acceptodds
Under review as a conference paper at ICLR 2027

Broadcast, Reinforce, Repeat: Attention Dynamics Behind Repetition Collapse in Diffusion Language Model Forecasting

Abstract

Repetition is a persistent failure mode of language generation, yet its mechanisms in diffusion language models (DLMs) remain poorly understood because bidirectional attention and iterative unmasking introduce dynamics absent from causal decoding. In this work, we study repetition collapse in diffusion-based time-series forecasting, a setting in which predictions can progressively lose temporal variation and converge to pathological numerical repetitions. Through systematic analysis across diffusion steps and Transformer depth, we identify a recurring attention geometry consisting of shared vertical stripes (anchor attention), shifted-diagonal predecessor feedback, and residual structure. We formalize these observations with an idealized reduced dynamical model. Rather than predicting monotonic collapse at every layer, the model shows that positive cumulative anchor–background gain across depth drives anchor concentration and, under appropriate alignment conditions, promotes cross-position representation homogenization. To test this mechanism interventionally, we construct two complementary training-free perturbations that act through distinct computational pathways: attention-level spectral SVD truncation and history-sparsified classifier-free guidance. Both interventions reduce the measured cumulative gain in the predicted direction and substantially suppress repetition while preserving forecasting accuracy. Together, these results connect observed attention geometry, cumulative collapse dynamics, and intervention-based validation into a closed mechanistic account of repetition collapse in diffusion-based time-series forecasting.

Primary area:
interpretability and explainable AI
Keywords:
diffusion language models, repetition collapse, attention dynamics, mechanistic interpretability, time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_6aFkOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies "repetition collapse" in diffusion language models applied to digit-wise time-series forecasting, evaluating LLaDA-8B and Dream-7B on the Monash benchmark. The authors identify a recurring attention geometry A(ℓ)≈1c(ℓ)⊤+γℓP+E(ℓ)A^{(\ell)} \approx \mathbf{1} c^{(\ell)\top} + \gamma_\ell P + E^{(\ell)} that combines shared anchor stripes and shifted-predecessor feedback. They derive a reduced replicator dynamics for anchor mass whose finite-depth concentration is governed by cumulative gain ∑ℓGℓ\sum_\ell G_\ell, and they obtain a cross-position homogenization corollary. They validate the mechanism with two training-free interventions, quarter-rank SVD truncation on attention and history-sparsified classifier-free guidance (HSCFG), both of which reduce measured gain by about 3% and substantially cut repetition while keeping NMAE comparable to vanilla. A Dream replication in the appendix reproduces the main qualitative findings.

Critical strengths
  • The work provides a clean, memorable empirical phenomenon. Look-ahead repetition climbs from 0.0450.045 to 0.4970.497 while draft NMAE flatlines near 1.011.01, and the slotwise H→T→UH \to T \to U cascade is quantified crisply.
  • The attention decomposition A≈1c⊤+γP+EA \approx \mathbf{1} c^\top + \gamma P + E is a non-obvious abstraction that productively ties the observed geometry to classical rank-collapse and concentration results.
  • The spectral-plateau argument genuinely justifies why low-rank SVD truncation targets the predecessor channel rather than the stripe, moving the intervention from a generic regularizer to a mechanism-targeted probe.
  • The two-pronged intervention design, with SVD acting within a forward pass and HSCFG acting between diffusion steps, is methodologically careful and the matched-pair protocol guards against obvious confounders.
Critical weaknesses
  • The quantitative link from theory to intervention is thin. The gain reductions are only about 3% with no reported variance, while HSCFG cuts value-level repetition from 1.0001.000 to 0.1320.132, so the gain pathway may not be doing most of the work.
  • The anchor-detection pipeline has several hyperparameters with no sensitivity analysis, so it is unclear how robust the measured gain curves and correlations are to those choices.
  • Causality of the slotwise cascade is not directly tested, PsameP_{\text{same}} is correctly described as not establishing causation, but no ablation with a randomized or round-robin scheduler is run to separate scheduler effects from the proposed feedback loop.
  • The forecasting-accuracy claim is operationally modest. NMAE sits near 1.01.0 across all conditions on both models, which is roughly persistence-baseline performance, and SVD/HSCFG slightly degrade NMAE on Dream.
  • Related DLM-repetition accounts (Zhao et al. 2026 on context tokens as anchors, Rulli et al. 2026 on attention sinks in DLMs) are cited but not quantitatively compared, which weakens the novelty claim given overlapping terminology.
Questions for authors
  • What are the standard errors or paired per-rollout difference distributions on the 3.22%3.22\% and 3.03%3.03\% cumulative-gain reductions and on the 4.569%4.569\% repeating-vs-non-repeating gap?
  • How sensitive are the gain measurements and gain-to-anchor correlations to the anchor-detection hyperparameters?
  • If you swap the scheduler for a randomized or round-robin policy and hold everything else fixed, does cumulative gain still rise and does repetition still emerge?
  • Does applying SVD only to shallow layers (where the stripe-plus-predecessor fit is cleanest) have a different effect than applying it uniformly or only to deep layers?
  • Are your "anchor" positions the same objects as the attention sinks and is your account complementary to or competing with Zhao et al. 2026?
3 comments

