes the nonlinear residual map at the tested perturbation scale. These diagnostics provide direct empirical support that the local regime underlying our analysis is realized in the actual trained model. The results are now included in Appendix E of the revised version.

[1] Yang, Xiao-Wen, Ziyu Han, Xi-Hua Zhang, Wen-Da Wei, Jie-Jing Shao, Lan-Zhe Guo, and Yu-Feng Li. “Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models.” Proceedings of the 43rd International Conference on Machine Learning (2026).

[2] Winston, Ezra, and J. Zico Kolter. “Monotone Operator Equilibrium Networks.” Advances in Neural Information Processing Systems 33 (2020).

[3] Bai, Shaojie, Vladlen Koltun, and J. Zico Kolter. “Stabilizing Equilibrium Models by Jacobian Regularization.” Proceedings of the 38th International Conference on Machine Learning (2021).

Q2. How does the continuous-time analysis relate to the finite-step recurrence?

Our continuous-time formulation is used specifically to characterize the local dynamics near an attracting fixed point, as stated in Lines 192–200, where late recurrent states are already close to equilibrium. Similar continuous-time dynamical formulations have been used to analyze recursive reasoning models [1, 2]. The implemented NDLM itself remains the native finite-step recurrence used by the original looped model.

To address the gap between this local continuous-time analysis and the implemented finite-step model, we directly examine whether the actual recurrent states enter the regime assumed by the theory. As shown in Table R4.1, the late recurrent states exhibit a relative residual of only about 2.2%, while approximately 99.2% of the eigenvalues of the symmetric residual Jacobian are negative, indicating that the actual finite-step trajectories enter a predominantly contractive near-fixed-point regime. These results support the native finite-step recurrence as a reasonable practical realization of the local dynamics characterized by our continuous-time analysis.

[1] Es'kin, Vasiliy A., and Mikhail E. Smorkalov. “Dynamical Systems Theory Behind a Hierarchical Reasoning Model.” arXiv preprint arXiv:2603.22871 (2026).

[2] Chen, Lizhang, Jonathan Li, Chen Liang, Ni Lao, and Qiang Liu. “Training-Free Looped Transformers.” arXiv preprint arXiv:2605.23872 (2026).

Q3. Why are the recurrent dynamics updated during training if the theory conditions on fixed dynamics?

The practical training procedure can be viewed as a coordinate-style alternating optimization between the task model and the noise network. Given the current recurrent dynamics FθF_\theta and residual Jacobian J, the noise network ϕ\phi is optimized through Lcov\mathcal L_{\mathrm{cov}} to adapt the perturbation covariance to the current local dynamics. The task model is then optimized through Ltask\mathcal L_{\mathrm{task}} under the resulting perturbations. Importantly, Lcov\mathcal L_{\mathrm{cov}} does not directly update the backbone. This procedure repeatedly applies the conditional covariance-design principle as the task model evolves, providing a practical counterpart to fixed-dynamics analysis.

We further examine whether the theoretical predictions remain valid after this joint optimization.

Table R4.2: Validation of the covariance-design predictions after joint training.

DiagnosticResult
Normalized Eq. (14) residual, EqR0.5267
Normalized Eq. (14) residual, NDLM0.081
HLH_L: mean negative eigenvalue ratio91.49%
HLH_L: mean near-zero eigenvalue ratio5.01%
Deterministic one-step ΔL\Delta L<0 at all evaluated depths

After joint training, NDLM substantially reduces the normalized Eq. (14) residual from 0.5267 to 0.081, showing that the learned covariance better satisfies the theoretically derived matrix condition. Meanwhile, the spectrum of HL=J⊤C+CJH_L=J^\top C+CJ is strongly dominated by negative directions, with an average of 91.49% negative eigenvalues across the evaluated depths. The deterministic one-step task-loss change is also consistently negative. These results show that the jointly trained model retains both the covariance-design condition and the task-loss descent behavior predicted by the conditional analysis. The results are now included in Appendix E of the revised version.

Q4. How sensitive is NDLM to the choice of K?

The role of K is precisely to restrict structured perturbations to recurrent states closer to the local regime analyzed theoretically. In the submitted implementation, structured perturbations are applied only during the final two recurrent steps. We additionally evaluate the sensitivity to the injection schedule:

Table R4.3: Ablation on the noise-injection schedule.

Noise Injection ScheduleAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
All recurrent steps97.398.598.799.1
Last 4 steps98.799.299.399.5
Last 2 steps (default)98.799.299.499.5
Last 1 step98.799.299.399.4

The results show that NDLM is not sensitive to the exact choice of the late-step injection schedule. Applying structured perturbations during the last 1, 2, or 4 recurrent steps yields very similar performance across all evaluated depths, with the default last-2-step setting achieving the best or tied-best accuracy throughout. In contrast, injecting structured noise at all recurrent steps consistently performs slightly worse. These results support restricting structured perturbations to late recurrent states, where the dynamics are closer to the local near-attractor regime analyzed in our theory, while also showing that performance is robust to moderate changes in K. The results are now included in Appendix E of the revised version.

Q5. How should the experimental results and the Maze-Hard breadth result be interpreted?

We agree that Table 1 does not clearly distinguish controlled comparisons from broader contextual baselines. Our primary controlled comparison is with EqR, while the remaining methods provide context across recent looped models. Detailed training configurations and hyperparameters for NDLM, EqR, and TRM are already provided in Appendix D.

We also agree that endpoint accuracy is close to saturation on some benchmarks. For this reason, we kindly note that our main empirical claim is not uniform superiority in peak accuracy, but improved test-time depth efficiency. Rather than relying on a single saturated endpoint, we compare the full accuracy–depth trajectory. We additionally evaluate three independent runs and report mean and standard deviation across three datasets:

Table R4.4: Test-time depth scaling across three independent runs.

DatasetMethodAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
Maze-UniqueEqR97.798.698.599.2
Maze-UniqueNDLM98.73 ± 0.4999.23 ± 0.4599.37 ± 0.3399.50 ± 0.29
Maze-HardEqR82.982.683.183.6
Maze-HardNDLM85.10 ± 0.0884.97 ± 0.2585.0 ± 0.2985.37 ± 0.19
Sudoku-LiteEqR83.888.491.993.2
Sudoku-LiteNDLM84.11 ± 0.3688.96 ± 0.3892.07 ± 0.2694.22 ± 0.10

Across all three datasets, NDLM consistently outperforms EqR at every evaluated recurrent depth, showing that the improvement extends beyond the Maze-Unique setting. Moreover, the small standard deviations across three independent runs indicate that the observed improvements are stable across random seeds. These results provide additional evidence that NDLM improves test-time depth scaling consistently across different structured reasoning tasks. The results are now included in Appendix E of the revised version.

We further clarify that breadth B and the learned perturbation are distinct. At B=1, NDLM still uses the learned structured covariance and samples one stochastic trajectory; increasing B only evaluates additional trajectories. Therefore, the strong Maze-Hard result at B=1 does not indicate a degenerate perturbation. Instead, it shows that the learned covariance is already effective with a single trajectory, while additional breadth provides only a modest gain on this dataset. Larger breadth gains are observed on Maze-Unique and Sudoku-Lite.

Reviewer_9HrAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies test-time depth scaling in looped models and argues that stable convergence of recurrent states does not necessarily imply continued improvement in task performance. Through a local analysis near an attracting fixed point, the authors formalize this discrepancy as a misalignment between residual contraction and task-loss descent. They then analyze how injected noise determines the stationary exploration covariance and propose a surrogate model in which noisy training adapts the local task geometry. Based on this analysis, the paper derives a noise-covariance design objective that maximizes a lower bound on the worst-direction task-loss descent rate. This objective is converted into a matrix-residual formulation and implemented as Noise Designed Looped Models (NDLM), which uses a lightweight conditional network to generate structured, input-dependent perturbations without explicitly constructing the full covariance matrix. Experiments on Sudoku-Extreme, Maze-Unique, and Maze-Hard show competitive exact accuracy. On Maze-Unique, NDLM improves depth-scaling efficiency over isotropic-noise and deterministic baselines and exhibits no observed overthinking over the evaluated depth range.

Critical strengths
  1. I think it is an interesting insight that this paper distinguishes recurrent-state convergence from task-loss improvement. This paper formalizes this issue as attractor-task misalignment. This provides a useful framework to understand why stable looped models still suffer from overthinking or ineffective additional computation.

  2. The covariance-design formulation connects local recurrent dynamics to task-loss descent and leads to an efficient noise-generation network that avoids explicitly constructing a full covariance matrix. The experiments further show meaningful depth-scaling gains on Maze-Unique, including reduced recurrent depth and no observed overthinking over the evaluated range.

Critical weaknesses
  1. The connection between the theoretical surrogate and the actual training dynamics is not clearly established. The main theory replaces realizable task-head adaptation with an unconstrained fixed-determinant positive-definite metric surrogate. It is therefore unclear to what extent the derived covariance optimum describes the Hessian induced by the actual noisy training procedure. The analysis is also conducted in a local continuous-time setting, whereas the implemented model uses finite-step discrete recurrence. Overall, it makes it questionable about how these results indeed explain practical observations.

  2. The empirical validation is limited. The strongest depth-efficiency and no-overthinking claims are concentrated on Maze-Unique, and the paper does not provide sufficient ablations or statistical variation across random seeds. In particular, comparisons against simpler anisotropic-noise baselines, alternative noise parameterizations, and different noise-injection schedules would be important to isolate the contribution of the proposed covariance objective.

Questions for authors
  1. How well does the fixed-determinant metric surrogate reflect the actual adaptation of the task head during noisy training? Can the authors provide empirical evidence that the learned task-head Hessian or local task geometry is aligned with (P_\infty^{-1}), as predicted by the theory?

  2. How does the continuous-time covariance-design analysis transfer to the finite-step discrete recurrence used in NDLM? In addition, please clarify whether the Gaussian-probe loss used to train the noise network is an unbiased or otherwise justified estimator of the stated matrix-residual objective.

  3. Can the authors provide stronger controlled comparisons to isolate the source of NDLM’s gains? In particular, comparisons with simpler anisotropic-noise parameterizations, matched random covariance baselines, alternative noise-injection schedules, and results over multiple random seeds would help determine whether the improvement comes specifically from the proposed covariance objective.

4 comments

We thank the reviewer for the careful assessment and for recognizing the value of distinguishing recurrent-state convergence from task-loss improvement, as well as the practical depth-scaling gains of NDLM. We address the concerns regarding the theory–practice connection, continuous-time approximation, controlled covariance ablations, and empirical robustness below.

Q1. How well does the theoretical surrogate reflect the practical nonlinear model?

We acknowledge that the fixed-determinant formulation in Eq. (11) serves as an analytical simplification to make the task-geometry optimization tractable, as stated in Lines 288–291. Such surrogate formulations are widely adopted in previous work to simplify difficult optimization problems [1, 2]. Specifically, fixing the determinant controls the overall volume of the positive-definite task metric, allowing us to isolate its directional structure and analytically characterize how stationary exploration affects task geometry. Accordingly, the fixed-determinant formulation provides a tractable theoretical model for deriving our covariance-design principle.

To examine whether the derived covariance-design principle remains effective in practical NDLM models despite this analytical simplification, we test whether the learned covariance and task geometry exhibit the behavior predicted by the surrogate. NDLM reduces the normalized Eq. (14) residual from 0.5267 to 0.0810, showing that the learned covariance substantially better satisfies the covariance-design condition derived in Theorem 2. We further analyze the eigenvalue spectrum of HLH_L.

