acceptodds
Under review as a conference paper at ICLR 2027

CoBiST: A Coupled Bidirectional Survival Transport Operator for Probabilistic Time-Series Generative Forecaster

Abstract

Probabilistic time-series forecasting targets the joint distribution of a segmentation of the future path. Modern generators approximate this distribution through sampling full trajectories. Most existing generative forecasters delegate the interaction between forecast positions to a generic backbone and typically lack a unified, normalized accounting of information routing across forecast positions. To capture such internal interaction during generation, we propose Coupled Bidirectional Survival Transport (CoBiST), a generator-agnostic cross-horizon transport operator with explicit probabilistic semantics for directional propagation, locality, and memory mechanisms. CoBiST can be plugged into multiple types of generator without changing the outer objective or sampler, where a shared gate produces directional preferences and one-step survival probabilities to jointly determine the budget assigned to causal and anti-causal recurrent scans and to a direct-local branch. The resulting transport operator is theoretically verified to admit an exact augmented row-stochastic kernel, a linear-time semiseparable realization, frozen-gate non-expansiveness, and a finite-horizon sensitivity bound for state-dependent gates. We instantiate the same operator in GAN, flow-matching, and diffusion forecasters, achieving average Projected CRPS reductions relative to capacity-matched dense controls of 6.72%, 21.48%, and 20.78% across eight datasets, respectively. CoBiST-Flow and CoBiST-Diffusion also achieve the best two average Projected CRPS, Marginal CRPS and Energy Score ranks among all evaluated forecasters.

Primary area:
generative models
Keywords:
Probabilistic time-series forecasting, Generative time-series models, Conditional generative modeling, Path-wise dependency modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Wu3tOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Coupled Bidirectional Survival Transport (CoBiST), a generator-agnostic cross-horizon transport operator designed for probabilistic time-series forecasting. Given an observed sequence history xx and a forecast horizon HH, the authors define a conditional generative forecasters to estimate the joint future distribution p⋆(⋅∣x)p^\star(\cdot \mid x) by mapping noise ϵ∼pϵ\epsilon \sim p_\epsilon to generated paths y^=Gθ(ϵ,x)\hat{y} = G_\theta(\epsilon, x). the method CoBiST replaces standard sequence backbones with a unified transport core operating on sequence states u=(u1,…,uH)u = (u_1, \dots, u_H).

Critical strengths

The methodology relies on rigourous Transport theory, the method CoBiST couples bidirectional scans, memory decay, and direct local routing into a unified, row-stochastic route simplex. The paper provides complete theoretical proofs: Theorem 1 uniquely characterizes the geometric-mean coupling budget qk=ρkpρkfq_k = \sqrt{\rho_k^p \rho_k^f} from axiomatic hazard principles; Theorem 2 & Proposition 2 derive an exact row-stochastic augmented kernel KkjK_{kj} with explicit source attributions; Theorem 3 proves rank-one semiseparability achieving exact O(Hw)O(Hw) time/memory complexity; and Theorem 4 guarantees frozen-gate non-expansiveness.

On the experimental side, I noted the The transport core is modular and easily inserted into GAN, Flow-matching, and Diffusion interfaces without altering outer training losses or samplers (Table 1). Across eight diverse benchmarks.

Critical weaknesses

I will appreciate to see a more formal explanantion distinction between Frozen-Gate Non-Expansiveness and State-Dependent Multi-Step Dynamics, I noted in Theorem 4 they established that the frozen-gate transport core is non-expansive (∥Tk−Tk′∥≤δ\|T_k - T_k'\| \le \delta), Proposition 4 (Appendix B.1) reveals that state-dependent gates introduce finite correction terms scaling with horizon length (2BHδρ+B(2δq+q⋆δα)2 B H \delta_\rho + B (2 \delta_q + q_\star \delta_\alpha)). While the authors explicitly document this qualification, the manuscript lacks an empirical quantification of how gate Lipschitzness affects perturbation growth across multi-step reverse ODE/SDE sampling trajectories in Flow and Diffusion models.

Questions for authors

Reflecting on what I have mentioned in the Critical Weaknesses, I want to emphasis on Proposition 4, the state-dependent bound contains a term 2BHδρ2 B H \delta_\rho that grows linearly with forecast horizon HH. In practice, do you use explicit gate regularization (e.g., spectral normalization, logit bounding, or weight decay on Γθ\Gamma_\theta) to control δρ\delta_\rho for longer horizons such as Wind (H=60H=60)?

