acceptodds
Under review as a conference paper at ICLR 2027

When Dynamics Are Suboptimal: Isolating Dynamics Inductive Bias in Representation Learning

Abstract

In many dynamical models, representations and dynamics are learned jointly, and the chosen dynamics model class and fitting objective thereby impose an inductive bias on representation learning. At a given training iterate, the dynamics parameters can be suboptimal for the representation under the chosen fitting criterion, so the induced representation gradient can differ from the corresponding gradient under a fitted dynamics reference. We theoretically characterize this representation-gradient discrepancy. In the canonical case where the dynamics-induced representation gradient is obtained by differentiating the dynamics-fitting objective with respect to the representation parameters, the discrepancy is exactly the gradient of the dynamics-parameter suboptimality gap. Controlled experiments in a sequential latent-variable model show that dynamics-parameter suboptimality alters representation formation, with consequences for learned dynamical quality. The effect is most pronounced during early representation formation, motivating Asymmetric Coupling Warm-up (ACW), a lightweight intervention that temporarily attenuates dynamics-to-representation coupling early in training. Applying ACW across distinct representation–dynamics paradigms yields consistent changes in optimization trajectories and downstream outcomes, providing cross-paradigm evidence that the identified mechanism has observable consequences beyond the controlled setting.

Primary area:
learning on time series and dynamical systems
Keywords:
Latent Dynamics, Representation Learning, Joint Optimization, Inductive Bias

Reviews

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

The paper studies how an underoptimized dynamics predictor affects representation learning during joint training. It characterizes the difference between representation gradients obtained using the current predictor and a predictor optimally fitted to the same representation. A controlled sequential-VAE study on a double pendulum examines the resulting changes in representations and prediction quality. The authors then propose Asymmetric Coupling Warm-up (ACW), which temporarily weakens dynamics-to-representation gradients, and evaluate it in three additional model–task settings.

Critical strengths
  • The controlled refitting experiment is useful. Freezing the learned representations and fitting fresh predictors under matched procedures helps distinguish changes in representation quality from differences in the original predictors’ optimization.

  • The intervention is simple and applicable to different model families. Results on Koopman-AE [1], PLDM [2], and STORM [3] provide preliminary evidence that regulating early dynamics feedback can affect downstream performance across different learning setups.

[1] Mori–Zwanzig Latent Space Koopman Closure for Nonlinear Autoencoder. Priyam Gupta et al. (2025)

[2] Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models. Uladzislau Sobal et al. (2025)

[3] STORM: Efficient Stochastic Transformer Based World Models for Reinforcement Learning. Weipu Zhang et al. (2023)

Critical weaknesses
  • Prior work, including DreamerV2 [1] and Tang et al. [2], already recognizes that predictor optimization and asymmetric gradient updates affect representation learning. Although the paper acknowledges these connections, it does not clearly establish what substantive understanding its analysis adds. The central theoretical characterization follows from an optimal-value decomposition and an application of Danskin’s theorem [3]. This provides a useful description of the gradient discrepancy, but does not identify conditions under which that discrepancy harms representation learning. My concern is therefore the additional explanatory value of the analysis, rather than the correctness of the identity.

  • The comparisons do not sufficiently isolate the value of ACW’s design. That is, relevant alternatives include keeping the dynamics-to-encoder gradient at a small, fixed weight throughout training; gradually increasing the weight of the entire dynamics loss, which scales both encoder and predictor gradients; and, where applicable, pretraining the encoder and decoder on reconstruction before starting joint dynamics training. These comparisons would clarify whether ACW’s particular timing and asymmetric gradient scaling are necessary. Section 4 already tests additional predictor-only updates through Catch-up, but Section 5 does not compare ACW against this alternative under comparable computational budgets.

  • several improvements are accompanied by substantial variability. For example, STORM’s greedy return changes from $1205.3\pm156.1 \quad\text{to}\quad 1471.5\pm314.0. $ The paper highlights a 22.1% improvement, but does not report uncertainty on that improvement or the individual paired performance differences. Since the runs are paired by seed, those differences are particularly relevant.

  • Generalization across tasks within each model family remains untested. Koopman-AE is evaluated only on cylinder flow, PLDM only on Two-Rooms, and STORM only on MsPacman. Because the model and task change together across experiments, the results establish improvements on three selected model–task combinations but do not demonstrate robustness across datasets for any individual model. Repeating each experiment with three seeds addresses training variability, not this limitation in task coverage.