Table R3.1: Empirical validation of task-geometry alignment.

StepMean Negative Eigenvalue Ratio (%)Mean Near-Zero Eigenvalue Ratio (%)Mean Deterministic ΔL\Delta L
3290.74965.89835-5.4623e-6
6493.61643.7536-1.2668e-6
12890.115.36738-1.5571e-6

Across all evaluated depths, approximately 90–94% of the eigenvalues of HLH_L are negative, while only a small fraction are near zero. This indicates that the learned task geometry is strongly biased toward directions in which recurrent refinement decreases the task loss. We further observe a negative deterministic one-step loss change at all evaluated steps. These results support the key prediction of the surrogate: the learned local geometry is biased toward task-loss descent, without requiring the nonlinear network to exactly solve the fixed-determinant optimization problem. The results are now included in Appendix E of the revised version.

Q2. How does the continuous-time analysis transfer to the finite-step recurrence, and how is the Gaussian-probe objective justified?

Our continuous-time formulation is used specifically to characterize the local dynamics near an attracting fixed point, as stated in Lines 192–200, where late recurrent states are already close to equilibrium. Similar continuous-time dynamical formulations have been used to analyze recursive reasoning models [1, 2]. The implemented NDLM itself remains the native finite-step recurrence used by the original looped model.

To address the gap between this local continuous-time analysis and the implemented finite-step model, we directly examine whether the actual recurrent states enter the regime assumed by the theory. For fair comparison, we use a standard backbone architecture also adopted by TRM, EqR, and GRAM, without imposing any explicit contraction constraint.

Table R3.2: Empirical validation of the local regime on Maze-Unique.

StepMean Relative Residual (%)Mean Negative Eigenvalue Ratio (%)Mean Near-Zero Eigenvalue Ratio (%)Mean Relative Linearization Error (%)
322.192499.2080.005080.4977
642.220399.1990.001720.5218
1282.312799.2080.002020.5201

Across Steps 32–128, the relative residual remains around 2.2%, showing that the actual finite-step model operates close to its fixed-point regime. Approximately 99.2% of the eigenvalues of the symmetric residual Jacobian are negative, indicating predominantly contractive local dynamics. Moreover, we quantify the relative linearization error as

$ E_{\mathrm{lin}} = \frac{|R(z+\delta)-R(z)-J(z)\delta|_2} {|J(z)\delta|_2+\varepsilon}. $

using random perturbations δ\delta. In particular, the mean errors of approximately 0.5% indicate that the first-order approximation accurately describes the nonlinear residual map. These results provide empirical evidence that the implemented finite-step model indeed reaches the near-fixed-point local regime targeted by our analysis, and that the first-order local characterization is accurate in this regime. The results are now included in Appendix E of the revised version.

For Eq. (16), random Gaussian probing is a standard technique for efficiently estimating implicit matrix quantities without explicitly constructing the underlying matrix, closely related to classical stochastic probing and trace-estimation methods [3, 4] and their use in neural-network models [5]. In our case, the Gaussian-probe objective is not an unbiased estimator of the Frobenius objective in Eq. (14), but it shares the same zero-residual optimum: its population loss is zero if and only if the matrix residual in Eq. (14) is zero.

[1] Es'kin, Vasiliy A., and Mikhail E. Smorkalov. “Dynamical Systems Theory Behind a Hierarchical Reasoning Model.” arXiv preprint arXiv:2603.22871 (2026).

[2] Chen, Lizhang, Jonathan Li, Chen Liang, Ni Lao, and Qiang Liu. “Training-Free Looped Transformers.” arXiv preprint arXiv:2605.23872 (2026).

[3] Hutchinson, M. F. "A Stochastic Estimator of the Trace of the Influence Matrix for Laplacian Smoothing Splines." Communications in Statistics - Simulation and Computation 18(3): 1059–1076 (1989).

[4] Avron, Haim, and Sivan Toledo. "Randomized Algorithms for Estimating the Trace of an Implicit Symmetric Positive Semi-Definite Matrix." Journal of the ACM 58(2): Article 8 (2011).

[5] Grathwohl, Will, Ricky T. Q. Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. "FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models." International Conference on Learning Representations (2019).

Q3. Do the gains specifically come from the proposed covariance design?

To directly examine this, we conduct a controlled ablation on Maze-Unique with progressively more expressive noise parameterizations.

Table R3.3: Ablation on noise parameterization and covariance structure.

Noise ParameterizationAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
Fixed Isotropic (EqR)97.798.698.599.2
Learnable Scalar93.49394.195.4
State-dependent Isotropic92.39394.395.7
State-dependent Diagonal91.191.99392.4
Matched Random Correlated Covariance95.996.396.997.1
NDLM (Ours)98.73 ± 0.4999.23 ± 0.4599.37 ± 0.3399.50 ± 0.29

The results further isolate the role of the proposed covariance design. The learnable scalar, state-dependent isotropic, and state-dependent diagonal variants all underperform the fixed isotropic baseline across the evaluated depths. Introducing a matched random correlated covariance improves over several adaptive variants, but still remains consistently below EqR and NDLM, indicating that correlation or anisotropy alone is insufficient. In contrast, NDLM achieves the highest accuracy at every evaluated depth, reaching 99.50% at depth 128. These results support the importance of designing the directional covariance according to the task and local recurrent dynamics, rather than simply increasing the flexibility or correlation structure of the injected noise.

We additionally evaluate the reviewer’s suggested noise-injection schedule ablation.

Table R3.4: Ablation on the noise-injection schedule.