Also, in the ablation studies (Table 5 and Section 5), the Neural-qq ablation performs comparably to the axiomatic geometric mean qk=ρkpρkfq_k = \sqrt{\rho_k^p \rho_k^f}. Did you observe any specific dataset regimes or horizon lengths where Neural-qq learned non-geometric coupling budgets that significantly improved or degraded performance?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Wu3t Part 1/2

We thank the reviewer for pointing out the gap between frozen-gate stability and state-dependent gates in iterative sampling.

W1 / Q1: State-dependent transport and multi-step dynamics.

Dynamic-gate theory under Lipschitz assumptions

We previously derive the frozen-gate theory because we didn't add assumption on the gate. We realized that the frozen-gate theory is less applicable, so we have replaced all related theory with a stronger conditional result under Lipschitz assumptions. Under Assumptions 1–2, for a fixed context and common boundaries, let embedded states have norm at most BB, retention logits have Lipschitz constant LbL_b, and the difference of direction logits have Lipschitz constant LℓL_\ell. For δ=max⁡j∥zj−zj′∥2\delta=\max_j\|z_j-z'_j\|_2, with gates recomputed for each input, Theorem 4 states

\le (1+2BL_b)\delta,$$ $$\max_k\{\|T _ k-T' _ k\| _ 2,\|D _ k-D' _ k\| _ 2\} \le \left(1+2BL _ b+\tfrac{B}{2}L _ \ell \right) \delta.$$ For realized gates, **Proposition 3** retains the kernel-level statement $$\|K _ {\mathrm{real}}\| _ \infty\le1, \qquad \rho(K _ {\mathrm{real}})\le1.$$ This supports the frozen-gate qualification with common boundaries. **Proposition 4** makes the additional state-dependent terms explicit: $$D_ZT[\dot Z] =K(\mathcal G(Z,c))\dot Z +(D _ ZK[\dot Z])Z +(D _ ZK_\partial[\dot Z])\xi.$$ The first term applies the realized kernel to the perturbation while the other two describe gate feedback. Theorem 4 bounds their combined effect. Appendix B.1 also gives a conditional complete-generator bound. If the embedding and positionwise readout have constants $L_E$ and $L_R$, and $L_{\mathrm{tr}}=1+2BL_b+(B/2)L_\ell$, then $$\|v(U)-v(U')\| _ {\infty,2} \le L _ RL _ E\sqrt{1+(2+4B^2)L _ {\mathrm{tr}}^2} \|U-U'\| _ {\infty,2}.$$ For a discrete sampler $U_{n+1}=a_nU_n+b_nv_n(U_n)+s_n\eta_n$ with state-independent scalar coefficients, matched additive noise, and full-field constants $L_{v,n}$ on the relevant domains, it also derives $$\|U _ N-U _ N'\|_{\infty,2} \le\prod _ {n=0}^{N-1}(|a _ n|+|b _ n|L _ {v,n}) \|U _ 0-U _ 0'\| _ {\infty,2}.$$ **Empirical perturbation growth and gate feedback** We add a new experiment to test whether CoBiST attenuates perturbations when its gates adapt to the perturbed input, within a single transport operation and through complete sampling. We first perturb the embedded horizon representation by $10^{-5}$ to $10^{-1}$ of its norm and recompute all gates. All dataset-wise median transport response ratios remain below $0.9$ and vary little with perturbation size (Figure 3 (a)--(c)). Next, we perturb the initial Flow or Diffusion sampling state by $1\%$ and track the distance between paired runs, reported in Figure 3 (d),(e). Generally, these results **support CoBiST's ability to combine adaptive cross-horizon routing with attenuation of the tested perturbations** across all three generator families, and most evaluated Flow and Diffusion samplers also attenuate initial-state perturbations over complete sampling (except amplification for Weather). We also examine the contribution of gate feedback through reference-gate replay controls in Appendix E.4. **Regularization at longer horizons** Our current CoBiST models do not use explicit regularization designed to control the gates or their sensitivity along the forecast horizon. In preliminary tests, we observed that some regularization strategies could be helpful, but their performance are sensitive to datasets. Therefore, currently we havn't identified a good regularizer to characterize the trade-off between forecasting accuracy and stability. We have written the horizon-wise regularization as a limitation of the current work into the manuscript and leave this for future work.
AuthorsOpenReview ↗

Response to Reviewer Wu3t Part 2/2

Q2: Neural-qq across datasets and horizons

It's worth noting that we follow the suggestions by other reviewers and revise the Nerual-qq experiments to a more convinced design. In the revision, the new revised Neural-qq now estimates qq directly by a network taken local and context embeddings as input. It is independently tuned fairly under the same protocol (Appendix D.3).