  • The writing and organization make the central experiments difficult to follow. Section 4 introduces several training procedures and diagnostics without a sufficiently clear explanation of their relationship. In particular, the extra predictor updates used during Catch-up training are easy to confuse with the separate optimization used to construct diagnostic references for both Joint and Catch-up checkpoints. Important notation also requires inference or consultation of the appendix: the Adam update vectors (Δϕcur\Delta\phi_{\mathrm{cur}}) and (Δϕref\Delta\phi_{\mathrm{ref}}) are introduced only implicitly, (IrϕI_{r_\phi}) is not explicitly identified as conditional mutual information when introduced, and the meaning of wrapped (θ1\theta_1) is explained in the appendix. Since interpreting these experiments is central to assessing the contribution, the main text needs a clearer experimental overview and explicit definitions at first use.

[1] Mastering Atari with Discrete World Models. Danijar Hafner et al. (2021)

[2] Understanding Self-Predictive Learning for Reinforcement Learning. Yunhao Tang et al. (2023)

[3] The Theory of Max-Min and Its Application to Weapons Allocation Problems. John M. Danskin. (1967)

Questions for authors

See weaknesses.

2 comments

We thank the reviewer for the constructive feedback. We have revised the manuscript and added experiments to address the concerns below.

W1: Mechanistic Novelty and Explanatory Value

We acknowledge prior insights from DreamerV2 and Tang et al. on predictor optimization and gradient asymmetry; our analysis uses standard mathematical tools. We distinguish the inductive bias specified by the dynamics model class and fitting objective from additional representation feedback caused by predictor suboptimality. This matters because the latter introduces optimization-state-dependent influence beyond the intended bias, with unknown consequences. Controlled Joint/Catch-up experiments and matched transition refits show that this influence changes representation formation and persists as differences in the dynamical quality supported by the final representation. We do not presume it to be universally harmful; rather, we isolate and empirically investigate its effects.

W2: ACW Design Ablations and Computational Cost

To examine whether ACW's gains depend on its specific coupling design, we added three ablations in the double-pendulum sequential VAE (Section 5.4): Fixed coupling (γ=0.2 throughout), Symmetric warm-up (scaling both representation- and predictor-side dynamics gradients), and Pretraining (optimizing non-transition objectives before joint training). All use three seeds and matched fresh transition refits.

Matched-refit latent NMSE ↓ (three-seed mean; Table 3)

MethodH=10H=25H=50H=100
Joint0.02330.25550.91611.5510
Catch-up0.02080.18570.76921.3845
ACW0.01520.16800.73141.3860
Fixed coupling0.04640.35801.04831.5620
Symmetric warm-up0.01650.18140.78431.4474
Pretraining0.01910.21310.84321.4655

Fixed coupling underperforms Joint, while the other two improve upon Joint but remain below ACW. These results support temporarily attenuating representation feedback while maintaining predictor optimization.

To assess schedule sensitivity, we also varied the attenuation onset and full-coupling restoration across nine configurations in the same setting. All improved upon Joint at four horizons (Figure 6; Appendix D.3).

We further compared ACW with Catch-up in performance and computational cost (Section 5.4; Appendix D.3). Under 10,000 outer updates, ACW achieves lower matched-refit NMSE at three horizons and comparable performance at H=100. For seed 1101, Catch-up requires 580,744 transition updates / 25,813 s, compared with ACW's 10,000 / 793 s (recorded training-loop times, excluding refitting/evaluation).

W3/W4: Robustness Across Seeds and Tasks

To address W3's concern about variability and paired improvements, we expanded PLDM and STORM to six paired seeds. The tables below report individual differences (ACW − Native) and their mean ± SD. To address W4's concern about within-model task generalization, we further evaluated STORM on Qbert, Alien, and BankHeist alongside MsPacman, testing multiple tasks within the same model family (Sections 5.2–5.3; Appendix C.2–C.3).