Noise Injection ScheduleAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
All recurrent steps97.398.598.799.1
Last 4 steps98.799.299.399.5
Last 2 steps (default)98.799.299.499.5
Last 1 step98.799.299.399.4

The results show that NDLM is not sensitive to the exact choice of the late-step injection schedule. Applying structured perturbations during the last 1, 2, or 4 recurrent steps yields very similar performance across all evaluated depths, with the default last-2-step setting achieving the best or tied-best accuracy throughout. In contrast, injecting structured noise at all recurrent steps consistently performs slightly worse. These results support restricting structured perturbations to late recurrent states, where the dynamics are closer to the local near-attractor regime analyzed in our theory, while also showing that performance is robust to moderate changes in K. The results are now included in Appendix E of the revised version.

Q4. Are the depth-scaling improvements statistically reliable and generalizable beyond Maze-Unique?

To further examine the generality and statistical reliability of our depth-scaling improvements, we evaluate accuracy as a function of recurrent depth on three datasets. We evaluate three independent runs and report the mean accuracy and standard deviation across seeds.

Table R3.5: Test-time depth scaling with three independent runs

DatasetMethodAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
Maze-UniqueEqR97.798.698.599.2
Maze-UniqueNDLM98.73 ± 0.4999.23 ± 0.4599.37 ± 0.3399.50 ± 0.29
Maze-HardEqR82.982.683.183.6
Maze-HardNDLM85.10 ± 0.0884.97 ± 0.2585.0 ± 0.2985.37 ± 0.19
Sudoku-LiteEqR83.888.491.993.2
Sudoku-LiteNDLM84.11 ± 0.3688.96 ± 0.3892.07 ± 0.2694.22 ± 0.10

Across all three datasets, NDLM consistently outperforms EqR at every evaluated recurrent depth, showing that the improvement extends beyond the Maze-Unique setting. Moreover, the small standard deviations across three independent runs indicate that the observed improvements are stable across random seeds. These results provide additional evidence that NDLM improves test-time depth scaling consistently across different structured reasoning tasks. The results are now included in Appendix E of the revised version.

Reviewer_9HrAOpenReview ↗

Response to the rebuttal

Thank you for the detailed response and the additional experiments. The new noise-parameterization and noise-injection schedule ablations, together with results across multiple datasets and random seeds, have substantially addressed my concerns about the empirical evaluation.

My remaining concern is the correspondence between the fixed-determinant surrogate and the task geometry learned by the actual model. The reduction in the Eq. (14) residual and the negative average one-step loss changes provide useful evidence for the effectiveness of the proposed design. However, these results do not directly establish whether the learned task-head Hessian or task metric aligns with P∞−1P_\infty^{-1}, which was the specific question raised in my review. Since this central connection between theory and practice remains only partially addressed, I maintain my rating.

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

This paper proposes Noise Designed Looped Models (NDLM) to improve test-time depth scaling in recurrent reasoning models. The authors identify a mismatch between convergence toward a stable fixed point and improvement in task performance, showing that decreasing the fixed-point residual does not necessarily imply decreasing task loss. To address this issue, they develop a theoretical framework connecting injected noise covariance, stationary exploration, and task-loss geometry. Under a local surrogate model, they characterize the optimal noise covariance and derive an equivalent matrix-residual optimization objective. Based on this analysis, NDLM employs a lightweight neural network to generate structured, state-dependent Gaussian perturbations. Experiments on Sudoku-Extreme, Maze-Unique, and Maze-Hard demonstrate competitive accuracy and improved inference-depth efficiency compared with isotropic Gaussian noise.

Critical strengths

Interesting theoretical analysis. The paper establishes a connection between noise covariance, recurrent dynamics, and task-loss descent. In particular, Theorem 1 characterizes how noisy training can align task geometry with recurrent dynamics under a surrogate model, while Theorem 2 provides a characterization of the optimal covariance. These results offer a principled motivation for structured noise injection beyond empirical heuristics.

Simple and practical algorithm. NDLM introduces a lightweight noise-generation network without substantially modifying the underlying recurrent architecture. The proposed covariance objective can be optimized using vector–Jacobian products, avoiding the explicit construction of high-dimensional Jacobian and covariance matrices.

Promising test-time scaling results. NDLM achieves competitive accuracy across three structured reasoning benchmarks. On Maze-Unique, it reaches EqR's reference accuracy with an estimated 82.7% reduction in recurrent depth. The experiments also demonstrate improved accuracy with increasing recurrent depth and no observed overthinking within the evaluated range, suggesting that structured noise can improve inference-depth efficiency.

Critical weaknesses

Limited algorithmic novelty. The main algorithmic modification is replacing isotropic Gaussian perturbations with a learned, state-dependent noise generator. Related approaches, such as GRAM, have already investigated learned state-dependent Gaussian perturbations in recurrent reasoning. While the theoretical derivation is interesting, the practical algorithm remains a relatively incremental extension of existing stochastic recurrent models.

Insufficient ablation studies. The paper compares NDLM primarily against fixed isotropic Gaussian noise, but does not investigate simpler learnable alternatives, such as a learned scalar noise magnitude or a state-dependent diagonal covariance. These baselines are important for determining whether the improvements arise specifically from learning correlated covariance structure, rather than simply adapting the noise distribution. Furthermore, the paper does not analyze the learned covariance matrices to establish whether they exhibit meaningful anisotropy.

Limited experimental evidence. The accuracy improvements over EqR are relatively modest, particularly on benchmarks where both methods already achieve nearly saturated performance. For example, NDLM improves Maze-Unique accuracy from 97.1% to 99.0%, while its Maze-Hard accuracy of 86.5% is only slightly higher than TRM's 85.3%. The paper also does not report multiple independent training runs, standard deviations, or confidence intervals, making it difficult to assess the statistical reliability of these improvements.