Table 3 gives the following PCRPS improvement rates (%), mean (std), across all eight datasets of CoBiST over the revised Neural-qq:

Model comparisonElec.Exch.TrafficM4WikiTaxiWindWeatherMean gain[95% CI]
CoBiST-GAN vs Neural-qq1.03 (7.36)12.30 (18.98)-1.99 (2.31)8.33 (6.71)0.92 (8.64)-0.46 (3.50)2.44 (28.74)-9.75 (4.57)1.60[-3.79, 7.24]
CoBiST-Flow vs Neural-qq-0.05 (0.47)4.40 (3.99)-0.71 (0.73)-0.20 (0.59)0.43 (0.34)0.18 (0.63)3.94 (1.93)-0.11 (1.79)0.98[-0.19, 2.46]
CoBiST-Diffusion vs Neural-qq0.03 (0.26)-0.63 (1.68)0.05 (0.20)-0.45 (1.24)-0.21 (0.47)0.05 (0.12)-0.89 (1.42)-1.49 (4.41)-0.44[-1.19, 0.09]

Generally, Neural-qq is close to CoBiST without significant difference, supporting the geometric mean as a simple, semantically interpretable choice with competitive accuracy. Detailed results of other scores and raw metrics are reported in Appendix E.1. Most differences between Neural-qq and CoBiST are not significant. We therefore can hardly statistically established regimes for possible significant improved or degraded performance.

I thank the authors for their extensive response, I appreciate the derivation about the new condition erivations under Lipschitz asumptions and the new empirical perturbation, this allow to effectively quantify how gat feedback behave across multu-step trajectories. This resolves my concerns regarding state-dependant dynamics. The revised q-ablation (using q ) was a a rigorous update whcih I needed to see to validate your architectural design choice.

Based on the rebruttal updates, I am increasing my score.

AuthorsOpenReview ↗

Thank you for recognizing our effort!

Thank you for recognizing our effort and your valuable comments. Your suggestions help us develop the newer and more appliable dynamic gate theory. Related sensitivity experiments provide further empirical support on it. Your comment regarding the geometric mean and regularization is also insightful. We would appreciate it if you could increase your rating as you mentioned, which is extremely important for us.

Reviewer_ve1XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoBiST proposes a generator-agnostic cross-horizon transport operator for probabilistic time-series forecasting. A shared gate determines directional preferences and two survival probabilities; their geometric mean sets the recurrent-versus-local budget. Causal and anti-causal survival scans, together with a direct-local path, yield an augmented row-stochastic source kernel and a semiseparable linear-time transport realization. The operator is inserted into GAN, flow-matching, and diffusion generators without changing their respective outer objectives or samplers. Across eight datasets, the paper reports average PCRPS reductions of 6.72%, 21.48%, and 20.78% relative to the corresponding approximately capacity-matched Dense controls, respectively. In a separate comparison with six external forecasters, the Flow and Diffusion variants achieve the best two average ranks on PCRPS, MCRPS, and ES. The paper also presents conditional frozen-gate stability results and intervention-based validation of transport attributions.

Critical strengths
  • The single gate couples directional preferences and survival probabilities with a normalized recurrent budget. The augmented kernel separates branch probabilities from actual past/local/future and boundary source masses, making the transport mechanism more auditable than an unqualified attention visualization.

  • The same operator can be embedded in GAN, flow-matching, and diffusion generators. Relative to their approximately capacity-matched Dense controls, the average PCRPS reductions are 6.72%, 21.48%, and 20.78% across eight datasets; the Flow/Diffusion variants also have strong average PCRPS, MCRPS and ES ranks among the nine forecasters evaluated.

  • The semiseparable scan gives linear-time transport computation; the authors explicitly distinguish frozen-gate non-expansiveness from state-dependent gate feedback. Forecasting results include four formal seeds, while a separate intervention analysis compares exact source masses with nonlinear-readout responses rather than merely plotting gate weights.

Critical weaknesses
  • The Dense control replaces the entire CoBiST block while allowing up to a 10% parameter-count difference; it does not disentangle bidirectionality, adaptive retention, shared routing, and the conserved coupling. Neural-q is a small 0.15-scale residual around the geometric-mean logit, not an unconstrained budget alternative. Its largely comparable accuracy supports simplicity but cannot establish that geometric-mean coupling is empirically necessary or optimal.