PLDM: Per-seed paired differences

SeedΔ Success rate ↑Δ Planning MSE ↓
101−0.01+17.52
102+0.11−65.06
103+0.17−57.91
2019+0.11−41.34
2021+0.16−24.20
2023+0.10−43.97
Mean ± SD+0.107 ± 0.064−35.83 ± 29.73

STORM: Per-seed paired return differences

SeedΔ MsPacman ↑Δ Qbert ↑Δ Alien ↑Δ BankHeist ↑
0+79.5+3421.25+276.0−76.5
1+386.5−410.00+9.5+314.0
2+332.5+3421.25−64.0+343.5
3+489.0+2666.25+67.5−95.5
4+645.0−2160.00+13.0+30.5
5+164.0+2492.50+41.0+82.5
Mean ± SD+349.4 ± 207.7+1571.9 ± 2312.6+57.2 ± 115.9+99.7 ± 189.5
Mean relative gain+26.6%+72.4%+8.4%+10.8%

The results show improvements in five of six PLDM seeds on both metrics and positive mean gains across all four STORM games, while explicitly revealing seed variability. Full Native/ACW scores and condition SDs are provided in Tables 10 and 12 (Appendix C.2–C.3), with task-level summaries in Table 2. Each STORM seed result averages 20 evaluation episodes.

W5: Experimental Clarity and Notation

We have revised the relevant passages to clarify the controlled experiments and their notation:

  • Experimental procedures (Section 4.1): We now clearly distinguish Catch-up's additional transition updates during training from retrospective reference fitting used solely for diagnostics, which leaves both training trajectories unchanged. We also explicitly define Δϕcur\Delta\phi_{\mathrm{cur}} and Δϕref\Delta\phi_{\mathrm{ref}} as hypothetical next-step updates computed using the same checkpointed Adam state.
  • Notation (Proposition 2; Figure 2): We clarify that IrϕI_{r_\phi} denotes conditional mutual information and that θ1\theta_1 is the first-link angle wrapped to (−π,π](-\pi,\pi].
AuthorsOpenReview ↗

A Gentle Reminder About Our Response

Dear Reviewer EnkT,

Thank you again for your time and review. As the discussion period closes on November 18 (AoE), we wanted to kindly check whether you have had a chance to read our response and revised manuscript.

Following your comments, we clarified our contribution relative to prior work and added several experiments. In particular, we (1) added gradient analyses and controlled experiments to examine the effect of ACW; (2) compared ACW with alternative coupling designs (ablation) and Catch-up, including computational cost; and (3) expanded STORM to four Atari games with six seeds each, and PLDM to six seeds. We also clarified the experimental procedures and notation in the manuscript.

We would appreciate it if you could take another look at our response and consider whether the revisions address your concerns. We would also be glad to clarify anything that remains unclear.

Reviewer_Qy3wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how suboptimal dynamics models affect representation learning during joint optimization. It provides a gradient decomposition and proposes Asymmetric Coupling Warm-up (ACW), evaluated across several dynamical learning settings.

Critical strengths

The research question is relevant.

The experiments cover several learning paradigms and report promising improvements.

Critical weaknesses

Very limited theoretical novelty. Proposition 1 follows almost directly from Danskin's theorem, while Proposition 2 is a standard information-theoretic decomposition. It does not meet the standards of a paper which claims to provide a theoretical characterization.

Lack of mathematical rigor in the exposition. Important quantities, such as g_\phi^{\mathrm{dyn}}, are initially introduced informally rather than precisely defined (only later in the paper is a specific definition given and presented as a particular case). Some notations are also insufficiently explained (for example I_{r_\phi} which should be clearly introduced as conditional mutual information).

Questions for authors

What is the theoretical novelty beyond standard applications of Danskin's theorem and information-theoretic identities?

2 comments

We thank the reviewer for the thoughtful feedback. We address the concerns below.

W1/Q1: Theoretical Novelty and Mechanistic Contribution

We acknowledge that Proposition 1 follows from Danskin's theorem and Proposition 2 builds on standard information-theoretic identities. Our contribution does not lie in deriving new mathematical identities, but in using these tools to characterize a previously underexamined mechanism: the distinction between the inductive bias prescribed by the chosen dynamics model class and fitting objective, and the additional representation feedback induced by predictor suboptimality. The latter introduces optimization-state-dependent influence whose consequences for representation learning are not understood a priori.

Through controlled Joint/Catch-up experiments and matched transition refitting, we further demonstrate that this influence not only alters the training process but also has lasting consequences for the dynamical quality of learned representations. Building on this mechanistic insight, we propose ACW, a lightweight intervention that regulates early dynamics-to-representation coupling without requiring additional predictor optimization, and demonstrate its effectiveness across three learning paradigms.