We thank the reviewer for the careful reading of our paper and the thoughtful, constructive feedback.

Weakness 1: Quantitative link between gain and HSCFG

We agree that the original comparison combines two effects of HSCFG: directly changing the selected output tokens, and allowing these changes to influence subsequent forward passes through the updated canvas. The direct effect can reduce repetition without changing the cumulative gain measured in the current forward pass.

To distinguish these effects, we conducted a new ablation. At each step, we record the HSCFG-selected tokens for output evaluation, but commit the unguided selections to the actual canvas used by subsequent forward passes. When guidance changes the selected position, we likewise record the guided position and token while retaining the unguided canvas update. Only after generation is complete do we replace the tokens at the recorded positions with their recorded HSCFG selections for output evaluation. This preserves the direct output effect while preventing the guided changes from influencing subsequent model inputs.

Table 1 reports the direct-output-only ablation. Its cumulative gain remains close to vanilla (42.298), while value-level repetition decreases from vanilla's 1.000 to 0.401. Full HSCFG achieves a cumulative gain of 41.016 and further reduces repetition to 0.132 when the guided changes also enter subsequent forward passes.

Table 1. HSCFG ablation separating direct output changes from their subsequent feedback.

SettingCumulative gainRHR_HRTR_TROR_ORtotalR_{\mathrm{total}}
Direct output changes only42.0720.5870.3680.3420.401

Your observation is therefore correct: the reduction to 0.132 reflects the combined effects of direct token changes and their feedback into subsequent forward passes.

Q1: Uncertainty in cumulative-gain differences

The standard errors of the paired per-rollout gain differences are 0.305 for quarter-rank SVD, 0.297 for HSCFG, and 0.338 for the repeating-versus-non-repeating comparison, all reported in the original cumulative-gain units. We calculated approximate 95% confidence intervals and expressed them as percentages of the corresponding reference means. The resulting intervals are [1.81%, 4.63%] for the SVD gain reduction, [1.65%, 4.41%] for the HSCFG gain reduction, and [2.93%, 6.21%] for the higher gain in repeating versus non-repeating rollouts. All three intervals lie entirely above zero, supporting both the reduction in cumulative gain under the interventions and the higher gain associated with repeating rollouts.

Q2: Sensitivity to anchor-detection hyperparameters

We further assessed the robustness of these measurements by varying the anchor-detection threshold, increasing the required excess above background from the default 20% to 40%, 70%, and 100%. The results are shown in Table 2.

Table 2. Anchor-detection threshold sensitivity.

Threshold above backgroundRepeating gainNon-repeating gainVanilla gainSVD gainHSCFG gainPearson rrSpearman ρ\rho
40%51.01449.48650.81549.50649.9000.83430.8787
70%57.90755.72057.83756.09356.4490.8190.824
100%67.21563.18567.11365.49266.0010.7890.837

The threshold affects the absolute scale of measured gain, but all tested settings preserve the comparisons central to our analysis: repeating rollouts have higher gain than non-repeating rollouts, and both SVD and HSCFG reduce gain relative to their corresponding vanilla baseline. Gain-to-anchor correlations also remain strongly positive, with Pearson rr ranging from 0.789 to 0.8343 and Spearman ρ\rho from 0.824 to 0.8787.

Thus, the relative gain comparisons and gain–anchor association supporting our mechanism remain robust across the tested thresholds, including substantially stricter anchor criteria. We will include these sensitivity results in the revision.