  • The paper proves O(Hw) transport work and gives the full block's theoretical complexity, but reports no wall-clock inference/training time, peak memory, throughput, or empirical horizon-scaling curves. The observed horizons are only 24–60, and multi-step Flow/Diffusion inference repeatedly invokes the block. It is therefore difficult to quantify an end-to-end practical advantage over Dense or modern structured alternatives .

  • The deletion study freezes generator states and gates, modifies transport coordinates while retaining the direct local coordinate, and uses seed 2027 rather than all four formal seeds. Its dataset-bootstrap intervals do not represent initialization variability. Source masses align with readout responses, but this does not establish faithful end-to-end attribution of final samples or predictive error.

  • DeepAR has the lowest mean PCRPS on four datasets; Flow is fourth by average VS rank, and GAN–Wind's VS is 44.92% worse than its Dense control despite a PCRPS gain. Four-seed standard deviations are valuable, but the study does not report paired confidence intervals for close model differences or test-origin resampling. The boundary-correlation text and Table 13 also disagree for GAN (0.83 vs 0.87) and Diffusion (0.90 vs 0.89).

Questions for authors
  • Could you isolate the benefits of bidirectionality, input-adaptive survival, the shared gate, and the conserved recurrent/local budget using controlled one-directional, fixed-retention, and decoupled-budget variants at matched parameter and training budgets? Since Neural-q is only a small residual around the proposed geometric mean, could a less constrained budget or a comparably tuned bidirectional structured mixer clarify which design choice matters most?

  • Could you report inference/training wall-clock time, peak memory, and horizon-scaling curves for matched CoBiST/Dense variants, including the full Flow/Diffusion sampling loop and the actual parameter counts/capacity gaps? How do the results change for substantially longer forecast horizons?

  • How stable are the source/deletion and memory/direction correlations across the four formal seeds? Could a follow-up intervention allow the gates and generator trajectory to respond, and report changes in final predictive scores, to delineate what the present frozen-state attribution does and does not explain?

  • Could you provide paired per-dataset differences with uncertainty when CoBiST-Flow/Diffusion and external methods are close, clarify the VS settings in the PDF, and reconcile the boundary Spearman numbers in the Appendix E.3 text (0.83/0.94/0.90) and Table 13 (0.87/0.94/0.89)?

7 comments
AuthorsOpenReview ↗

Response to Reviewer ve1X Part 1/2

We thank the reviewer for the detailed suggestions on controls, computation, interventions, and uncertainty.

W1 / Q1: Isolating architectural components

To verify the contribution of coupling, we added an independently tuned, approximately capacity-matched Bi-LSTM control to each generator family, where Bi-LSTM worked as a sequence-to-sequence mixer without ground-truth future observations inside the forecast horizon, and CoBiST's corresponding average improvements over Bi-LSTM are 14.58%, 4.16%, and 5.36% for GAN, Flow and Diffusion, respectively with significance.

Table 3 gives the following PCRPS improvement rates (%), mean (std), across all eight datasets of CoBiST over Bi-LSTM:

Model comparisonElec.Exch.TrafficM4WikiTaxiWindWeatherMean gain[95% CI]
CoBiST-GAN vs Bi-LSTM7.05 (13.35)31.23 (9.74)13.13 (13.58)22.58 (5.15)7.82 (4.36)14.03 (8.93)-0.76 (15.35)21.59 (35.46)14.58[6.63, 22.78]
CoBiST-Flow vs Bi-LSTM0.74 (0.75)2.95 (2.18)-1.07 (0.79)3.23 (1.62)1.55 (1.38)1.89 (1.07)25.49 (0.47)-1.53 (2.37)4.16[0.30, 11.09]
CoBiST-Diffusion vs Bi-LSTM1.60 (0.55)0.73 (2.53)0.85 (0.38)4.65 (1.55)1.56 (0.68)-0.05 (0.47)35.11 (2.04)-1.54 (2.65)5.36[0.20, 15.06]

These comparisons support CoBiST's advantage over bidirectional recurrent mixer.

We agree that the original Neural-qq control that learns the residual was less convincing. In the revision, the new revised Neural-qq now estimates qq directly by a network taken local and context embeddings as input. It is independently tuned fairly under the same protocol (Appendix D.3).

