acceptodds
Under review as a conference paper at ICLR 2027

Early Signatures of Memorization in Diffusion Models via Basin Geometry and Cyclic Denoising

Abstract

Diffusion models generalize early in training and later reproduce individual training samples. Standard tests detect memorization only once one-shot generation already produces near-copies, which leaves a released model unaudited until its outputs fail. We show that memorization is encoded in the geometry of the learned energy landscape before it appears in generated samples, a state we call latent memorization. Using score divergence and basin volume, we find that localized basins form around individual training samples and separate them from held-out samples before the first memorized sample appears, with an onset that follows the same scaling as the memorization time. We probe these basins with cyclic denoising, the repeated application of partial noising followed by denoising. Under the exact empirical score, we prove that cycling started near an isolated training sample recovers it and returns to it over any finite number of cycles with high probability, with a bound controlled by the cycling noise and the separation from competing samples. In trained models, cycling recovers training images from CelebA and CIFAR-10 checkpoints whose one-shot samples contain no copies, and at a CelebA checkpoint with 0.1% one-shot copies, 500 cycles raise the memorized fraction above 30%. Cycling also reveals degenerate attractors that match no single training image. They are prevalent early in training and fade as training proceeds, so residence in a basin does not by itself imply memorization. These findings hold on a Gaussian mixture, CelebA, and CIFAR-10 across optimizers, architectures, noise schedules, and training-set sizes. They also extend to off-the-shelf Stable Diffusion v1.4, where the cycled conditional–unconditional divergence gap separates memorized from non-memorized prompts with an AUC of and a TPR of at FPR. Altogether, probing the learned landscape uncovers memorization that eludes one-shot generation. More broadly, what a diffusion model has memorized is a property of the geometry and stability of its learned distribution, and assessing it requires examining this structure rather than generated outputs alone.

Primary area:
interpretability and explainable AI
Keywords:
Diffusion models, Interpretability, Memorization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_eS5COpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper argues that memorization in diffusion models is encoded in the geometry of the learned landscape before one-shot sampling produces copies of the training data (“latent memorization”). The authors propose to study score divergence and a basin-volume proxy to unveil these phenomena. Cyclic denoising (CD) recovers training images from checkpoints whose samples contain no copies. It also reveals “degenerate attractors” that match no training image. A simple theorem bounds residence under the exact empirical score. The approach transfers to Stable Diffusion v1.4 as a memorization detector.

Critical strengths

The question is relevant for memorization studies, and the latent-memorization picture is interesting.

The procedure and the insights gained from it are new and valuable.

The experiments are broad and consistent across datasets, architectures, optimizers and schedules.

The paper is honest about limitations, including the comparison with previous works.

Critical weaknesses

The theory part is weak but that is not a critical weakness per se. The numerical experiments is insightful and enough for the conclusions.

Questions for authors

Fig. 1C is misleading. It shows a single energy landscape, but the energy E_t is time-dependent. A CD cycle goes through the whole family of landscapes. Large γ does not necessarily mean barrier hopping; it means the training-point basins have merged at that noise level. Propose an alternative figure.

The critical amplitude is not well understood. Both the γ_c of the basin-volume sweep and the amplitude where residence disappears are treated phenomenologically. What fixes its scale and what does it mean in terms of the scenario put forward in this work? How does it evolve during training?

3 comments
AuthorsOpenReview ↗

Response to Reviewer eS5c(Part 1/1)

We address each point below, using W for weaknesses and Q for questions.

Q1. One fixed landscape suggests that larger noise pushes an image over a static barrier. The figure should show how the landscape itself changes with diffusion time.

We propose the updated Fig. 1C in the revised manuscript, using a 3-dimensional schematic with γ\gamma on an explicit axis to remove the misleading ambiguity in the previous illustration. The additional axis makes the noise dependent landscapes explicit.

We agree with the reviewer’s interpretation that larger γ\gamma does not necessarily imply barrier crossing. We retain the paths in the figure to illustrate possible residence and transitions between basins. They are not intended to imply that a larger cycling amplitude guarantees a transition.

W1/Q2. What controls the noise scale for recovering a target in one reverse pass versus remaining near it through repeated cycles? How do those scales move as training proceeds? Extension to Theory.