Gap between theoretical analysis and practical implementation. The theoretical results rely on local linearization near a stable fixed point and a surrogate model of task-geometry adaptation, which may not accurately characterize the training dynamics of nonlinear neural networks. In particular, the assumption that task geometry adapts according to a fixed-determinant optimization problem is not empirically justified. The paper does not verify whether the learned covariance approximates the theoretically optimal covariance or whether the predicted task-geometry alignment actually occurs during training.

Questions for authors

Simpler noise parameterizations. Have the authors considered comparing NDLM against simpler alternatives, such as a learnable scalar noise magnitude (\beta), state-dependent isotropic noise (\beta_\phi(x,z)\epsilon), or a learned diagonal covariance? Such comparisons would help clarify whether structured covariance provides additional benefits beyond adaptive noise magnitude.

Properties of the learned covariance. What does the learned covariance look like in practice? Does it exhibit significant anisotropy, or does it approach a scaled identity matrix? Could the authors report its eigenvalue distribution or condition number, and investigate whether the learned covariance aligns with the theoretical optimum characterized by Theorem 2?

Statistical reliability of experimental improvements. Could the authors report results over multiple independent training seeds, including standard deviations or confidence intervals? This would be particularly valuable for Maze-Hard, where the accuracy differences between competing methods are relatively small, and for the test-time depth-scaling experiments.

Validation of theoretical assumptions. Can the authors provide empirical evidence that noisy training induces the task-geometry alignment predicted by Theorem 1? For example, measuring the local task Hessian and recurrent Jacobian in a small model could help determine whether the theoretical conditions hold approximately in practice and whether the learned covariance produces the predicted task-loss descent behavior.

3 comments

We thank the reviewer for the thoughtful and constructive feedback, and especially for recognizing the value of our task-geometry analysis, the covariance-design objective, and the practical VJP-based implementation. We address the concerns regarding novelty, ablations, statistical reliability, and the connection between our theory and practical models below.

Q1. Precise novelty relative to existing stochastic recurrent models

Our central contribution is to establish a task-loss-oriented principle for noise covariance design in looped models. GRAM learns state-dependent Gaussian perturbations through a variational formulation, while retaining an isotropic covariance. This leaves unexplored how the directional structure of injected noise can be optimized to align recurrent attractors with task-loss geometry. We address this gap by developing a dynamics-aware covariance-design objective and translating it into a practical algorithm that learns structured, anisotropic perturbations guided by local recurrent dynamics. Thus, NDLM advances noise learning from simple perturbation generation to principled, task-oriented covariance design.

Specifically, our contributions are threefold:

  1. Attractor–task misalignment (Proposition 1). We show that fixed-point residual contraction does not necessarily imply task-loss descent, even under local contraction.

  2. Noise-induced task-geometry adaptation (Theorems 1 and 2). We characterize how noise covariance shapes stationary exploration and task geometry, deriving an objective that maximizes a lower bound on the worst-direction task-loss descent rate.

  3. Principled and efficient covariance learning (Section 3.3). We derive a tractable matrix-residual objective and implement it using a lightweight network to generate structured perturbations.

Q2. Do the gains specifically come from the proposed covariance design?

To directly examine this, we conduct a controlled ablation on Maze-Unique with progressively more expressive noise parameterizations, ranging from adaptive noise magnitude to direction-dependent variance.

Table R2.1: Ablation on noise parameterization.

Noise ParameterizationAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
Fixed Isotropic (EqR)97.798.698.599.2
Learnable Scalar93.49394.195.4
State-dependent Isotropic92.39394.395.7
State-dependent Diagonal91.191.99392.4
NDLM (Ours)98.73 ± 0.4999.23 ± 0.4599.37 ± 0.3399.50 ± 0.29

The results show that increasing the flexibility of the noise parameterization alone does not lead to better performance. The learnable scalar, state-dependent isotropic, and state-dependent diagonal variants all underperform the fixed isotropic baseline across the evaluated depths. In contrast, NDLM consistently achieves the highest accuracy, reaching 99.50% at depth 128. This suggests that the gains arise from the proposed task- and dynamics-aware covariance design, which explicitly optimizes the directional structure of the perturbations, rather than from simply making the noise adaptive or more expressive. The results are now included in Appendix E of the revised version.

Q3. What does the learned covariance look like?

We analyze the learned covariance on representative late-iteration states of Maze-Unique to determine whether it collapses toward a scaled identity and whether it is consistent with the covariance structure characterized by Theorem 2.

Table R2.2: Properties of the learned covariance on Maze-Unique.

MethodMean Condition NumberMean Off-diagonal Absolute-value Ratio (%)Mean Normalized Eq. (14) Residual
EqR100.5267
NDLM (Ours)1.78316.95980.081

The learned covariance has a mean condition number of 1.783, compared with 1 for isotropic noise, indicating a clear departure from isotropy. Moreover, its mean off-diagonal absolute-value ratio reaches 16.9598%, demonstrating non-negligible cross-dimensional correlations in the learned perturbations. More importantly, NDLM reduces the normalized Eq. (14) residual from 0.5267 to 0.0810, indicating that the learned covariance substantially better satisfies the matrix optimality condition derived in Theorem 2. These results show that NDLM learns a nontrivial directional covariance structure that is closely aligned with our theoretical design objective. The results are now included in Appendix E of the revised version.

Q4. Statistical reliability of the improvements

As stated in Lines 152–155 and Lines 423–454, our evaluation focuses on test-time depth efficiency, i.e., obtaining predictive gains from additional recurrent computation with fewer iterations. To further examine the generality and statistical reliability of this behavior, we evaluate accuracy as a function of recurrent depth on three datasets. We evaluate three independent runs and report the mean accuracy and standard deviation across seeds.