Table 3 gives the following PCRPS improvement rates (%), mean (std), across all eight datasets of CoBiST over the revised Neural-qq:

Model comparisonElec.Exch.TrafficM4WikiTaxiWindWeatherMean gain[95% CI]
CoBiST-GAN vs Neural-qq1.03 (7.36)12.30 (18.98)-1.99 (2.31)8.33 (6.71)0.92 (8.64)-0.46 (3.50)2.44 (28.74)-9.75 (4.57)1.60[-3.79, 7.24]
CoBiST-Flow vs Neural-qq-0.05 (0.47)4.40 (3.99)-0.71 (0.73)-0.20 (0.59)0.43 (0.34)0.18 (0.63)3.94 (1.93)-0.11 (1.79)0.98[-0.19, 2.46]
CoBiST-Diffusion vs Neural-qq0.03 (0.26)-0.63 (1.68)0.05 (0.20)-0.45 (1.24)-0.21 (0.47)0.05 (0.12)-0.89 (1.42)-1.49 (4.41)-0.44[-1.19, 0.09]

Generally, Neural-qq is close to CoBiST without significant difference. Detailed results of other scores and raw metrics are reported in Appendix E.1.

We further isolate the components inside CoBiST and compare the performance. We summarize the result in the following points:

  • Bidirectionality. We compare the bidirectional design with past-to-future only and future-to-past only transport. The preferred direction varies across datasets: some may prefer future-to-past route, and perform worse on past-to-future transport, while some are just the opposite. In comparison, the bidirectional design avoid pre-selection on the transport direction. It may not be the best one but it is more stable and reliable.
  • Fixed-retention. The performance of this variant is highly sensitive to the pre-determined retention. It's much more efficient if the retention can be set correctly, but it faces high risk of degradation as it's hard to specify the retention values without further tuning and validation.
  • Decoupled-budget. Refer to the revised Neural-qq ablation study. We don't expect the geometric-mean design to outperform the complex neural networks, but it's performance is competitive, supporting the geometric mean as a simple, semantically interpretable choice with competitive accuracy.

W2 / Q2: Time and memory diagnostics with longer horizons

We examine the computational efficiency of CoBiST on simulated AR(1)AR(1) data. A CoBiST training update takes longer than Dense or Bi-LSTM at shorter horizons, but its slower growth rate makes it lower than Dense's and close to Bi-LSTM's at larger horizons. Inference is slowest for Dense, followed by CoBiST and Bi-LSTM. Bi-LSTM has the highest peak memory usage. CoBiST's memory usage grows more slowly than Dense's and is more efficient at larger horizons. The results generally agree with the theoretical analysis in Table 2. Details are reported in Appendix E.5.

AuthorsOpenReview ↗

Response to Reviewer ve1X Part 2/2

W3 / Q3: Four-seed stability and dynamic-gate intervention

We repeated the diagnostics using selected checkpoints on all four formal seeds in this version for the intervention-based validation experiment. Figure 4 gives dataset-level four-seed means, and Table 17 gives dataset-macro means and 95% hierarchical intervals over datasets and seeds. For GAN, Flow, and Diffusion, respectively, source-mass/readout-response Spearman correlations are 0.86 [0.81, 0.90], 0.69 [0.62, 0.75], and 0.73 [0.66, 0.79], while memory correlations are 0.71, 0.34, and 0.51, and direction correlations are 0.83, 0.64, and 0.70. These correlations indicate positive associations between the transport quantities and the measured responses.

We change the frozen gate analysis in Figure 2(d) and Appendix D.4 to allow the states and gates to respond to previous deletions, and report the final PCRPS score change. At intermediate deletion fractions, high-mass deletion produces a larger change in final-sample PCRPS on average than random or low-mass deletion. Thus, the kernel can rank source contributions whose removal has a greater effect on the final forecasting score, providing a practical basis for targeted ablations and sensitivity analysis, and therefore for studying the internal interaction within the future generated path.

W4 / Q4: Forecasting uncertainty, VS, and boundary correlations

We agree that average ranks do not establish superiority on every dataset. DeepAR has the best mean PCRPS on Electricity, Exchange, Traffic, and Taxi, while CoBiST-Flow and CoBiST-Diffusion have the best two average PCRPS ranks (1.88 and 2.38). We also revised the GAN design: the discriminator now uses the same mixer architecture as the generator, and the discriminator's parameters are included in tuning and capacity. The Wind-GAN average VS improvement rate is now positive but with high variance. We acknowledge the high instability of GAN models (for all of Dense, Bi-LSTM, Neural-qq, CoBiST architectures) and such instability becomes serious for harder task like the Wind dataset.