Thus, the contribution of our work lies in mechanistic characterization, controlled validation, and the resulting algorithmic intervention, rather than mathematical identities alone.

W2: Mathematical Definitions and Notation

We appreciate the reviewer's suggestion regarding mathematical clarity. In Section 3, we initially introduce gϕdyng_\phi^{\mathrm{dyn}} as the dynamics-induced contribution to the representation gradient under the model's native gradient routing, rather than directly restricting it to ∇ϕEfit\nabla_\phi E_{\mathrm{fit}}. This is because the objectives and gradient-routing mechanisms differ across learning paradigms. For example, in STORM, the dynamics-related representation gradient arises from the complete weighted KL objective with asymmetric stop-gradient routing, rather than the ordinary gradient of a single predictor-fitting objective (Appendix C.3).

Section 3.1 therefore considers the specific setting gϕdyn=∇ϕEfitg_\phi^{\mathrm{dyn}}=\nabla_\phi E_{\mathrm{fit}}, under which the gradient decomposition is established. This organization distinguishes the general cross-paradigm concept from the specific conditions under which the theoretical result holds.

Additionally, we have revised Proposition 2 to explicitly clarify that IrϕI_{r_\phi} denotes conditional mutual information under rϕr_\phi.

AuthorsOpenReview ↗

A Gentle Reminder About Our Response

Dear Reviewer Qy3w,

Thank you again for your time and feedback. As the discussion period closes on November 18 (AoE), we wanted to kindly check whether you have had a chance to read our response.

In brief, we clarified that our main contribution lies in identifying and validating the effects of predictor suboptimality on representation learning, and in developing ACW as a lightweight, mechanism-guided intervention, rather than deriving new mathematical identities. We also explained why the dynamics-induced gradient is introduced generally before considering a specific mathematical setting, and added an explicit definition of the conditional mutual information notation following Proposition 2.

We would appreciate it if you could review our response and consider updating your assessment if your concerns have been addressed.

Reviewer_AWRbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies models that learn a representation and its dynamics jointly, and shows that a dynamics model that is not yet well fitted adds an extra, unintended signal to representation learning, separate from the intended inductive bias of the dynamics model. A controlled sequential-VAE study shows that reducing this suboptimality changes the learned representation and improves rollout accuracy. The authors then propose Asymmetric Coupling Warm-up (ACW), which attenuates the dynamics-to-representation gradient early in training, and report gains on Koopman-AE, PLDM and STORM.

Critical strengths
  1. In Joint and Catch-up runs share the model seed, outer minibatches and reparameterization noise. The matched transition refit (Table 1), with identical fresh initializations and validation-based model selection, isolates the quality of the learned representation from the quality of the final transition model. This provides direct causal evidence, within this setting, that dynamics-parameter suboptimality during training alters the learned representation.
  2. Separating the intended inductive bias of the dynamics model from the additional effect of dynamics-parameter suboptimality is a precise and general formulation of a phenomenon that is otherwise addressed heuristically (stop-gradient targets, asymmetric KL routing, staged training). The proposed intervention is evaluated on three paradigms with different dynamics parameterizations (linear operator, deterministic recurrent predictor, stochastic Transformer) and different downstream uses (prediction, planning, control).
Critical weaknesses
  1. The framework implicitly treats the representation gradient under fully fitted dynamics as the desirable learning signal, and the suboptimality-induced part as a deviation from it. This assumption could be discussed further: in self-predictive and JEPA-style models, following the fitted-dynamics gradient can favour collapsed or low-information representations [1], which is why models such as PLDM rely on additional variance/covariance regularizers. Clarifying when this signal is beneficial would strengthen the framing.
  2. ACW is not tested in the double-pendulum sequential-VAE setting, where the mechanism is measured and Catch-up provides a natural reference. Showing that ACW reproduces the Catch-up effects there (update discrepancy, CKA, refit NMSE) would connect the two parts of the paper.

[1] Y. Tian, X. Chen, S. Ganguli. Understanding self-supervised learning dynamics without contrastive pairs. ICML 2021.