Q3: Randomized and round-robin schedulers

To examine the contribution of scheduling to the slotwise cascade, we evaluated both round-robin and randomized schedulers while holding the remaining settings fixed. Table 3 reports the results.

Table 3. Scheduler ablations with the remaining settings fixed.

SchedulerCumulative gainRHR_HRTR_TROR_ORtotalR_{\mathrm{total}}NMAE
Round-robin40.58330.42760.12280.10500.21801.372
Randomized40.3890.418670.116430.10380.21991.440

Repetition persists under both schedulers, but is substantially reduced. Cumulative gain also decreases from the vanilla value of 42.298 to 40.5833 and 40.389, respectively. Thus, disrupting the original reveal order weakens both cumulative reinforcement and repetition, consistent with our model's prediction.

These results support the reviewer's observation that scheduling contributes to the slotwise cascade. In particular, the original preference for revealing hundreds digits first facilitates early shared predictions that can reinforce subsequent repetition. The scheduler therefore affects the strength of the proposed feedback process, and the feedback–repetition relationship remains evident under both alternative policies.

However, using these schedulers to suppress repetition comes with a substantial forecasting cost: NMAE reaches 1.372 for round-robin and 1.440 for randomized scheduling, both exceeding the persistence baseline of 1. These scheduler changes therefore do not provide the same balance between repetition suppression and forecasting accuracy as our interventions. We will include these ablations in the revision.

Q4: SVD placement across layers

We also evaluated deep-only and uniform SVD placement against our original shallow-layer setting. For the uniform setting, we intervened on one head per layer to keep the total number of modified head–layer pairs approximately matched to the original intervention budget. Table 4 summarizes the comparison.

Table 4. Effect of SVD layer placement.

SVD placementCumulative gainValue-level repetition
Shallow layers (original)40.9360.701
Deep layers only41.2730.837
Uniform, one head per layer42.3060.992

Deep-only SVD provides a weaker reduction in repetition, with RH=0.9508R_H=0.9508, RT=0.817R_T=0.817, and RO=0.836R_O=0.836. Uniform placement produces little change in repetition, and its cumulative gain remains close to vanilla (42.306 versus 42.298).

These results support our choice to apply SVD in the shallow layers, where the stripe-plus-predecessor model provides the best fit and the intervention is most effective.

Q5: Anchors, attention sinks, and Zhao et al. (2026)

Finally, regarding attention sinks, our anchors differ from the conventional picture of a few dominant sink tokens. In our forecasting rollouts, these stripes appear on meaningful content tokens, rather than being tied to BOS or the first sequence position, and are observed across all layers. Moreover, attention is distributed across multiple positions: in many attention matrices, roughly one sixth of the key positions form elevated columns.

Zhao et al. (2026) provide related empirical background. Their setting concerns cache-induced repetition in diffusion multimodal models, which differs substantially from our uncached forecasting setting. Their information-flow account is primarily an empirical interpretation based on attention patterns and entropy trajectories. Their anchors describe nearby context tokens that aggregate information, without a formal characterization of their role in repetition dynamics or a theoretical derivation of how that role produces repetition. Their observations therefore provide contextual background instead of a competing explanation of the mechanism studied here.

Reviewer_aHR8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates repetition collapse in diffusion language models for time-series forecasting. It proposes a simplified framework for attention dynamics to explain this phenomenon and introduces two training-free interventions to mitigate repetition and test the proposed mechanism.