Finally, thank you also for catching the original boundary-correlation mismatch because of the previous wrong version control. The revised four-seed analysis consistently reports 0.94, 0.93, and 0.88 for GAN, Flow, and Diffusion in both the text and Table 17 (different from before after the above-mentioned modifications).

We thank the authors for the constructive discussion. The revision clearly improves several points: the four-seed intervention and dynamic source deletion address the frozen-state limitation; the boundary correlations are now consistent between the text and Table 17; and the added controls and horizon-scaling curves are useful.

However, the central concern remains. I accept that the bidirectional design avoids pre-selecting a transport direction; precisely for that reason, the key question is how efficiently and interpretably bidirectional information is used. Bi-LSTM is a valid recurrent control; a more informative baseline would be an attention-based or SSM-style mixer, which lets content implicitly decide which positions to read from, parallelizes better at short horizons, and is closer to modern practice.

Second, Neural-q only replaces the estimator of q; the conservation π0 = 1 − q remains. It therefore ablates the geometric-mean coupling, not the conservation itself, so whether the conserved coupling is useful is still unanswered.

Third, Table 3 reveals an asymmetry the paper does not discuss: relative to Bi-LSTM, Wind accounts for about 77% (Flow) and 82% (Diffusion) of the summed per-dataset gains, with only ~1.1% mean gain on the other seven datasets, whereas GAN is the opposite (Wind −0.76%, non-Wind mean 16.78%). Relative to Dense, Flow/Diffusion gains are broad, but GAN's 4.91% is about 65% from Wind and its CI crosses zero. The paper only reports a special Wind GAN tuning grid and does not explain this asymmetry; the response's note on GAN instability does not explain the Flow/Diffusion–Bi-LSTM gap. I would ask for a Wind-excluded sensitivity analysis and an explanation of why Bi-LSTM is especially poor on Wind for Flow/Diffusion.

Finally, I would encourage the authors to add the requested component ablations, and to reposition the coupling as an interpretability contribution rather than an accuracy claim.

I also note for the AC that the authors uploaded a substantially revised PDF; the large scope of changes should be factored into the decision.

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

This paper proposes Coupled Bidirectional Survival Transport (CoBiST), a generator-agnostic sequence-mixing operator for probabilistic time-series forecasting. CoBiST combines causal and anti-causal recurrent scans with a direct-local branch through a normalized routing mechanism. Specifically, the geometric mean of two directional retention probabilities determines the recurrent transport budget, while learned directional preferences distribute this budget between the two scans.

The authors establish several theoretical properties, including an augmented row-stochastic transport kernel, a rank-one semiseparable representation, frozen-gate non-expansiveness, and a finite-horizon sensitivity bound for state-dependent gates. The proposed operator is integrated into GAN, flow-matching, and diffusion-based forecasters.

Experiments on eight datasets report average PCRPS improvements of 6.72%, 21.48%, and 20.78% over capacity-matched Dense controls for the three respective generator families. CoBiST-Flow and CoBiST-Diffusion also achieve competitive average rankings against six existing probabilistic forecasting methods.

Critical strengths

The paper presents a coherent framework that combines directional propagation, adaptive retention, and local routing within an explicitly normalized operator. The augmented kernel representation (Theorem 2 and Proposition 3) provides exact source-level coefficients, while Theorem 3 establishes an efficient semisepearable realization. This theoretical accounting provides a transparent description of a sequence-mixing mechanism that would otherwise be represented primarily through implicit neural interactions.

The evaluation demonstrates improvements over the specified Dense controls across GAN, flow-matching, and diffusion-based generators. In particular, the Flow and Diffusion implementations show substantial improvements over their matched controls on several datasets (Table 5), and achieve competitive average PCRPS rankings against the evaluated forecasting baselines (Table 3)

This cross-generator applicability suggests that the proposed operator may be useful as a reusable architectural component rather than a modification specific to one generative objective.

Critical weaknesses

W1. Limited novelty and insufficient ablation of the proposed coupling mechanism. The paper does not sufficiently distinguish its contribution from existing bidirectional recurrent and structured sequence-mixing methods. Bidirectional propagation, gated retention, and efficient recurrent mixing have already been extensively studied in prior works, including Mamba-2 and Hydra.

The main contribution of CoBiST therefore appears to be the specific coupling between directional retention probabilities and the recurrent routing budget. However, the experiments do not adequately demonstrate the necessity of this coupling.