Table R2.3: Test-time depth scaling with three independent runs

DatasetMethodAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
Maze-UniqueEqR97.798.698.599.2
Maze-UniqueNDLM98.73 ± 0.4999.23 ± 0.4599.37 ± 0.3399.50 ± 0.29
Maze-HardEqR82.982.683.183.6
Maze-HardNDLM85.10 ± 0.0884.97 ± 0.2585.0 ± 0.2985.37 ± 0.19
Sudoku-LiteEqR83.888.491.993.2
Sudoku-LiteNDLM84.11 ± 0.3688.96 ± 0.3892.07 ± 0.2694.22 ± 0.10

Across all three datasets, NDLM consistently outperforms EqR at every evaluated recurrent depth, showing that the improvement extends beyond the Maze-Unique setting. Moreover, the small standard deviations across three independent runs indicate that the observed improvements are stable across random seeds. These results provide additional evidence that NDLM improves test-time depth scaling consistently across different structured reasoning tasks. The results are now included in Appendix E of the revised version.

Q5. Empirical validity of the local stability assumptions

To examine whether this local regime is meaningful in practice, we evaluate the residual magnitude, the spectrum of the symmetric residual Jacobian, and the accuracy of the first-order linearization on Maze-Unique. For fair comparison, we use a standard backbone architecture also adopted by TRM, EqR, and GRAM, without imposing any explicit contraction constraint.

Table R2.5: Empirical validation of the local theoretical regime on Maze-Unique.

StepMean Relative Residual (%)Mean Negative Eigenvalue Ratio (%)Mean Near-Zero Eigenvalue Ratio (%)Mean Relative Linearization Error (%)
322.192499.2080.005080.4977
642.220399.1990.001720.5218
1282.312799.2080.002020.5201

Across the evaluated late iterations, the mean relative residual remains around 2.2%, indicating that the recurrent states are already close to the fixed-point regime considered in our local analysis. Despite the absence of explicit stability regularization, approximately 99% of the eigenvalues of the symmetric part of the residual Jacobian are negative, with only about 0.005% near zero, indicating predominantly contractive local dynamics.

Moreover, we quantify the relative linearization error as

$ E_{\mathrm{lin}} = \frac{|R(z+\delta)-R(z)-J(z)\delta|_2} {|J(z)\delta|_2+\varepsilon}. $

using random perturbations δ\delta. In particular, the mean errors of approximately 0.5% indicate that the first-order approximation accurately describes the nonlinear residual map.

These results provide empirical evidence that the near-attractor, locally linear regime assumed by our analysis is a reasonable approximation of the dominant late-stage dynamics. The results are now included in Appendix E of the revised version.

Q6. Does the predicted task-geometry alignment occur in practice?

We acknowledge that the fixed-determinant formulation in Eq. (11) serves as an analytical simplification to make the task-geometry optimization tractable, as stated in Lines 288–291. Such surrogate formulations are widely adopted in previous work to simplify difficult optimization problems [1, 2]. Specifically, fixing the determinant controls the overall volume of the positive-definite task metric, allowing us to isolate its directional structure and analytically characterize how stationary exploration affects task geometry. Accordingly, the formulation provides a tractable theoretical model for deriving our covariance-design principle.

To examine whether the derived covariance-design principle remains effective in practical NDLM models despite this analytical simplification, we analyze the eigenvalue spectrum of HLH_L. Negative eigenvalues correspond to local directions along which the recurrent dynamics favor task-loss descent, whereas positive eigenvalues correspond to directions that can increase the task loss.

Table R2.6: Empirical validation of task-geometry alignment.

StepMean Negative Eigenvalue Ratio (%)Mean Near-Zero Eigenvalue Ratio (%)Mean Deterministic ΔL\Delta L
3290.74965.89835-5.4623e-6
6493.61643.7536-1.2668e-6
12890.115.36738-1.5571e-6

Across all evaluated depths, approximately 90–94% of the eigenvalues of HLH_L are negative, while only a small fraction are near zero. This indicates that the learned task geometry is strongly biased toward directions in which recurrent refinement decreases the task loss. We further observe a negative deterministic one-step loss change at all evaluated steps, directly confirming that additional recurrent refinement locally decreases the task loss. The small magnitude of ΔL\Delta L is expected because these measurements are performed at late recurrent states close to the fixed-point regime. The results are now included in Appendix E of the revised version.

Together with the substantially lower Eq. (14) residual reported in Q3, these results provide empirical support for both components of our theoretical prediction: the learned covariance better satisfies the derived covariance-design condition, while the resulting local task dynamics are biased toward task-loss descent. This evidence supports the practical relevance of the analytical surrogate without assuming that the full nonlinear training dynamics exactly implement Eq. (11).

[1] Jiang, Qijia, Olaoluwa Adigun, Harikrishna Narasimhan, Mahdi Milani Fard, and Maya Gupta. "Optimizing Black-box Metrics with Adaptive Surrogates." Proceedings of the 37th International Conference on Machine Learning (ICML), pp. 4784–4793 (2020).

[2] Bottou, Léon, Frank E. Curtis, and Jorge Nocedal. "Optimization Methods for Large-Scale Machine Learning." SIAM Review 60(2): 223–311 (2018).

Reviewer_Aj3HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces a new method for improving the performance of looped models by introducing learnable noise covariance in an SDE formulation of the looped models, and use this to design a new type of looped model (NDLM) incorporating these ideas and generating structured perturbations. The theory behind the design is thoroughly explored, and on the experimental side, NDLM is compared against a wide range of existing baselines on standard datasets used for this setting. NDLM achieves competitive performance with existing methods on those datasets, with a claimed up to 82.7% reduction in recurrent depth.

Critical strengths

1)The rigorous analysis of the design choices provides a strong foundation for the choices made, and helps explain their performance.