Critical strengths
  1. A coherent mathematical formulation of the observed dynamics. The paper introduces a simplified attention model that captures shared-anchor broadcasting and predecessor feedback. Under the stated assumptions, the derivations provide a coherent explanation of how cumulative gain amplifies anchor attention and, with additional alignment conditions, promotes representation homogenization.
  2. Effective mitigation of the observed repetition phenomenon. Both SVD truncation and history-sparsified guidance substantially reduce repetition in the evaluated repetition-prone rollouts. Results on both LLaDA and Dream support the effectiveness of these training-free interventions in mitigating the studied phenomenon.
Critical weaknesses
  1. Insufficient justification of the problem’s practical significance. Repetition is not necessarily undesirable in time-series forecasting. The paper acknowledges that legitimate repetition exists in real-world data, yet it does not sufficiently explain when the observed repetitive behavior becomes pathological or which aspects of forecasting quality it harms.
  2. Limited demonstrated benefits beyond reducing repetition. The proposed interventions reduce the measured repetition rate, but their benefits for forecasting quality remain unclear. NMAE is largely unchanged or only marginally improved on LLaDA, while it worsens on Dream.
  3. Limited empirical scope and incomplete reproducibility details. The evaluation covers two models and uses data from the Monash Time Series Forecasting Archive, but the manuscript does not clearly specify which constituent datasets were used or the overall evaluation size. These omissions hinder reproducibility and make it difficult to assess the breadth of the evaluation.
  4. Incomplete figure annotations hinder interpretation. Figure 1 is particularly problematic: the caption refers to panels (a), (b), and (c), but the corresponding labels are absent from the figure. The faint gray curves in the top-row panels and the orange markers in the upper-right panel are also not clearly explained in the legend or caption. These omissions make the presentation difficult to follow.
Questions for authors
  1. Consequences of repetition. Could the authors clarify what harm the observed repetition causes in time-series forecasting? Specifically, which aspects of forecasting quality are degraded, and what empirical evidence distinguishes this behavior from legitimate repetition in the data?
  2. Dataset selection and evaluation details. Which constituent datasets from the Monash Time Series Forecasting Archive were used? Since Monash contains multiple datasets, please provide the specific dataset names, the number of evaluated series, and the overall evaluation size.
  3. Broader evaluation and forecasting benefits. Could the authors provide results on additional datasets and a more comprehensive evaluation of forecasting performance beyond repetition rate? Such evidence would help establish whether suppressing repetition produces more useful forecasts and whether the findings hold consistently across datasets.
4 comments

We thank the reviewer for their time and their feedback.

Response to Q1.

The observed collapse harms two aspects of forecasting quality: the accuracy of temporal changes and the preservation of variability. Our measurements show substantial variation in the ground-truth forecast windows, indicating that the observed near-constant repetition does not reflect the underlying temporal dynamics.

To quantify these consequences beyond repetition rate, we report ΔMAE, which measures errors in predicted changes between consecutive time points:

ΔMAE = (1 / (H − 1)) · Σ_{t=2}^{H} | (ŷ_t − ŷ_{t−1}) − (y_t − y_{t−1}) |,

where H is the forecast horizon, ŷ_t is the prediction, and y_t is the corresponding ground truth. We also report temporal forecast variance over the corresponding prediction windows and compare it with ground-truth variance.

Table 1. Temporal change error and forecast variance for Dream and LLaDA. Lower ΔMAE\Delta\mathrm{MAE} is better. Forecast variance is compared with ground-truth variance over the corresponding prediction windows.

ModelVanilla ΔMAE↓\Delta\mathrm{MAE}\downarrowVanilla VarianceHSCFG ΔMAE↓\Delta\mathrm{MAE}\downarrowHSCFG VarianceSVD ΔMAE↓\Delta\mathrm{MAE}\downarrowSVD VarianceGround truth Variance
Dream183.807392.75443.7836774.979150.651752.59511306.728
LLaDA166.718340.04953.7119626.553137.932639.63011306.728

Across both models, the forecasts generated by vanilla decoding exhibit extensive repetition, and their mean temporal variance is only 3.0%–3.5% of the corresponding ground-truth variance, indicating severe suppression of temporal variability. The high ΔMAE\Delta\mathrm{MAE} further shows that the forecasts fail to capture the target's consecutive changes accurately.

Both interventions lower ΔMAE\Delta\mathrm{MAE} and move forecast variance toward the ground-truth value. HSCFG reduces ΔMAE\Delta\mathrm{MAE} by 76.2% on Dream and 67.8% on LLaDA, while SVD reduces it by 18.0% and 17.3%, respectively. These results provide evidence that the interventions improve temporal change accuracy alongside reducing repetition.


Response to Q2.

We evaluated all 7,578 time series across the 16 constituent datasets used in our Monash Time Series Forecasting Archive evaluation. The per-dataset counts were weather (3,010), covid_deaths (266), solar_weekly (137), tourism_monthly (366), australian_electricity_demand (5), pedestrian_counts (66), traffic_hourly (862), hospital (767), fred_md (107), tourism_yearly (518), tourism_quarterly (427), us_births (1), nn5_weekly (111), traffic_weekly (862), saugeenday (1), and cif_2016 (72).