We agree that these amplitudes are characterized phenomenologically in our experiments. They mark the measured limits of recovery or residence, rather than theoretically established transition points. Their scale depends on how the noise schedule converts amplitude into perturbation strength and how strongly the learned reverse dynamics return perturbed states to a basin (Section 2.2, lines 212–224). Theorem 1 relates recovery and residence to noise strength relative to training-sample separation under exact empirical reversal. The new inexact-score theorem in Appendix K (Theorem 2, lines 2559–2569) instead relates residence to the margin between the neighbourhood from which the reverse pass returns and the smaller region containing the cycle endpoints. The measured thresholds also depend on the recovery tolerance and observation horizon (Section 2.1, lines 204–209; Appendix H, lines 2214–2225). The basin-volume sweep tests single-pass recovery, whereas cycling tests repeated persistence, so their thresholds need not coincide.

During training, Figure 2C shows increasing recovery-volume proxies for training samples, while held-out samples separate. Section 3.3 and Figure 3 show how residence depends on checkpoint and cycling amplitude. Together with the higher-amplitude observations in Section 3.3 (lines 423–426), these results indicate that training-associated basins become more robust and can sustain residence at higher amplitudes as training proceeds and memorization increases. A higher recovery threshold means that a sample can withstand stronger noising and still be reconstructed; residence at higher amplitudes means that repeated perturbations still return the trajectory to the same basin. The theoretical results give sufficient conditions for these behaviours; predicting how the measured thresholds evolve during training remains open.

I thank the authors for the reply. I find the revised figure satisfactory. I keep my good score.

AuthorsOpenReview ↗

Thank you!

Dear Reviewer eS5C,

Thank you so much for your fruitful suggestions and support for the manuscript.

Thank you, Authors.

Reviewer_T4nJOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether memorization in diffusion models can be detected before it becomes visible through one-shot generation, introducing the notion of latent memorization. The authors show that localized basins around individual training examples emerge in the learned score field before generated samples reproduce those examples, using score divergence and a recovery-based basin-volume proxy to characterize this transition. They further use cyclic denoising (CD) as a dynamical probe that repeatedly perturbs and denoises samples, revealing training-sample attractors that ordinary sampling may not reach; notably, CD recovers memorized images even at checkpoints with zero or negligible one-shot memorization. The paper complements these observations with an idealized theoretical result showing that, under exact empirical-score reversal, sufficiently isolated training points exhibit high-probability repeated recovery depending on cycling noise and inter-sample separation. The phenomenon is demonstrated across a Gaussian mixture model, CelebA, CIFAR-10, multiple architectures, optimizers, noise schedules, and dataset sizes, and is further validated on Stable Diffusion v1.4, where cycling substantially strengthens a conditional–unconditional divergence signal for detecting memorized prompts. Overall, the paper provides a compelling geometric and dynamical account of how memorization develops before it becomes apparent from generated outputs.

Critical strengths

1. A strong conceptual contribution to audit memorization.
The most consequential contribution is the demonstration that memorization can become encoded in the geometry of the learned distribution before it is visible through one-shot generation. This distinction between latent and expressed memorization is interesting, as well as practically relevant: output-only audits can declare a checkpoint apparently clean while sample-specific basins have already formed. The agreement between the temporal emergence of score-field geometry, its O(n)O(n) scaling, and the later onset of explicit memorization makes this considerably stronger than an isolated empirical correlation.

2. Comprehensive and consistent validation.
The empirical, dynamical, and theoretical components reinforce one another. Score divergence and basin volume identify emerging geometric structure; cyclic denoising demonstrates that this hidden structure is dynamically accessible; the exact-score analysis explains why sufficiently isolated training points can exhibit persistent residence. Off-the-shelf Stable Diffusion checkpoint shows a practical carry-over. Together with detailed supplementary material and released code, reproducibility of this work appears good.

Critical weaknesses

I have minor concerns.

1. The distinction between detecting latent memorization and attributing it to a particular training example could be stated more sharply. Establishing that a particular attractor corresponds to a specific memorized training example still requires candidate training data. I think the paper would benefit from a slightly clearer statement separating model-level detection of emerging memorization from sample-level attribution. This is primarily a matter of framing and does not require additional experiments.