  1. I find the task-geometry adaptation, as well as the way of training the network to generate the noise perturbations using independent Gaussian probes, to be quite an interesting idea that seems to work well.
Critical weaknesses
  1. In general, the level of novelty and improvement in the paper is less significant than claimed. The dynamical systems perspective has already been introduced and explored by [1] and [2], and the stochastic differential equation interpretation was already explored by [1]. Additionally, there seems to be a lot of similar styles of work (either relating to dynamical system interpretations or learned noise, etc). In the experimental comparisons, this model is competitive but doesn't strongly outperform [2] overall ([2] performs slightly better on Sudoku-Extreme, while this model performs slightly better on the other two tasks). But considering the small amount of parameters used by each model, the extra parameters this model introduces could be sufficient to boost the score enough (i.e. it may not be the strength of the learnable noise that helps, but just the extra parameters).

  2. There are multiple points in the theory which I think need better clarification/justification, and in general, much of the theory is just using standard dynamical systems results. On the assumptions side, the main results assume that the fixed point being studied is stable (but as far as I could see, there is no demonstration that this actually holds for practical models and on practical datasets). Additionally, is there any evidence (empirically) that the residual Jacobian is actually strictly negative definite (i.e. that J is nonsingular)?

[1] Es' kin, Vasiliy A., and Mikhail E. Smorkalov. "Dynamical Systems Theory Behind a Hierarchical Reasoning Model." arXiv preprint arXiv:2603.22871 (2026).

[2] Huang, Benhao, Zhengyang Geng, and Zico Kolter. "Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning." arXiv preprint arXiv:2605.21488 (2026).

Questions for authors
  1. Regarding the theoretical assumptions, is there any empirical evidence that shows that those assumptions actually do hold in practical setups?
  2. How does this align with the results in the EqR paper which tested a learnable noise term, and found that it didn't actually help much?
  3. Can you clarify the precise novelty compared to the existing dynamical systems exploration of looped models (including the ones I have listed in the weaknesses section)?
3 comments

Q1. Novelty relative to existing dynamical-systems approaches

Our central contribution is to establish a principled connection between noise covariance design and attractor–task alignment in looped models. To our knowledge, this is the first work to explicitly investigate how noise can be designed to align recurrent dynamics with task-loss geometry. This distinguishes our work from [1] and [2]. In [1], stochastic dynamics are studied in the context of dynamical stability and representation diversity, while [2] investigates stochastic perturbations as a means of facilitating exploration. Both works leave unexplored how the directional structure of injected noise influences the alignment between recurrent attractors and task-loss geometry. We address this gap by developing a task-geometry-aware covariance-design principle and translating it into a practical noise-learning algorithm that improves test-time depth efficiency on structured reasoning benchmarks. The discussion on [1] and [2] is included in the revised Related Work section.

Specifically, our contributions are threefold:

  1. Attractor–task misalignment (Proposition 1). We show that fixed-point residual contraction does not necessarily imply task-loss descent, even under local contraction.

  2. Noise-induced task-geometry adaptation (Theorems 1 and 2). We characterize how noise covariance shapes stationary exploration and task geometry, deriving an objective that maximizes a lower bound on the worst-direction task-loss descent rate.

  3. Principled and efficient covariance learning (Section 3.3). We derive a tractable matrix-residual objective and implement it using a lightweight network to generate structured perturbations.

[1] Es'kin, Vasiliy A., and Mikhail E. Smorkalov. "Dynamical Systems Theory Behind a Hierarchical Reasoning Model." arXiv preprint arXiv:2603.22871 (2026).

[2] Huang, Benhao, Zhengyang Geng, and Zico Kolter. "Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning." arXiv preprint arXiv:2605.21488 (2026).

Q2. Why does NDLM improve when learnable noise in EqR showed limited benefits, and could the gains simply come from additional parameters?

We thank the reviewer for raising these two closely related questions. The key distinction in NDLM is how the perturbation structure is parameterized and optimized, instead of noise learnability alone. EqR results suggest that simply making the noise learnable does not necessarily provide a substantial improvement over fixed isotropic noise. Our work instead optimizes the directional structure of the perturbation through Eq. (16), which is motivated by the bi-level optimization problem described in Eq. (9). Thus, our claim is that the benefit comes from task- and dynamics-aware covariance design, rather than merely introducing trainable noise parameters.

To isolate the contribution of the proposed noise parameterization from model capacity, we conduct a controlled ablation over different noise parameterizations on Maze-Unique. For the parameter-matched control, we use EqR with a slightly larger backbone so that its total parameter count is matched to NDLM. This isolates whether the observed gains can be explained simply by increased model capacity.

Table R1.1: Ablation on noise parameterization.

Noise ParameterizationAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
Fixed Isotropic (EqR)97.798.698.599.2
Learnable Scalar93.49394.195.4
State-dependent Isotropic92.39394.395.7
State-dependent Diagonal91.191.99392.4
Parameter-matched Control97.898.398.698.7
NDLM (Ours)98.73 ± 0.4999.23 ± 0.4599.37 ± 0.3399.50 ± 0.29

The results show that simply introducing learnable noise does not improve performance. The learnable scalar, state-dependent isotropic, and diagonal variants all underperform fixed isotropic noise across the evaluated depths, indicating that additional flexibility alone is insufficient. Likewise, the parameter-matched control remains comparable to EqR and does not reproduce the gains of NDLM, suggesting that the improvement cannot be explained by the increased parameter count. In contrast, NDLM consistently achieves the best accuracy across all evaluated depths, reaching 99.50% at depth 128. These results support the importance of the proposed task- and dynamics-aware covariance design, rather than noise learnability or additional model capacity alone. The results are now included in Appendix E of the revised version.

Q3. Empirical validity of the theoretical assumptions

Our theory focuses on the local regime near a stable fixed point. To assess whether this regime is reflected in trained models, we evaluate both the residual magnitude and the local residual Jacobian on Maze-Unique at late recurrent states.

