acceptodds
Under review as a conference paper at ICLR 2027

COROS: Context-Robust Time-Series Forecasting with Score-Based Diffusion Models

Abstract

Diffusion models have emerged as an effective approach for probabilistic time-series forecasting, generating future trajectories from historical observations via iterative denoising. However, in many real-world applications the conditioning history may be contaminated by measurement noise or anomalies, which can degrade conditional generation. We propose COROS (COntext-RObust time-series forecasting with Score-based diffusion models), which extends score-based diffusion to the concatenated past–future sequence. By jointly denoising both the forecast horizon and the (possibly corrupted) history during sampling, COROS improves robust forecasting even under noisy conditions. Because noise is often introduced during observation, our method reduces sensitivity to corrupted historical data, thereby enhancing forecasting accuracy. Across five multivariate probabilistic forecasting benchmarks, COROS achieves competitive or best-in-class performance compared with 13 baselines. Furthermore, evaluation on standard LTSF benchmarks, where robustness to corrupted history is most transparently reflected in point-forecast degradation, demonstrates that COROS exhibits substantially smaller error growth compared to representative baselines as input noise increases.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, Diffusion models, Robustness, Noisy conditioning, Probabilistic forecasting

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 3 reviews, above 27% of the others.
Reviewer_AfS8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes COROS, a score-based diffusion framework for time-series forecasting under corrupted historical inputs. COROS applies diffusion to the concatenated historical and future sequence while retaining the observed history as a condition. It further introduces an imbalanced loss for the historical and future coordinates and optionally uses classifier-free guidance. Experiments are conducted on five probabilistic forecasting datasets and three LTSF datasets with synthetic input corruption.

Critical strengths
  1. The paper studies a practically relevant problem: historical observations may be corrupted by measurement noise or anomalies, while most conditional forecasting methods assume reliable context.
  2. The empirical evaluation covers datasets with substantially different dimensionalities, and some robustness improvements.
Critical weaknesses
  1. Limited methodological novelty Most components of COROS are already established. Full-sequence diffusion has been explored in methods such as TSDiff; conditional diffusion, VP-SDE, DiffWave-style networks, denoising score matching, and classifier-free guidance are standard techniques. Assigning different weights to historical and future coordinates is also a natural engineering choice. The main contribution is therefore the combination of conditional full-sequence diffusion, latent-history generation, and an imbalanced loss for corrupted-context forecasting. While reasonable, I do not find this combination sufficiently novel without stronger technical or empirical evidence.

  2. Presentation and reproducibility The manuscript contains multiple grammatical, notation, algorithm-description, and bibliography issues. Some variables are undefined or unused, several experimental hyperparameters are insufficiently explained, and the reference list contains duplicated or corrupted entries. Some citations also provide only indirect support for the associated claims. These issues are not the primary reason for rejection, but they indicate that the paper requires substantial revision and make the method harder to reproduce.

  3. The claimed robustness mechanism is not established. COROS still parameterizes the score as sθ(t,xthist,xtpred,c)s_\theta(t,x_t^{\mathrm{hist}},x_t^{\mathrm{pred}},c), so the corrupted condition cc remains directly available to the forecast score at every reverse step. The architecture does not require the forecast to use the supposedly denoised latent history. Moreover, the theoretical factorization relies on xtpred⊥ct∣xthistx_t^{\mathrm{pred}}\perp c_t\mid x_t^{\mathrm{hist}}, which does not generally follow after independently noising variables with a shared clean-history ancestor. The experiments do not demonstrate that the generated history is actually cleaner than the corrupted input or isolate purification from auxiliary-task regularization.

Questions for authors

1.What prevents the forecast-coordinate score from relying directly on the potentially corrupted condition cc and ignoring the evolving latent history xthistx_t^{\mathrm{hist}}? Is this dependence constrained architecturally or only expected to emerge from training?Can the authors quantitatively verify the claimed purification mechanism by measuring whether the final generated history is closer to the clean history than the corrupted input is? Is the generated history quantitatively closer to the clean history than the corrupted input? When γ=0\gamma=0, is the historical state still updated during sampling? Under what additional assumptions does the claimed conditional independence hold after forward noising? Does the robustness improvement remain when the forecast score is prevented from directly accessing cc?The architecture does not enforce the claimed denoising mechanism。The paper argues that prediction-only diffusion repeatedly uses a corrupted condition c=xhistc=x^{\mathrm{hist}} during sampling. However, COROS still parameterizes the score as

sθ(t,xthist,xtpred,c),s_\theta(t,x_t^{\mathrm{hist}},x_t^{\mathrm{pred}},c),

so the original condition remains directly available to the forecast score at every reverse step. COROS adds the indirect path

c→xthist→xtpred,c\rightarrow x_t^{\mathrm{hist}}\rightarrow x_t^{\mathrm{pred}},

but it does not remove the direct path

c→xtpred.c\rightarrow x_t^{\mathrm{pred}}.