Questions for authors
  1. In the diagnostic of Fig. 5b, when the coupling is fully attenuated, the ACW distance to the reference gradient equals the norm of the reference gradient itself. what is the norm of the reference gradient and the cosine similarity between the Native and reference gradients, to show whether ACW improves the direction of the update or mainly reduces its magnitude?
  2. The standard deviations in Table 2 overlap considerably with three seeds, and STORM is evaluated on a single Atari game. Could the authors report per-seed or paired results, results with additional seeds and games, and the sensitivity of the results to the ACW warm-up schedule?
3 comments

We thank the reviewer for the feedback. We address the concerns below with additional analyses, experiments, and manuscript revisions.

W1: Is the Fitted-Dynamics Gradient Necessarily Beneficial?

We agree that the fitted-dynamics gradient is not necessarily beneficial. It characterizes the intended inductive bias prescribed by the dynamics model class and fitting objective, rather than an inherently desirable signal. Under joint training, predictor suboptimality introduces an additional influence that is not part of this intended bias, yet still shapes representation learning. We seek to isolate this influence to distinguish the effects of dynamics modeling from those of predictor suboptimality and investigate its consequences empirically.

W2/Q1: Mechanistic Analysis and Controlled Validation of ACW

We revised Section 5.4 and Figure 5 to analyze full representation gradients in PLDM and the double-pendulum sequential VAE (Appendix D.1). At matched Native checkpoints, ACW reduces both Euclidean and cosine distances to a better-fitted reference during early training. In PLDM at 9.2% progress, cosine distance decreases from 0.949 to 0.096 while the full-gradient norm increases from 0.534 to 1.130. Thus, the improved alignment is not merely due to magnitude reduction (Table A).

We also added ACW to the original double-pendulum study alongside Joint and Catch-up (revised Section 5.4, Table 3, Appendix D.2). Final Joint–ACW CKA is 0.742 ± 0.095 (new Figure 8). After matched transition refitting, ACW reduces latent NMSE by 10.6%–35.0% versus Joint, outperforming Catch-up at three horizons and achieving comparable performance at H=100 (Table B). These results connect ACW's gradient-level effects with representation formation and dynamical quality.

Table A. Full-gradient diagnostics (three-seed means; distances to reference, ↓).

SettingProgressNorm: Native / ACW / Ref.Cosine distance: Native → ACWL2: Native → ACW
PLDM1%3.129 / 2.241 / 1.8181.014 → 0.0373.641 → 0.690
PLDM9.2%0.534 / 1.130 / 1.4670.949 → 0.0961.537 → 0.653
Pendulum1%507.165 / 508.231 / 507.7010.001151 → 0.00002924.469 → 3.735

Cosine distance = 1 − cosine similarity, as in revised Figure 5.

Table B. Double-pendulum matched-refit latent NMSE (↓; mean ± SD, three seeds).

HJointCatch-upACW
100.0233 ± 0.00220.0208 ± 0.00590.0152 ± 0.0038
250.2555 ± 0.00890.1857 ± 0.01900.1680 ± 0.0330
500.9161 ± 0.01050.7692 ± 0.03700.7314 ± 0.0826
1001.5510 ± 0.08061.3845 ± 0.04451.3860 ± 0.0782

Two matched transition refits per seed.

Q2: Robustness Across Seeds, Tasks, and Coupling Schedules

We expanded PLDM to six seeds and STORM to four Atari games with six paired seeds each (revised Table 2, Appendix C). ACW improves both PLDM metrics in 5/6 seeds and mean STORM returns in all four games (Tables C–D).

We also tested nine coupling schedules in the double-pendulum VAE (new Figure 6, Appendix D.3). All improve matched-refit latent NMSE over Joint at all four horizons, with gains of 9.5%–42.4% (Table E).

Table C. Koopman-AE and PLDM paired results (Native → ACW).

ModelSeedMetric 1Metric 2
Koopman991.8854 → 0.93122.0463 → 1.2555
1001.3179 → 0.79210.7655 → 0.7120
1010.7176 → 0.72461.1838 → 1.5165
PLDM1010.55 → 0.54147.29 → 164.81
1020.42 → 0.53242.89 → 177.83
1030.44 → 0.61146.43 → 88.52
20190.55 → 0.66136.93 → 95.59
20210.48 → 0.64142.28 → 118.08
20230.45 → 0.55160.41 → 116.44

Koopman: 20-step NMSE (×10⁻⁴, ↓) / 1000-step NMSE (×10⁻², ↓). PLDM: success rate (↑) / planning MSE (↓). Mean ± SD in revised Table 2.