For a fair comparison, we adopt a standard backbone architecture also used in existing baselines, including TRM, EqR, and GRAM, without imposing any additional contraction constraints. Nevertheless, our empirical analysis shows that approximately 99% of the eigenvalues of the symmetric part of the residual Jacobian are negative, with only approximately 0.005% being near zero. This indicates that local dynamics exhibit contraction in the overwhelming majority of directions.

Table R1.2: Empirical validation of the local theoretical regime on Maze-Unique.

StepMean Relative Residual (%)Mean Negative Eigenvalue Ratio (%)Mean Near-Zero Eigenvalue Ratio (%)
322.192499.2080.00508
642.220399.1990.00172
1282.312799.2080.00202

These observations provide strong empirical support for the local contraction assumption, suggesting that it is a mild and practically reasonable assumption for analyzing trained looped models. The results are now included in Appendix E of the revised version.

Q4. Overall performance and inference-depth efficiency

As stated in Lines 152–155 and Lines 423–454, our evaluation focuses on test-time depth efficiency, i.e., obtaining predictive gains from additional recurrent computation with fewer iterations. To further examine the generality and statistical reliability of this behavior, we evaluate accuracy as a function of recurrent depth on three datasets. We evaluate three independent runs and report the mean accuracy and standard deviation across seeds.

Table R1.3: Test-time depth scaling with three independent runs.

DatasetMethodAcc@16 ↑Acc@32 ↑Acc@64 ↑Acc@128 ↑
Maze-UniqueEqR97.798.698.599.2
Maze-UniqueNDLM98.73 ± 0.4999.23 ± 0.4599.37 ± 0.3399.50 ± 0.29
Maze-HardEqR82.982.683.183.6
Maze-HardNDLM85.10 ± 0.0884.97 ± 0.2585.0 ± 0.2985.37 ± 0.19
Sudoku-LiteEqR83.888.491.993.2
Sudoku-LiteNDLM84.11 ± 0.3688.96 ± 0.3892.07 ± 0.2694.22 ± 0.10

Across all three datasets, NDLM consistently outperforms EqR at every evaluated recurrent depth, showing that the improvement extends beyond the Maze-Unique setting. Moreover, the small standard deviations across three independent runs indicate that the observed improvements are stable across random seeds. These results provide additional evidence that NDLM improves test-time depth scaling consistently across different structured reasoning tasks. The results are now included in Appendix E of the revised version.

These responses and the revised paper have resolved some but not all of my concerns. I appreciate the new experiments and Appendix additions. The parameter-matched comparison is now resolved, and I do agree that it properly rules out model size as any cause of the improvement over prior works. However, Table 5 in the text states that the condition number is only ~1.8, which indicates only mild anisotropy. It may be the case that this mild anisotropy is all that is needed to solve these tasks, so I do not consider it a negative result specifically, but I would request the authors be careful to avoid overstating these results.

Regarding the empirical validity of the assumptions, and specifically the Jacobian properties, while the results show that the vast majority (99.2%) of eigenvalues are in fact negative, that still leaves 0.8% of eigenvalues positive. Meanwhile, the theory relies on strict negative definiteness, so even one positive eigenvalue is enough to separate from the theory. I would request that you verify carefully whether the theory can in fact hold under these "predominantly contractive" settings and provide at least some kind of explanation of how the theory can be adjusted.

Additionally, it seems that Table 1 in the main text has been revised on the Maze-Unique setting, to use significantly smaller models for both EqR and NDLM, but it is not clear why that was done, or why such a smaller model improves the metrics for both models. That change causes the improvement of NDLM over EqR to become quite narrow (0.3 points). That also affects the claimed reduction in depth (from 82.7% to 75.0%).

This claimed reduction in depth seems to be quite overstated too, and could be simply caused by noise. A threshold of 99.7% is used, while EqR reaches 99.6% at depth 16. From my understanding of this dataset, that means that EqR solves one less maze at 16 depth (out of 1000 total mazes) which is not a significant enough gap to make broad claims about. If I understand correctly, EqR solving a single additional maze would have caused the claimed depth improvement to only be ~13%.

Finally, on the novelty side, I do agree that for this specific task, there does seem to be enough novelty and differentiating factors compared to the prior work. However, although you state "To our knowledge, this is the first work to explicitly investigate how noise can be designed to align recurrent dynamics with task-loss geometry.", this is not in fact the first that explores this in general recurrent dynamics. Both [3] and [4] consider recurrent or discretized continuous-depth models with random noise injections, and how those injections affect training. They do not study a designed or learned noise setting though, but it does affect the framing on the SDE side, and I recommend that the authors revise their framing to clarify this.

If these points can be clarified, I do believe that it can be sufficient enough to be an interesting work. For now, I will raise my Soundness to 2, but I think that more needs to be resolved before I can raise my score.

[3] Lim, Soon Hoe, et al. "Noisy recurrent neural networks." Advances in Neural Information Processing Systems 34 (2021): 5124-5137. [4] Liu, Xuanqing, et al. "How does noise help robustness? explanation and exploration under the neural sde framework." 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020.

Comments on the paper

1 comment

Dear ACs and SACs,

Thank you very much for your time and effort in handling our submission.

As the author–reviewer discussion period comes to a close, we would like to kindly follow up regarding the discussion. We have submitted detailed responses to all four reviewers, addressing their questions and providing additional experimental results and clarifications. However, we have not yet received any follow-up from the reviewers during the discussion period.

We would simply like to ensure that our responses and the additional evidence have been adequately considered. If possible, we would greatly appreciate your help in checking whether the reviewers might have any further questions or points requiring clarification.

We remain ready and willing to address any remaining concerns.

Thank you again for your time, consideration, and support.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.