2. The role of the theoretical result relative to the latent-memorization regime could be clarified. I would encourage the authors to state more explicitly which aspects of the empirical regime - which occurs before complete memorization under a learned approximate score - the theorem is intended to explain (in the Stable Diffusion experiments), and which remain outside its scope.

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer T4nJ(Part 1/1)

We address each point below, using W for weaknesses and Q for questions.

W1. The distinction between detecting latent memorization and attributing it to a particular training example could be stated more sharply.

We agree and have sharpened the framing in the Introduction and Section 3.3. We now explicitly distinguish model-level evidence of emerging memorization from sample-level attribution. Geometric probes and cycling can reveal stable attractors without training set access, but residence alone does not confirm memorization because degenerate attractors also occur. Establishing that an attractor reproduces a specific training example requires candidate training data.

Our empirical observations nevertheless suggest a useful detection signal. Across the trajectories we examined at different checkpoints (including those within the generalization window), resident attractors that retained recognizable high-level features at higher cycling amplitudes matched memorized training samples. The complementary work of Sharma and Martiniani [1] demonstrates the utility of unconditional cycling in off-the-shelf models, including Stable Diffusion v1.4. Larger cycling amplitudes reveal structured memorized attractors, whereas low amplitudes can produce trivial states. Their procedure identifies candidates through sustained residence and high similarity between consecutive cycles, but verification still uses reverse-image search or previously identified memorized training examples.

Our Stable Diffusion v1.4 experiments provide quantitative evidence specifically for detecting memorized prompts. The cycled conditional–unconditional divergence gap reaches an AUC of 0.944 and a TPR of 0.866 at 1% FPR on the evaluated prompt benchmark. Computing this score requires no training images. These results establish detection performance in the conditional setting, while our observations and results motivate unconditional detection through cycling and basin geometry. Quantifying the reliability of such unconditional detection remains future work.

[1] R. Sharma and S. Martiniani. Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models. arXiv:2606.24000, 2026. Paper.

W2. Theorem Extension and Intepretation

Theorem 1 bounds recovery and residence under exact empirical reversal; it does not directly cover a learned approximate score. We have therefore added a separate conditional residence theorem and proof in Appendix K, Conditional residence under an inexact score.

The key condition is local: the complete deterministic reverse pass returns every input within a larger neighbourhood of a reference state to a smaller neighbourhood. This allows reconstruction error and does not require the model to reproduce the full empirical training distribution. For neighbourhood radii R>rR>r and additive forward-noise variance qq, a trajectory starting in the smaller neighbourhood remains there at all MM clean cycle endpoints with probability at least

\qquad Z\sim\mathcal N(0,I_d).$$ Each forward perturbation $\sqrt q\,Z$ of norm at most $R-r$ keeps the noisy state within the larger neighbourhood, after which the reverse pass returns it to the smaller one. Independence of the noise draws gives the bound. In the appendix, we establish the return condition throughout the neighbourhood for a constructed inexact score and compare the bound with simulated residence. **For Stable Diffusion, this provides a conditional explanation of persistence after recovery**, as illustrated by the `hippo` chain remaining at its matched image through $c_{200}$ (§3.4). Applying the result requires a fixed prompt, guidance setting and deterministic reverse sampler in latent coordinates. We have not established the uniform return condition for Stable Diffusion, so we claim no numerical guarantee there. An informative guarantee would require a return margin large enough relative to the noise, with the required margin depending on the latent dimension, cycle count and desired confidence. **Outside its scope.** The new theorem gives sufficient conditions for residence; recovery guarantees in Theorem 1 require exact empirical reversal. Neither result quantitatively predicts when the local return condition develops during training, which distant initializations reach a particular basin, or the arrival and transition times observed in Stable Diffusion. Recovering training images from initially non-copying Stable Diffusion generations is therefore supported by our experiments rather than established by these theorems. **We believe the revision addresses each of your concerns. If any remain, please let us know and we will be happy to address those during the discussion period. Otherwise, we would be grateful if you would reconsider your score in light of these changes. Looking forward to your response.**

Thank you to the authors for their response. Since there are only 4 levels for the overall ranking - I would keep my score.

AuthorsOpenReview ↗

Thank you!