Nothing prevents the network from ignoring the latent history and continuing to rely directly on the corrupted condition. Moreover, training uses the same clean history as both the condition and the historical generation target. The model is not explicitly trained to recover a clean history from a corrupted condition. The reported gains could therefore arise from auxiliary-task regularization or the enlarged diffusion state rather than actual history purification. The paper does not quantitatively demonstrate that the generated history is cleaner than the corrupted input.

  1. Does the robustness improvement remain without classifier-free guidance?

  2. In Algorithm 1, the condition c=xhistc=x^{\mathrm{hist}} is identical to the clean target of the historical diffusion coordinates. Therefore, q(xthist∣c)q(x_t^{\mathrm{hist}}\mid c) is a Gaussian centered at αtc\alpha_t c, and its optimal score appears to reconstruct cc exactly. Why should this objective remove a test-time corruption in cc, rather than reproduce it in the generated history? Is the condition independently corrupted or noised during training in a way not described in Algorithm 1?

5 comments

We have uploaded the revised manuscript; all changes are highlighted in blue.

We thank the reviewer for the detailed and technically careful review. New results are in Appendix A.7 (Table 11) and Appendix I (Figure 7, Table 17).

W1. Methodological novelty.

We agree that the individual components (VP-SDE, DiffWave-style backbone, DSM, CFG) are standard and that conditional full-sequence diffusion has precedents; we do not claim these components as new. The contribution is (i) the setting: diffusion forecasters trained on clean data and evaluated when the observed context is corrupted at test time; (ii) a model that keeps the observed context c\mathbf c as the condition while jointly evolving a latent history with the forecast, with a history-loss weight γ\gamma; and (iii) empirical evidence that this design choice matters.

The new control in Table 11 provides this evidence. With γ=1\gamma=1, the history and forecast losses are weighted equally, which is the full-sequence conditional objective of TSDiff-Cond (Kollovieh et al., 2023), here with our backbone. Under the same corruption protocol, training procedure, and sampler (Traffic MSE):

Methodσ=0\sigma=0σ=1.5\sigma=1.5
COROS (selected γ\gamma)0.250.78
Prediction-only0.280.82
COROS, γ=1\gamma=10.340.95
COROS, γ=1\gamma=1 + condition purification0.891.02

The selected weight is more accurate at both noise levels and reduces the corruption-induced error growth from 0.61 to 0.53, whereas the equal-weight objective is worse than prediction-only even on clean inputs. (The last row is discussed under Q1(c) below.)

W2. Presentation and reproducibility.

We agree and have revised the manuscript accordingly.

  • References. Merged the duplicated entries (Torres et al.) and fixed the corrupted one (Lütkepohl). Citations that gave only indirect support were replaced: DSM is attributed to Vincent (2011) and Song et al. (2021), observation self-guidance to Kollovieh et al. (2023), and the effect of forward noising on anomalies additionally to the purification literature (Yoon et al., 2021; Liu et al., 2025).
  • Notation. The observed history is denoted c\mathbf c throughout (it was previously also written xhist\mathbf x^{\mathrm{hist}}, yhist\mathbf y^{\mathrm{hist}}, yobs\mathbf y_{\mathrm{obs}}, or hh), with xthist\mathbf x_t^{\mathrm{hist}} reserved for the latent state. Previously undefined symbols (λ(t)\lambda(t), the VP schedule βmin⁡\beta_{\min}, βmax⁡\beta_{\max}, αt\alpha_t, σt\sigma_t, and the constants of Theorem 1) are defined in the main text, and overloaded symbols (TT, tt, σ\sigma, ww) are disambiguated.
  • Algorithm description. Algorithm 1 now includes λ(t)\lambda(t) and no longer samples an unused variable (η\eta).
  • Hyperparameters. The abbreviated columns of Table 16 are now defined.
  • Typos. We also corrected typographical errors.

W3 / Q1. The robustness mechanism.