Response to Q3.

We have extended the evaluation beyond Monash to all four ETT datasets (ETTh1, ETTh2, ETTm1, and ETTm2), comprising 1,200 forecasting examples, together with eight Darts series. We evaluate both LLaDA and Dream and report pooled results below. Beyond repetition, we measure ΔMAE\Delta\mathrm{MAE} and temporal forecast variance, as defined in our response to Q1, alongside NMAE.

Repetition mitigation generalizes to the additional evaluation data.

As shown in Table 2, both interventions reduce all three repetition metrics (RHR_H, RTR_T, and ROR_O) on both models.

Table 2. Repetition metrics on the additional evaluation data, pooled across the four ETT datasets and eight Darts series.

ModelVanilla RHR_HVanilla RTR_TVanilla ROR_OSVD RHR_HSVD RTR_TSVD ROR_OHSCFG RHR_HHSCFG RTR_THSCFG ROR_O
Dream0.9850.8710.9460.8980.7510.7000.4730.2090.257
LLaDA0.9200.8930.8870.8690.7730.7280.4750.2410.275

These reductions are accompanied by improvements in temporal forecasting quality.

Tables 3 and 4 show that both interventions lower ΔMAE\Delta\mathrm{MAE} on both models. HSCFG reduces ΔMAE\Delta\mathrm{MAE} by 16.7% on LLaDA and 40.6% on Dream; SVD reduces it by 6.8% and 9.5%, respectively.

Table 3. Additional forecasting evaluation on LLaDA, pooled across the four ETT datasets and eight Darts series. Lower ΔMAE\Delta\mathrm{MAE} and NMAE are better.

MethodΔMAE↓\Delta\mathrm{MAE}\downarrowForecast varianceNMAE↓\mathrm{NMAE}\downarrow
Vanilla18.3491945.5710.861
SVD17.0992193.4040.861
HSCFG15.2825718.2630.880

Table 4. Additional forecasting evaluation on Dream, pooled across the four ETT datasets and eight Darts series. Lower ΔMAE\Delta\mathrm{MAE} and NMAE are better.

MethodΔMAE↓\Delta\mathrm{MAE}\downarrowForecast varianceNMAE↓\mathrm{NMAE}\downarrow
Vanilla40.6671245.4160.833
SVD36.7931370.5790.840
HSCFG24.1485574.2280.846

The mean ground-truth temporal variance over the corresponding prediction windows is 5597.435 for both models' evaluations. The extensively repetitive vanilla forecasts exhibit suppressed temporal variance relative to this value. HSCFG brings the mean forecast variance close to the ground-truth value on both models, while SVD produces a smaller increase. Together with the lower ΔMAE\Delta\mathrm{MAE}, these results show benefits beyond reducing repetition: the forecasts better capture temporal changes and retain more of the target variability.

NMAE remains unchanged for SVD on LLaDA and increases slightly in the other settings, by 0.007–0.019. Thus, the additional evaluation supports improvements in temporal change accuracy and variability, with a small trade-off in NMAE. This pattern holds across both models in the pooled evaluation.

Figure annotations.

We have added the missing panel labels (a), (b), and (c) and updated the legend in Figure 1.

The faint gray curves show look-ahead forecast trajectories from earlier denoising steps, including provisional predictions at positions that were still masked at those steps. The orange squares mark fully revealed time-series values within the look-ahead forecasts.

Thank you for the response. My main concerns for the consequences of repetition and evaluation scope are well addressed, and I would like to increase the score to 3 accordingly.

Thank you for following up. We appreciate your feedback and are glad that our responses addressed your concerns.

Reviewer_EK6WOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the research question in repetition collapse in diffusion language models for time-series forecasting. The paper further studies in two axes: diffusion step and transformer layer of DLM repetition evolving, and model this progressive repetition collapse in digit wise diffusion forecasting and identify a recurring shared anchor and predecessor attention geometry. It proposes an attention based mechanism involving anchor concentration and cumulative gain, and evaluates two training-free interventions for more insights: attention SVD truncation and history-sparsified classifier free guidance (HSCFG), both can reduce cumulative gain together with repetition while preserving forecasting accuracy.