Dear Reviewer T4nJ,

Thank you for your engagement and for supporting the manuscript. We really appreciate it.

Thank you, Authors

Reviewer_23GDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies how memorization develops in diffusion models before copies are detected in generated samples. It tracks score divergence statistics and a recovery based basin volume proxy across training, and uses cyclic denoising to probe the accessibility and stability of training-image attractors. The authors report that geometric separation between training and held-out samples precedes observed one-shot copying, with an onset that approximately scales with dataset size. Cycling substantially increases the fraction of outputs classified as memorized at some checkpoints. The paper also identifies stable degenerate attractors and proves a local recovery and finite-horizon residence bound under exact reversal of the empirical distribution. Experiments cover a Gaussian mixture, CelebA, CIFAR-10, and an application to image recovery and prompt-level memorization detection in Stable Diffusion v1.4.

Critical strengths
  1. The longitudinal experiments provide a useful connection between geometric structure and extractable memories. Following training, held-out, and generated samples across checkpoints makes it possible to examine when the geometric signatures emerge relative to observed copying. The experiments span synthetic and image data, different architectures and optimizers, and several cycling amplitudes. In particular, recovering training images from checkpoints with zero observed one-shot copies is an interesting empirical finding that motivates looking beyond ordinary sample inspection.

  2. The paper carefully distinguishes basin residence from training-data memorization. The nontraining attractors demonstrate that high consecutive-cycle similarity is insufficient evidence of copying, which is consequential for interpreting a model-based audit. The local residence theorem appears sound under its stated assumptions, and the authors explicitly acknowledge that it does not explain discovery from distant initializations or establish the same guarantee for learned scores. The appendix also reports informative uncertainty estimates and limitations for the Stable Diffusion comparison.

Critical weaknesses
  1. Insufficient experimental validation and interpretation. Although the paper shows that CD can substantially increase the memorization fraction, it lacks a comparison with independent sampling under the same computational budget, making it difficult to assess the practical advantage of CD. Moreover, Section 3.2 reports that generated samples have almost the same score divergence as training samples, even when one-shot memorization is not observed. I think this observation is interesting, but more explanation and analysis are needed.

  2. Limitations in presentation. I found some parts of the paper difficult to follow. Several notations and quantities are not clearly introduced before use in the main paper, and the CD procedure could be presented more systematically.

Questions for authors
  1. Comparison with independent sampling. Have the authors compared CD with independent one-shot generation under the same computational budget? Would generating additional independent samples with an equivalent number of denoising function evaluations recover a comparable number of memorized training images?

  2. Geometric similarity between generated and training samples. In Section 3.2, why does the score divergence of generated samples closely follow that of training samples, even when the one-shot memorization fraction is zero?

  3. Presentation and notation. Could the authors provide a clearer and more self-contained description of the CD algorithm, including its forward perturbation, reverse sampling procedure, and cycling amplitude? It would also help to define important notation more clearly before it is used and clarify the relationships among different geometric and dynamical quantities. For example, τ\tau is not explicitly defined until Appendix A, and the mathematical definitions of A^(x,t)\hat{\mathcal{A}}(x,t) could be presented more clearly in Section 2.1. I also noticed some notation inconsistencies. On page 7, the definition of ΔA1\Delta\mathcal{A}_1 uses ϵ\epsilon for the unconditional condition, while the accompanying explanation refers to ϕ\phi. Moreover, Appendix A uses ΔA^\Delta\hat{\mathcal{A}}, which is inconsistent with the notation in the main text. In addition, the definition of residence onset in Section 2.2 could be more precise, particularly regarding the early-cycle and late-cycle levels and the meaning of sustained crossing. There are also some formatting inconsistencies. For example, some mathematical symbols such as A\mathcal{A} are unexpectedly bolded on page 4, and the formatting of paragraph headings switches between bold and italic styles starting from page 6.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 23GD(Part 1/3)

We address each point below, using W for weaknesses and Q for questions.

W1/Q1. Comparison with independent sampling.

We ran a compute-matched experiment on our CelebA model (n=1024n=1024, U-Net, SGD), comparing independent one-shot generation, Cyclic Denoising or CD (512 trajectories ×\times 500 cycles, γ=0.1\gamma=0.1), and a restart baseline. Our results show that CD recovers 1.7–6.0$\times$ as many distinct training images compared to independent sampling at equal compute. We discuss the details below.