(a) Direct access to c\mathbf c. We agree that the architecture does not enforce reliance on the latent history: the forecast score retains a direct path from c\mathbf c. This is by design, since COROS uses the observed context and the evolving latent history jointly, and the existence of a direct path does not by itself determine the overall sensitivity to corruption. What matters is how the trained network uses the two inputs. Appendix B (Table 12) measures this with paired Jacobian diagnostics: at the same diffusion time and latent state, switching the context from clean to corrupted decreases the forecast score's sensitivity to the observed context (median 0.76× on Traffic, 0.66× on Weather) while maintaining or increasing its sensitivity to the latent history (1.02×, 1.12×). On Electricity both increase, the latent-history sensitivity more (1.15× vs. 1.07×).

(b) Is the generated history cleaner than the corrupted input? On real benchmarks the clean history is itself unknown due to its possible inherent noise, so we answer where the ground truth is exact: two ODE systems with N(0,I)\mathcal N(0,I) spikes injected into 1/8 of the context (Appendix I). With the selected γ\gamma, the generated history largely does not follow the injected spikes while the forecast matches the ground truth (Figure 7). Forecast errors are in Table 17 (2D oscillator: MSE 0.2953 at γ=0.01\gamma=0.01 vs. 0.0006 at γ=0.1\gamma=0.1; 3D harmonic: 0.0210 vs. 0.0002; means over five seeds). Regarding auxiliary-task regularization and the enlarged diffusion state: the γ=1\gamma=1 model in Table 11 has the same enlarged state and the strongest auxiliary history-reconstruction loss, yet it is less accurate and less robust than the selected weight. The gains are therefore not explained by the enlarged state or by more auxiliary history training alone.

(c) Forecasting without direct access to c\mathbf c. Removing the direct path would change the conditioning design rather than isolate a component of it. The closest test we ran goes the other way: we replace c\mathbf c with a purified estimate c^\widehat{\mathbf c} (Appendix A.7, following MadSGM), so the forecast never sees the raw corrupted context. For the same model, this increases MSE from 0.34 to 0.89 on clean inputs and from 0.95 to 1.02 at σ=1.5\sigma=1.5 (Table 11). The observed context therefore carries details that the forecast should retain.

(d) γ=0\gamma=0. With γ=0\gamma=0 the history coordinates receive no training signal and COROS reduces to the prediction-only model, which is the prediction-only control reported in Tables 3, 9, 10, and 11.

(e) Conditional independence after forward noising. We agree that xtpred⊥ct∣xthist\mathbf x_t^{\mathrm{pred}}\perp \mathbf c_t \mid \mathbf x_t^{\mathrm{hist}} does not hold in general; a smooth Gaussian example satisfying our regularity assumptions (A1)–(A3) already violates it, so no such assumption restores it. We no longer rely on it. Lemma 2 now uses the exact decomposition pt(xthist,xtpred∣c)=pt(xthist∣c) pt(xtpred∣xthist,c)p_t(\mathbf x_t^{\mathrm{hist}},\mathbf x_t^{\mathrm{pred}}\mid\mathbf c)=p_t(\mathbf x_t^{\mathrm{hist}}\mid\mathbf c)\,p_t(\mathbf x_t^{\mathrm{pred}}\mid\mathbf x_t^{\mathrm{hist}},\mathbf c), which requires no independence assumption, and Theorem 1 is re-derived so that it accounts for both the direct dependence on c\mathbf c and the dependence through the latent history.

Q2. Does the robustness improvement remain without classifier-free guidance?

Yes. All corrupted-history results in Tables 3, 9, 10, and 11 are sampled without guidance. Guided results are labeled COROS-CFG (Table 2), and every unlabeled COROS entry is unguided; we now state this convention in Section 5.

Q3. If c=xhist\mathbf c=\mathbf x^{\mathrm{hist}} is also the target, why would the model remove rather than reproduce corruption?

We agree at the level of the ideal score on the training distribution: with c=xhist\mathbf c=\mathbf x^{\mathrm{hist}} the optimal history score reconstructs c\mathbf c, so the objective alone does not guarantee that corruption is removed. We therefore present the suppression of corrupted context as an empirical property of the trained model rather than a consequence of the objective. Three points qualify the conclusion that the model will reproduce the corruption.