In particular, the Neural-q ablation in Appendix D learns a residual around the proposed geometric-mean budget rather than using a fully independent routing mechanism. Furthermore, Section 5 mentions that alternative coupling methods were omitted because preliminary experiments showed only marginal differences.

Consequently, it remains unclear whether the reported improvements arise from the proposed coupling or simply from bidirectional recurrent information exchange. Comparisons against a simple bidirectional recurrent baseline and independently learned routing would be important for establishing the novelty and significance of the proposed method.

Questions for authors

Q1. Necessity of the proposed coupling mechanism.

Could the authors provide ablation studies comparing CoBiST against a simple bidirectional recurrent baseline and a model with an independently learned routing budget?

The current Neural-q ablation learns a residual around the proposed geometric-mean budget, rather than using a fully independent routing mechanism. Therefore, it remains unclear whether the performance improvements originate from the proposed coupling mechanism or simply from bidirectional recurrent information exchange.

How do the authors justify the necessity of the proposed geometric-mean coupling over these simpler alternatives?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ugtp

We thank the reviewer for recognizing our logical coherence and applicability, and highlighting the need to distinguish the coupled mechanism from existing bidirectional and structured mixers.

Novelty and bidirectional recurrence

We agree that bidirectionality, gated retention, and efficient recurrent or structured mixing are established ideas. However, such models normally work on a known past sequence and are used for future forecasting. We emphasize the bidirectionality inside the forecast horizon for the generator, and our contribution is the specific shared accounting that connects directional survival, the recurrent-versus-local budget, and an exact augmented source kernel.

To verify the contribution of coupling, we added an independently tuned, approximately capacity-matched Bi-LSTM control to each generator family, where Bi-LSTM worked as a sequence-to-sequence mixer without ground-truth in future observations inside the forecast horizon, and CoBiST's corresponding average improvements over Bi-LSTM are 14.58%, 4.16%, and 5.36% for GAN, Flow and Diffusion, respectively with significance.

Table 3 gives the following PCRPS improvement rates (%), mean (std), across all eight datasets of CoBiST over Bi-LSTM:

Model comparisonElec.Exch.TrafficM4WikiTaxiWindWeatherMean gain[95% CI]
CoBiST-GAN vs Bi-LSTM7.05 (13.35)31.23 (9.74)13.13 (13.58)22.58 (5.15)7.82 (4.36)14.03 (8.93)-0.76 (15.35)21.59 (35.46)14.58[6.63, 22.78]
CoBiST-Flow vs Bi-LSTM0.74 (0.75)2.95 (2.18)-1.07 (0.79)3.23 (1.62)1.55 (1.38)1.89 (1.07)25.49 (0.47)-1.53 (2.37)4.16[0.30, 11.09]
CoBiST-Diffusion vs Bi-LSTM1.60 (0.55)0.73 (2.53)0.85 (0.38)4.65 (1.55)1.56 (0.68)-0.05 (0.47)35.11 (2.04)-1.54 (2.65)5.36[0.20, 15.06]

These comparisons support CoBiST's advantage over bidirectional recurrent mixer. Detailed results of other scores and raw metrics are reported in Appendix E.1.

Independent routing budget

We agree that the original Neural-qq control that learns the residual was less convincing. In the revision, the new revised Neural-qq now estimates qq directly by a network taken local and context embeddings as input. It is independently tuned fairly under the same design (Appendix D.3).

Table 3 gives the following PCRPS improvement rates (%), mean (std), across all eight datasets of CoBiST over the revised Neural-qq:

Model comparisonElec.Exch.TrafficM4WikiTaxiWindWeatherMean gain[95% CI]
CoBiST-GAN vs Neural-qq1.03 (7.36)12.30 (18.98)-1.99 (2.31)8.33 (6.71)0.92 (8.64)-0.46 (3.50)2.44 (28.74)-9.75 (4.57)1.60[-3.79, 7.24]
CoBiST-Flow vs Neural-qq-0.05 (0.47)4.40 (3.99)-0.71 (0.73)-0.20 (0.59)0.43 (0.34)0.18 (0.63)3.94 (1.93)-0.11 (1.79)0.98[-0.19, 2.46]
CoBiST-Diffusion vs Neural-qq0.03 (0.26)-0.63 (1.68)0.05 (0.20)-0.45 (1.24)-0.21 (0.47)0.05 (0.12)-0.89 (1.42)-1.49 (4.41)-0.44[-1.19, 0.09]

Generally, Neural-qq is close to CoBiST without significant difference, supporting the geometric mean as a simple, semantically interpretable choice with competitive accuracy. Detailed results of other scores and raw metrics are reported in Appendix E.1.