Each method uses about 25.7 million denoiser evaluations (equal to 256,512 independent DDIM samples) and the same memorization criterion, d1/d2<1/3d_1/d_2<1/3. Full details are in Appendix L of the revised paper.

Distinct training images recovered (each image counted once)

Checkpoint (training-log one-shot fmemf_{\mathrm{mem}}, mean ± s.d.)IndependentCDRestarts
95k (0.0% ±\pm 0.00%)12726
100k (0.0% ±\pm 0.02%)26879
115k (0.1% ±\pm 0.10%)9115331

Cumulative fraction of memorized outputs (including repeated copies and initialization)

CheckpointIndependentCDRestarts
95k0.008%9.77%0.05%
100k0.028%10.08%0.40%
115k0.10%24.13%0.68%

Near the onset of memorization, CD recovers 1.7–6.0$\times$ as many distinct training images as independent sampling at equal compute. Because each image is counted once, repeated copies do not explain the gap. After only 50 cycles (about 10% of the budget), CD already exceeds independent sampling at full budget. Appendix L discusses the experiment in detail.

AuthorsOpenReview ↗

Response to Reviewer 23GD(Part 2/3)

Q2. Geometric similarity between generated and training samples.

Training-like score divergence does not require memorized generations. Across three checkpoints, 88.57–96.39% of generations lie within the central 90% of training-sample A1\mathcal A_1 values, including 92.97% at a checkpoint with fmem=0f_{\mathrm{mem}}=0. Meanwhile, the held-out fraction falls from 82.42% to 0.10%. We have added Appendix J to the revised manuscript, which includes these results. Additionally, it includes a theoretical analysis, a sample-level plot of score divergence against the relative distance (d1/d2d_1/d_2) memorization score, and visual comparisons of generations with their nearest training images (Figure 26), alongside quantitative results (Table 8). These additions complement Figure 2 and the membership-inference analysis. The theory explains how training-like score divergence can arise without near-replication. It also gives sufficient conditions for this agreement under learned, inexact scores.

For the exact empirical score, we obtain

∇x⋅st(x)=−dσt2+tr⁡Ct(x)σt4,\nabla_x\cdot s_t(x)=-\frac{d}{\sigma_t^2}+\frac{\operatorname{tr}C_t(x)}{\sigma_t^4},

where Ct(x)C_t(x) is the covariance of the training samples multiplied by αˉt\sqrt{\bar\alpha_t}, weighted by their posterior probabilities at xx. For evaluation points xgx_g and xrx_r obtained from generated and training images at the same diffusion time, ∇x⋅st(xg)−∇x⋅st(xr)=[tr⁡Ct(xg)−tr⁡Ct(xr)]/σt4\nabla_x\cdot s_t(x_g)-\nabla_x\cdot s_t(x_r)=[\operatorname{tr}C_t(x_g)-\operatorname{tr}C_t(x_r)]/\sigma_t^4. Similar covariance traces therefore give similar score divergence without requiring the underlying images to be near-copies. For a learned score, an additional residual-divergence term appears. Appendix J derives sufficient conditions for agreement between generated and training averages without assuming that individual score residuals are small.

The sample-level comparison uses CelebA with Adam and 4,096 training samples. We evaluate 1,024 generated and 1,024 held-out samples at each checkpoint. Table 8 reports:

Checkpoint τ\taufmemf_{\mathrm{mem}}Generated within training A1\mathcal A_1 central 90%Held-out within training A1\mathcal A_1 central 90%
10,0000.00%92.97%82.42%
30,0004.49%96.39%31.15%
300,00084.57%88.57%0.10%

Figure 26A–C shows how A1\mathcal A_1 varies with d1/d2d_1/d_2 for individual generations. Figure 26D shows a selected nonmemorized generation with d1/d2=0.990d_1/d_2=0.990. It is visibly distinct from its nearest training image, yet their A1\mathcal A_1 values are −711.55-711.55 and −709.44-709.44. Panels E and F show memorized generations at later checkpoints. These comparisons illustrate similar score divergence both with and without near-replication.