  1. A corrupted test-time context lies outside the training distribution, so the network's behavior there is not determined by the population-optimal identity. At the forecast level, COROS reduces error growth by 58% relative to the matched prediction-only control on Electricity (Table 10) and has a lower CRPS$_{\mathrm{sum}}$ at every noise level on Electricity and Weather (Table 9).
  2. The condition is not corrupted during training: training and model selection use clean data, and corruption is applied only at test time (Appendix D; we have made this explicit).
  3. The toy experiment in Appendix I tests exactly this question. There the training trajectories are noise-augmented and the same augmented sequence is used as both condition and history target, so c=xhist\mathbf c=\mathbf x^{\mathrm{hist}} holds and the model never sees a clean target for a noisy condition. Consistent with the reviewer's analysis, with γ=1\gamma=1 the generated history reproduces the injected spikes almost exactly (Figure 7, lower panels). With γ=0.1\gamma=0.1 the same architecture suppresses most of them, and the forecast error falls (Table 17). Since a positive γ\gamma does not change the population-optimal score, we attribute this to the finite trained model.

We thank the reviewer again. We believe the corrected theory, the clarified scope, and the new evidence on equal weighting, condition purification, and the toy systems address the main soundness concerns, and we would be grateful if the reviewer would reconsider the assessment in light of these revisions.

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

This paper proposes COROS for probabilistic time-series forecasting with corrupted historical inputs. Instead of generating only the future, it jointly denoises the concatenated history and forecast window while retaining the observed history as a fixed condition. A weighted loss balances historical reconstruction and forecasting. The paper provides a robustness analysis and evaluates forecasting accuracy under several types and levels of input corruption.

Critical strengths

1.The method is simple and addresses a useful problem. Joint history–future generation improves forecasting under corrupted inputs in the reported experiments, including comparisons with a matched prediction-only diffusion model. 2.The evaluation includes different corruption strengths and types, history-loss weights, and sampling budgets. These experiments provide useful evidence beyond clean-data forecasting and help assess when the proposed method is effective.

Critical weaknesses

1.The explanation for robustness is not fully supported by the theory or experiments. The theoretical analysis does not clearly establish why jointly generating history and future should remove noise from the observed history. In particular, the conditional independence used in the analysis needs further justification, and comparing upper bounds on sensitivity does not by itself establish that one method is less sensitive than another. The experiments also lack a simple “denoise the history first, then forecast” baseline. Such a comparison would help determine whether joint generation provides an advantage beyond cleaning the historical inputs. 2. The historical loss weight in the provided code differs from the paper. The paper defines the loss as a weighted sum of squared residuals, with weight γ on the historical coordinates. However, in main_model.py, the historical residual is multiplied by a hard-coded value of 0.1 before squaring, making the effective weight 0.01. This differs from the values reported in Table 15, including 0.1 for several datasets and 0.5 for Electricity. Please clarify which implementation was used for the reported results and whether this difference affects the results and the γ ablation.

Questions for authors

1.Can you justify or revise the conditional-independence claim in Lemma 2 and the strict sensitivity comparison in Theorem 3? Does the reconstructed history become closer to the uncorrupted history during sampling? 2.How does COROS compare with a separate history-denoising stage followed by the same forecasting backbone, considering both accuracy and computational cost?

4 comments

We have uploaded the revised manuscript; all changes are highlighted in blue.

We thank the reviewer for the positive assessment and the concrete suggestions, including the careful check of our code.

W1 / Q1. Theoretical support for robustness.

Conditional independence (Lemma 2). We agree that the claim needed revision: after independent forward noising, xtpred⊥ct∣xthist\mathbf x_t^{\mathrm{pred}}\perp\mathbf c_t\mid\mathbf x_t^{\mathrm{hist}} does not hold in general, because the noised history does not screen off the clean history shared by the condition and the future. Lemma 2 now uses the exact decomposition pt(xthist,xtpred∣c)=pt(xthist∣c)pt(xtpred∣xthist,c),p_t(\mathbf x_t^{\mathrm{hist}},\mathbf x_t^{\mathrm{pred}}\mid\mathbf c)=p_t(\mathbf x_t^{\mathrm{hist}}\mid\mathbf c)p_t(\mathbf x_t^{\mathrm{pred}}\mid\mathbf x_t^{\mathrm{hist}},\mathbf c), which requires no independence assumption and keeps the dependence of the forecast on the observed context c\mathbf c.