Critical strengths

S1: Well motivated studies in interesting and timely topics: study on repetition collapse of DLM, I know several successful works in AR, like “The curious case of neural text degeneration [ICLR 2020]” in reference, help back up some successful decoding in AR setup. I encourage the authors to keep working in this direction and share more insights on DLM to the community.

S2: Good initial empirical results with theory behind on DLM, though I did not fully understand every detail behind the theory here.

Critical weaknesses

W1: Lack of results of the impact on normal rollouts for DLM: this is directly related to whether the proposed method and findings can generalize

W2: Lack of novelty in some way: similar methods have been proposed in AR, such as “Learning to Break the Loop: Analyzing and Mitigating Repetitions for Neural Text Generation, NeurIPS 2022”, while the paper does not bring or clarify unique challenges or novelty into DLM.

Questions for authors

Q1:(regarding W1). How could the method affect normal rollout of DLM, how it generalizes?

Q2:(regarding W2): Could you elaborate more of this method novelty to similar methods in AR LLM, and any unique challenge for DLM to the community?

3 comments

Thank you for raising these questions about generalization and novelty.

Q1. Do the interventions transfer to open-ended text generation?

We tested both interventions on the same 100 ChatGPT 6.1 generated story prefixes of 300–350 words each, using 128 denoising steps per continuation.

Table 1. Repetition in open-ended story continuations.

MetricVanillaSVDHSCFG
4-gram repetition in the continuation79.6%71.8%4.5%

The vanilla model entered a loop after its first sentence, while SVD entered a loop after two sentences, and HSCFG produced a coherent continuation with new events that advanced the plot.

Q2. What is the novelty, and what changes in diffusion?

The central novelty of our work is a mathematical framework that explains how attention–representation feedback accumulates across Transformer layers to reinforce repetition and motivates two training-free interventions to mitigate it. To our knowledge, this is the first measurable, system-level framework that mathematically connects attention–representation feedback across Transformer layers to cumulative reinforcement, and tests its predictions through measurements and interventions.

When several positions attend to shared anchors, their updated representations can strengthen attention to those anchors in the next layer. Our recurrence links cumulative gain to this amplification, which can align representations across positions and favor repeated predictions. In our experiments, repeating rollouts have higher cumulative gain on average than non-repeating rollouts. Both interventions reduce cumulative gain and repetition, in line with our theory. Our framework offers a way to trace a sequence-level failure back to measurable feedback across Transformer layers.

The methodological novelty lies in the history-sparsified reference branch we construct for HSCFG and the theoretical justification our mathematical model provides for the SVD intervention.

HSCFG. We introduce a partially conditioned reference branch for repetition mitigation. Standard LLaDA CFG masks the entire prompt in the reference branch. HSCFG retains a subset of the observed history while giving both branches the same partially revealed forecast. This changes the guidance contrast to capture the additional support provided by the withheld observations. As denoising progresses, the extra guidance for a repeated candidate can diminish when both branches increasingly agree on its relative preference. In our experiments, HSCFG reduces cumulative gain together with repetition, consistent with our theory.

SVD. We use our mathematical model to identify a specific attention component to suppress. Our spectral analysis shows how low-rank truncation can attenuate shifted-predecessor feedback while largely preserving the dominant shared-anchor direction in the reduced model. This provides a theoretical basis for selectively weakening a source of cumulative reinforcement through shallow-layer attention truncation.

Thank you for your efforts, please consider highlighting the contribution and novelty more and adding new experiments for open-ended story continuations for the final revision. I am raising the score from 2 to 3.

Thank you for your feedback and updated assessment. We will incorporate these changes into the revised manuscript.

Comments on the paper

2 comments

Dear reviewers,

A reminder that the discussion period will end on November 18 at 11:59pm AoE. I encourage you to interact with the authors' rebuttal to clarify any remaining issue (and how well the rebuttal addressed your concerns) before the next phase.

Thanks, The AC

Dear Reviewers,

Thank you once again for your review comments! As the discussion period draws to a close, we wanted to follow up on our detailed responses posted on November 12.

If you have any remaining questions or concerns, we’d appreciate hearing from you soon so we have time to respond before the discussion closes on November 18. If our responses address the points you raised, we would be grateful if you would consider updating your overall assessment.

Thank you again for your time and feedback!

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.