Regions carrying more probability mass under the learned distribution contribute more generations on average. Fig. 2 shows that generated samples share training-like geometry while held-out samples separate. This is consistent with generation through regions containing training-specific structure. The complementary work of Garnier-Brun et al. [1] describes biased generalization, where generated samples increasingly reflect features specific to the finite training set even while test loss improves. This influence can be shared across several nearby training examples without reproducing an individual example. In the complementary kernel density estimation (KDE) analysis of Merger and Goldt [2], overlapping kernels around training examples support interpolation, while sparse coverage favors retrieval of isolated examples. This provides a mechanism for novel generations to retain training-associated geometry. Similar score divergence therefore need not imply memorized generations, and when all evaluated generations remain outside the relative-distance memorization threshold, fmem=0f_{\mathrm{mem}}=0.

The same analysis in Appendix J also connects concentrated training influence to more negative divergence. As one posterior weight approaches one, the posterior covariance vanishes, so exact-score divergence approaches its lower bound −d/σt2-d/\sigma_t^2 at fixed training set and noise variance. Figure 15 shows a compatible pattern: the reported divergence estimate decreases as trajectories approach training-associated basins and enter sustained residence in memorized states. This decrease can begin before the replication onset. Thus, shared geometry can precede memorized generations, while more negative divergence can accompany residence in memorized states.

[1] J. Garnier-Brun, L. Biggio, D. Beltrame, M. Mézard, L. Saglietti. Biased Generalization in Diffusion Models. 2026. OpenReview. [2] C. Merger, S. Goldt. Local Coverage Governs Memorization in Diffusion Models. 2026. OpenReview.

AuthorsOpenReview ↗

Response to Reviewer 23GD(Part 3/3)

Q3. Presentation and notation.

We have revised Section 2.2 (lines 212–219) to describe the forward perturbation, reverse sampling, and cycling amplitude explicitly. Appendix E.3 adds the CD algorithm (lines 950–970) and an example trajectory. We now introduce τ\tau earlier (line 102) and give explicit formulas for both A\mathcal A and A^\widehat{\mathcal A} in Section 2.1 (lines 191–200). We have standardized the empty prompt notation and the conditional–unconditional difference as ΔA1\Delta\mathcal A_1 (lines 336–340 and Appendix A, line 789). Appendix H (lines 2214–2225) specifies the early and late cycle levels, sustained-crossing rule, and distinction between residence and detected onset. We have also standardized mathematical symbol formatting and paragraph headings throughout the manuscript.

We believe the revision addresses each of your concerns. If any remain, please let us know and we will be happy to address those during the discussion period. Otherwise, we would be grateful if you would reconsider your score in light of these changes. Looking forward to your response.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all the reviewers and AC

Thank you for the positive feedback and constructive suggestions. We have carefully addressed each concern and revised the manuscript accordingly. All revisions are highlighted in blue in the updated manuscript. The principal changes are summarized below:

  1. Added a compute-matched comparison with independent sampling and restarts (Appendix L). CD recovers 1.7–6.0x as many distinct training images as independent sampling across three checkpoints. Each image is counted once, so repeated copies do not explain the improvement.

  2. Extended the theory beyond exact empirical scores (Appendix K). We added a conditional residence theorem for inexact scores, supported by an analytical example and numerical tests. We also clarified its scope for learned models and Stable Diffusion.

  3. Explained why geometric similarity can precede copying (Appendix J; Figure 26 and Table 8). Added theoretical analysis, sample level measurements, and paired visualizations showing how generated samples can share training-like score divergence without being memorized copies.

  4. Clarified detection versus attribution (Introduction and Section 3.3). Revised the framing to make the distinction between model-level detection and attribution to a specific training example clearer.

  5. Clarified the CD procedure and residence criteria (Sections 2.1–2.2; Appendices E.3 and H). We revised the main-text description for clarity and added algorithm pseudocode, an example trajectory, and detailed residence-onset criteria in the appendices.

  6. Revised the landscape illustration (Figure 1C). The updated schematic includes an explicit noise-level axis to show the changing landscapes traversed during cycling.

  7. Fixed notation and formatting inconsistencies throughout the manuscript. Introduced key symbols earlier and standardised notation, mathematical formatting, and paragraph headings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.