Strict sensitivity comparison (Theorem 3). We agree that a smaller upper bound does not by itself establish a smaller sensitivity, and we no longer draw this inference. The result is re-derived as Theorem 1, proved in Appendix G.2. For two runs with conditions c,c′\mathbf c,\mathbf c', the same initial state, and the same reverse noise, the forecast difference is bounded by zp(G)∥c−c′∥z_p(G)\Vert\mathbf c-\mathbf c'\Vert, where (zh,zp)(z_h,z_p) solves a linear comparison system containing the direct path from c\mathbf c, the path through the latent history, and their coupling. This is a guarantee for every pair of contexts, and it identifies the properties of the trained score that control it: the bound is non-increasing in the history-score curvature mhm_h and also depends on the forecast curvature mpm_p and the sensitivities Lh→pL_{h\to p} and Lc→pL_{c\to p}. The comparison with prediction-only sampling is stated as a sufficient condition for COROS to have the smaller bound. Appendices B–C already measure these quantities along the sampling trajectories (Tables 12–13), and the actual forecast sensitivity is measured directly by the error growth under corruption (Tables 3, 9, 10, and 11).

Why joint generation helps. Since the history target equals the condition during training, the objective alone does not force the generated history to be cleaner. We therefore present this as a property of the trained model and test it directly, as in the next point.

Does the reconstructed history become closer to the uncorrupted history? On real benchmarks the uncorrupted history is itself unknown, so we answer on two ODE systems with exactly known trajectories and N(0,I)\mathcal N(0,I) spikes injected into 1/8 of the context (Appendix I). With the selected weight (γ=0.1\gamma=0.1), the generated history largely does not follow the injected spikes while the forecast matches the ground truth (Figure 7). With γ=1\gamma=1, the generated history reproduces the spikes. Forecast MSE is 0.0006 at γ=0.1\gamma=0.1 vs. 0.2953 at γ=0.01\gamma=0.01 (2D oscillator; 3D harmonic: 0.0002 vs. 0.0210; five seeds; Table 17).

W1 / Q2. Separate history-denoising stage followed by the same forecasting backbone.

We implemented this two-stage baseline following MadSGM (Lim et al., 2023b), using one model for both stages so that the forecasting backbone is identical. The model is trained with γ=1\gamma=1, so that its unconditional branch learns the full history. Stage 1 purifies the observed history with MadSGM's default settings: it noises the history to an intermediate time τ\tau and denoises it with the unconditional branch. Stage 2 uses the purified history c^\widehat{\mathbf c} as the condition in standard sampling with the same model (Appendix A.7, Table 11; Traffic MSE):

Methodσ=0\sigma=0σ=1.5\sigma=1.5
COROS (selected γ\gamma)0.250.78
Same γ=1\gamma=1 model, observed context0.340.95
Same γ=1\gamma=1 model, purified context (two-stage)0.891.02

Accuracy. Purification degrades the same model at both noise levels (0.34 → 0.89 on clean inputs, 0.95 → 1.02 at σ=1.5\sigma=1.5), and COROS is more accurate than either variant. The two-stage pipeline replaces the observed context, which also removes fine-grained structure that the forecast relies on. COROS instead keeps the observed context as the condition and uses it jointly with the evolving latent history.

Cost. The two-stage pipeline requires an additional purification pass before forecasting. Because this pass starts from the intermediate time τ\tau, the overhead is modest, but COROS requires no such pass: its sampling cost is that of standard one-shot diffusion sampling (Table 15).

W2. Historical loss weight in the code.

Thank you for checking the code carefully. The supplementary archive inadvertently contained an early version of the loss, in which the history residual was scaled by a hard-coded 0.1 before squaring. All reported results, including the γ\gamma ablation in Figure 2, were obtained with the loss as written in the paper, γ∥m⊙Δ∥22+∥(1−m)⊙Δ∥22,\gamma\Vert\mathbf m\odot\Delta\Vert_2^2+\Vert(1-\mathbf m)\odot\Delta\Vert_2^2, i.e., with γ\gamma multiplying the squared norm and the dataset-specific values in Table 16 (0.1, and 0.5 for Electricity). The difference therefore does not affect the reported results or the ablation.

We thank the reviewer again for the constructive review.

Dear Reviewer XnyX,

Thank you again for your constructive review. Before the discussion period ends, we would be grateful to know whether our response and the revised manuscript address your concerns. In particular, we re-derived the robustness bound with explicit condition-dependent terms (Theorem 1, Appendix G.2) and added the two-stage baseline that purifies the history before forecasting with the same backbone (Table 11). We would be happy to clarify any remaining points.

Dear Reviewer XnyX,

We apologize for reaching out again as the discussion period draws to a close, and sincerely appreciate your time and thoughtful review.

We have updated the revised manuscript with additional experiments and clarifications addressing your concerns. We would be grateful if you could review these updates and share any remaining feedback.

Thank you again for your time and consideration.