Table D. STORM greedy returns, six paired seeds (Native → ACW, ↑).

SeedMsPacmanQbertAlienBankHeist
01029.5→1109855→4276.3610→8861202→1125.5
11259→1645.54150→3740674.5→684826→1140
21327.5→16601262.5→4683.8720.5→656.5761→1104.5
31602→20911335→4001.3601.5→669980.5→885
41330→19754257.5→2097.5712.5→725.5898→928.5
51343→15071171.3→3663.8743.5→784.5878.5→960.5
Gain+26.6%+72.4%+8.4%+10.8%
Improved6/64/65/64/6

Each seed averages 20 evaluation episodes. Mean ± SD in revised Table 2.

Table E. Schedule sensitivity: matched-refit NMSE improvement over Joint (%). Within each cell, values correspond to full-coupling steps 1100 / 2000 / 2900.

StartH=10H=25H=50H=100
036.2/36.8/26.837.3/41.8/36.117.9/23.1/18.69.5/10.0/10.9
20037.8/35.0/21.942.4/34.2/33.820.5/20.2/17.69.8/10.6/10.6
50032.4/34.3/22.131.9/37.5/31.016.6/17.8/15.59.8/11.4/10.6

Default schedule: start = 200, full coupling = 2000. Three seeds and two refits per seed.

Reviewer_AWRbOpenReview ↗

Response to the rebuttal

I thank the authors for their detailed response and for additional experiments, which address most of my concerns. It is not clear from the response which parts of the manuscript have been revised. It would help if the authors could highlight the changes in the revision using different colors, so that the new text and results could be easily identified. But given these revisions, I would be happy to raise my score after discussion with the other reviewers.

Thank you for considering raising your score. We look forward to your discussion with the other reviewers. Your constructive comments have been very helpful in improving our work. Following your suggestion, we have highlighted the revised and newly added content in blue throughout the manuscript and posted a separate comment summarizing the main revisions.

Comments on the paper

3 comments

Dear Reviewers, the authors’ responses are now available. If you haven’t already, please review them and share your comments before the discussion period closes. Thank you for your time and effort.

AuthorsOpenReview ↗

Manuscript Revision and Highlighted Changes

We have revised the manuscript in response to the reviewers' comments. The main changes are summarized below:

  • Clarifications and presentation: We refined the Abstract, Introduction, and Conclusion to better reflect the contributions and findings. We also clarified the reference-based diagnostic protocol and the definition of representation updates in Section 4.1.

  • Expanded experimental evaluations: We added training seeds for PLDM and STORM, extended STORM to four Atari games, and updated the corresponding results and discussions in Sections 5.2–5.3 and Appendix C.

  • Additional mechanism analyses and ablations: We substantially revised Section 5.4, adding full representation-gradient analyses (Figure 5), controlled validation and comparisons with alternative training strategies (Table 3), schedule sensitivity analysis (Figure 6), and computational cost comparisons. The corresponding experimental details and additional results are provided in Appendix D.

To make these revisions easier to identify, we have now highlighted the newly added and substantially modified content in blue throughout the manuscript.

We thank the reviewers and AC for their feedback and consideration.

AuthorsOpenReview ↗

Final Clarification for the Area Chair

Dear Area Chair,

We sincerely thank you and the reviewers for your time and consideration. We especially appreciate Reviewer AWRb's constructive feedback and willingness to consider raising their score after further discussion. However, we have not yet received follow-up responses from the other two reviewers. As the discussion period approaches its end, we are concerned that our substantial revisions and additional evidence may not be fully reflected in the final assessment.

We would therefore like to briefly clarify our contributions: (1) mechanism identification and formalization, (2) controlled experimental investigation, and (3) mechanism-guided algorithmic intervention. Regarding the concerns about theoretical novelty raised by Reviewers EnkT and Qy3w, we emphasize that established mathematical tools are used to rigorously characterize the identified mechanism, rather than to claim new mathematical identities. Our novelty lies in isolating and characterizing the influence of dynamics-parameter suboptimality, experimentally validating its consequences, and developing ACW based on these findings.

We hope the Area Chair will consider the revised manuscript and our responses in assessing the work as a whole. Thank you again for your consideration!

Sincerely, 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.