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

COROS targets multivariate probabilistic forecasting when the observed history is corrupted at test time. It retains the observed history as a fixed conditioning input while applying VP diffusion to the concatenated history and forecast window. A weighted denoising score-matching objective adjusts the historical reconstruction loss, and classifier-free guidance is an optional enhancement. Validation includes five probabilistic forecasting datasets, three corrupted-history benchmarks, matched prediction-only controls, weight and sampling ablations, and computational measurements. The results show lower forecast errors in several settings, but the additional reduction in corruption-induced error growth varies substantially across datasets.

Critical strengths
  1. The empirical improvements are meaningful in some settings. Relative to the matched prediction-only control, the Electricity MSE increase from noise scale 0 to 1.5 falls from 0.3370 to 0.1412, a reduction of approximately 58.1% (Tables 3 and 10). This is consequential because it demonstrates a potentially useful robustness benefit beyond simply comparing unrelated architectures.

  2. The experimental analysis includes relevant controls and practical measurements. The paper compares prediction-only, Replacement, and OSG under a shared backbone, includes both equal and unequal history weights, reports five-seed results for several evaluations, and measures training cost, inference latency, and memory. These analyses provide a substantive basis for an empirical contribution; the work is not supported only by headline benchmark results.

Critical weaknesses
  1. The central robustness guarantee is not established. Appendix G.2 infers Yt⊥Ct∣HtY_t\perp C_t\mid H_t after independently noising variables that initially satisfy Y⊥C∣HY\perp C\mid H. Let H,U,VH,U,V be independent standard Gaussians, Y=H+UY=H+U, and C=H+VC=H+V. Applying independent VP noise with signal coefficient aa, where 0<a<10<a<1, yields Cov⁡(Yt,Ct∣Ht)=a2(1−a2)>0\operatorname{Cov}(Y_t,C_t\mid H_t)=a^2(1-a^2)>0. Thus Lemma 2 fails even in a regular Gaussian setting. Moreover, Proposition 2 fixes the condition and future trajectory, whereas Theorem 3 applies its contraction across different conditions and trajectories without accounting for the resulting forcing terms. The proof also infers a strict ordering of actual worst-case sensitivities from an ordering of upper bounds, which is not valid. These problems directly undermine a stated main contribution.

  2. The remaining methodological contribution and mechanism are insufficiently distinguished. Conditional full-sequence diffusion with observed-region reconstruction already appears in SSSD-D0 and TSDiff-Cond, the latter described in Appendix B.4 of the cited TSDiff paper. Figure 2 does provide a useful γ=1\gamma=1 control, but the corrupted-history experiments do not establish the additional robustness benefit of the selected weight over this equal-weight formulation. Furthermore, with the actual training condition c=hc=h and independent forward noise, the ideal base conditional model satisfies pt(ht,yt∣c)=qt(ht∣c)pt(yt∣c)p_t(h_t,y_t\mid c)=q_t(h_t\mid c)p_t(y_t\mid c) on the training support. Its future score therefore need not use latent history. Finite-model auxiliary learning may still help, but the proposed historical purification mechanism has not been separated from that explanation. This limitation applies to the base conditional branch; CFG introduces an additional unconditional branch and requires separate analysis.

Questions for authors
  1. Clarification / correction of the current claim: How do Lemma 2 and Theorem 3 accommodate the Gaussian construction above? Please provide a corrected statement and argument that address the condition-dependent forcing and distinguish a smaller upper bound from a smaller actual worst-case sensitivity.

  2. Additional evidence needed for the current contribution claim: What improvement remains over equal-weight conditional full-sequence diffusion under the same corrupted-history protocol, training budget, and sampling budget? Please relate this control explicitly to SSSD-D0 and TSDiff-Cond, and isolate the effects of the history weight and CFG.

  3. Clarification and additional evidence needed for the current mechanism claim: Given c=hc=h during training, what evidence shows that the evolving history becomes a better estimate of the clean history and that its updates improve the forecast? Reconstruction errors against the available clean reference, or a matched intervention that freezes or removes history updates while retaining auxiliary training, would materially affect this assessment.

8 comments

We have uploaded the revised manuscript; all changes are highlighted in blue.

We thank the reviewer for the careful review and for the precise counterexample, which identified the gap in Lemma 2. The revised analysis is in Appendix G.2 (the result previously numbered Theorem 3 is now Theorem 1, proved there); new results are in Appendix A.7 (Table 11) and Appendix I (Figure 7, Table 17).

W1 / Q1. Lemma 2, Theorem 3, and upper bounds versus actual sensitivities.

Lemma 2. The reviewer is correct. Independent forward noising does not preserve the conditional independence: in the Gaussian construction, Cov(Yt,Ct∣Ht)=a2(1−a2)>0\mathrm{Cov}(Y_t,C_t\mid H_t)=a^2(1-a^2)>0, because the noised history HtH_t does not screen off the clean history shared by YY and CC. We have removed this claim. Lemma 2 now states the exact decomposition pt(xthist,xtpred∣c)=pt(xthist∣c)pt(xtpred∣xthist,c),p_t(\mathbf x_t^{\mathrm{hist}},\mathbf x_t^{\mathrm{pred}}\mid\mathbf c)=p_t(\mathbf x_t^{\mathrm{hist}}\mid\mathbf c)p_t(\mathbf x_t^{\mathrm{pred}}\mid\mathbf x_t^{\mathrm{hist}},\mathbf c), which keeps the dependence of the forecast on c\mathbf c and requires no independence assumption, and the score identity in Lemma 3 retains c\mathbf c accordingly. The Gaussian construction is consistent with both.

Condition-dependent forcing. We agree that Proposition 2 compared runs with the same condition and future trajectory, while Theorem 3 applied it across different conditions. The revised analysis couples two runs with conditions c,c′\mathbf c,\mathbf c', the same initial Gaussian state, and the same reverse noise, and keeps the forcing explicitly. In the normalized reverse time u=H(t)u=H(t), Proposition 2 now reads dh(u)≤Kh(u,0)dh(0)+∫0uKh(u,v)[Lp→h(v)dp(v)+Lc→h(v)δ]dv,δ=∥c−c′∥,d_h(u)\le \mathcal K_h(u,0)d_h(0)+\int_0^u \mathcal K_h(u,v)\big[L_{p\to h}(v)d_p(v)+L_{c\to h}(v)\delta\big]dv,\qquad \delta=\Vert\mathbf c-\mathbf c'\Vert, and, combined with the corresponding forecast inequality, Theorem 1 bounds the forecast difference by ∥xεpred−xεpred′∥≤zp(G)δ\Vert\mathbf x_\varepsilon^{\mathrm{pred}}-\mathbf x_\varepsilon^{\mathrm{pred}\prime}\Vert\le z_p(G)\delta, where (zh,zp)(z_h,z_p) solves z˙h=−κhzh+Lp→hzp+Lc→h,z˙p=ℓpzp+Lh→pzh+Lc→p,zh(0)=zp(0)=0,\dot z_h=-\kappa_h z_h+L_{p\to h}z_p+L_{c\to h},\qquad \dot z_p=\ell_p z_p+L_{h\to p}z_h+L_{c\to p},\qquad z_h(0)=z_p(0)=0, with κh=mh−12\kappa_h=m_h-\tfrac12 and ℓp=12−mp\ell_p=\tfrac12-m_p. The direct path from c\mathbf c, the path through the latent history, and their coupling all enter, and the rates may change sign over time. Assumptions (A1)–(A3) are reformulated as the corresponding cross-coordinate Lipschitz bounds and one-sided curvature bounds of the history and forecast scores; the Gaussian construction satisfies them, and the bound holds there.

Upper bound versus actual sensitivity. We agree that an ordering of upper bounds does not by itself order the actual sensitivities, and we no longer draw this inference. We use the theorem for what a bound does establish. (i) It is a guarantee for every pair of contexts under (A1)–(A3): however the context is corrupted, the forecast changes by at most zp(G)δz_p(G)\delta. (ii) It identifies the properties of the trained score that control this guarantee: the history curvature mhm_h (the bound is non-increasing in mhm_h), the forecast curvature mpm_p, and the sensitivities Lc→pL_{c\to p} and Lh→pL_{h\to p}. The comparison with prediction-only sampling, which has its own coefficients LPOL_{\mathrm{PO}} and ℓPO\ell_{\mathrm{PO}}, gives a sufficient condition (ℓp≤ℓPO\ell_p\le\ell_{\mathrm{PO}} and Lh→pzh+Lc→p≤LPOL_{h\to p}z_h+L_{c\to p}\le L_{\mathrm{PO}} almost everywhere, strict on a set of positive measure) for COROS to have the smaller guarantee. Whether the trained model has these properties is checked by our diagnostics, described next, and the actual forecast sensitivity is measured directly by the error growth under corruption (Tables 3, 9, 10, and 11).

Relation to the existing diagnostics. The revised assumptions are written in terms of the same quantities that our diagnostics already measure along the sampling trajectories, so these results remain in place and now map directly onto the theory:

  • Appendix B (Table 12) reports Lh→pL_{h\to p} and Lc→pL_{c\to p}, the sensitivities of the trained forecast score to the latent history and to the observed condition. Under corruption, the direct sensitivity Lc→pL_{c\to p}, which enters the bound linearly, decreases (median 0.76× on Traffic, 0.66× on Weather), while the sensitivity to the latent history is maintained or increases (1.02×, 1.12×).
  • Appendix C (Table 13) reports mheffm_h^{\mathrm{eff}}, computed from the Jacobian of the network's full history score sh(t,xthist,xtpred,c)s_h(t,\mathbf x_t^{\mathrm{hist}},\mathbf x_t^{\mathrm{pred}},\mathbf c). This is exactly the quantity in the revised (A2). On Electricity and Weather its median exceeds 12\tfrac12 at every retained diffusion time, i.e., κh>0\kappa_h>0 and the history block contracts locally.
  • Table 13 also reports that the forecast curvature mpm_p changes sign during the reverse process; this observation is why the revised (A3) allows a time-dependent sign.
  • Figure 6 perturbs only the observed condition at a fixed noised input, an empirical check of bounded condition sensitivity.

We read these as local estimates along the trajectories rather than global constants.

W2 / Q2. Equal-weight conditional full-sequence diffusion (SSSD-D0, TSDiff-Cond) and the effect of CFG.

SSSD-D0 and TSDiff-Cond train a conditional full-sequence model with equal weight on the observed and forecast regions; in objective, both correspond to COROS with γ=1\gamma=1. We now discuss TSDiff-Cond in Appendix H and compare against this equal-weight objective with our backbone in Table 11, under the same corruption protocol, training procedure, and sampler (Traffic MSE):

Methodσ=0\sigma=0σ=1.5\sigma=1.5
COROS (selected γ\gamma)0.250.78
Prediction-only0.280.82
COROS, γ=1\gamma=1 (equal weight)0.340.95

The selected weight is more accurate at both noise levels and reduces the corruption-induced error growth from 0.61 to 0.53. The equal-weight model is worse than prediction-only even on clean inputs, consistent with full history reconstruction competing with forecasting.

All models share the same condition-dropout training, and all corrupted-history results (Tables 3, 9, 10, and 11) are sampled without guidance. The robustness comparisons therefore concern the base conditional branch and do not involve the CFG combination, so Table 11 isolates the effect of the history weight.

W2 / Q3. Evidence for the mechanism given c=xhist\mathbf c=\mathbf x^{\mathrm{hist}} during training.

We agree with the population-level argument: with c=xhist\mathbf c=\mathbf x^{\mathrm{hist}} and independent forward noise, the ideal base conditional model factorizes, its future score need not use the latent history, and the objective alone does not guarantee purification. As the reviewer anticipates, we present the mechanism as a property of the finite trained model and support it as follows.

  1. The trained forecast score uses the latent history. As above, Table 12 shows that the trained network responds to the latent history and relies on it relatively more under corruption.
  2. Comparison with a clean reference. Real benchmarks lack a clean history, so we use two ODE systems with exactly known trajectories and N(0,I)\mathcal N(0,I) spikes injected into 1/8 of the context (Appendix I). Training keeps the structure c=xhist\mathbf c=\mathbf x^{\mathrm{hist}}: the noise-augmented sequence is used as both condition and history target. With γ=1\gamma=1 the generated history reproduces the injected spikes (Figure 7), as the population argument predicts. With γ=0.1\gamma=0.1 it suppresses most of them while the forecast matches the ground truth, and the forecast MSE is 0.0006, compared with 0.2953 at γ=0.01\gamma=0.01, where the history is only weakly trained (2D oscillator; 3D harmonic 0.0002 vs. 0.0210; five seeds; Table 17). Better generation of the history thus coincides with a much more accurate forecast.
  3. Auxiliary training alone. The γ=1\gamma=1 model in Table 11 has the same latent state and the strongest auxiliary history-reconstruction loss, yet it is less accurate and less robust than the selected weight. Auxiliary history training by itself therefore does not account for the gains.

We thank the reviewer again for the precise and constructive critique. We believe the corrected Lemma 2, the re-derived bound with explicit forcing, the equal-weight control, and the clean-reference experiment address the main concerns, and we would be grateful if the reviewer would reconsider the assessment in light of these revisions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.