Why the geometric mean?

We use geometric mean to keep the designed probabilistic semantics in survival and hazard analysis. In empirical part, the Neural-qq comparison shows that geometric mean is a simple, semantically interpretable choice with competitive accuracy.

Thank you for the detailed response and the new Bi-LSTM and independent Neural-q controls, which directly address the ablations I requested. However, these results also clarify the source of the gains. Excluding Wind, CoBiST improves over Bi-LSTM by only about 1% for Flow and Diffusion, and Bi-LSTM is better on Traffic and Weather. The independently estimated routing budget also performs on par with the geometric-mean coupling. This suggests that the accuracy gains come mainly from bidirectional mixing, while the proposed coupling contributes mainly interpretability. Comparisons with structured bidirectional mixers such as Hydra are still absent. Since my concern about the necessity of the coupling remains, I keep my rating of 2.

Thank you for your reply. We do admit our improvement is marginal over baselines. However, our model, CoBiST-Flow and CoBiST-Diffusion have already got the best 2 average ranks among all evaluated modern probabilistic time-series forecasters.

This is because using a mixer for latent or intermediate representations across the generated forecast horizon is itself a novel aspect of our contribution. CoBiST becomes moderate in comparison to the ablation variants, like Bi-LSTM, Transformer, and Hydra or something else because some of our contributions have been absorbed into the mixing mechanism within the forecast horizon.

We may be not able to finish more ablation comparison during the discussion period, and currently, we use the result for Bi-LSTM and Transformer for reference. CoBiST still outperforms Bi-LSTM on 6 of 8 datasets for both Flow and Diffusion, and 7 of 8 for GAN. Against Transformer, the corresponding counts are 5 of 8 for both Flow and Diffusion, and 6 of 8 for GAN, with significant advantages in time and memory efficiency. Therefore, CoBiST still has potential for future applications, with interpretability as an additional advantage.

Interpretability is our key contribution, and this is useful in time series forecasting. Compared with mainstream generative tasks such as image and video generation, time-series forecasting often places greater emphasis on statistical inference. For example, it is difficult to interpret the meaning of a parameter in an image-generation model, whereas parameters in time-series models can have more concrete meanings and can be used for subsequent diagnostics and statistical inference.

CoBiST offers these advantages. Its metrics have more concrete meanings, and, as demonstrated in our experiments, we can verify how well they align with time-series features through visualization. In comparison, for complex Transformer networks, it is almost impossible to identify the temporal meaning of the parameters in Q, K, and V.

Interpretability also offers the following potential applications:

  • Prediction auditing: By examining the retention, past-to-future, and future-to-past weights at each prediction step, we can determine whether the model relies primarily on long-term history, local information, or a specific transmission direction.
  • Monitoring: We can empirically estimate the statistical ranges of relevant quantities. After deployment, we can diagnose whether serious anomalies have occurred by monitoring these quantities.
  • Error diagnostics: When predictions deteriorate over a specific period, users can investigate whether the issue stems from excessively low retention, an imbalance between transmission directions, or an insufficient local budget. Such diagnostics help users determine whether the problem arises from historical dependence, directional selection, or short-term variations.
  • Model and configuration selection: The associated interpretable parameters are also helpful in determining whether bidirectional transmission, stronger historical memory, or greater local correction capability is needed.

Thank you for your time and valuable comments to help us improve. This response is intended to clarify our contributions and demonstrate the future potential of CoBiST. We hope this helps you better position our contributions and reconsider your assessment accordingly.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Thank you all for helping us improve!

We want to express our heartfelt gratitude to the AC and all the reviewers for their insightful comments and questions, which have helped us improve our theoretical framework and the experimental design, thereby making the paper more comprehensive and convincing.

We made the following key changes (major changes in the main text are marked in red):

Fair and efficient experimental design

  • We optimized the CoBiST and ablation models using CUDA Graphs and reran corresponding experiments.
  • We revised the GAN design: the discriminator now uses the same mixer architecture as the generator, and the discriminator's parameters are included in tuning and capacity.
  • Some results may be different from the original version due to the above-mentioned modifications.

More ablation study

  • We add a generator with bidirectional LSTM (Bi-LSTM) mixer as an ablation baseline to verify the contribution of coupling. CoBiST's corresponding average improvements over Bi-LSTM are 14.58%, 4.16%, and 5.36% for GAN, Flow and Diffusion, respectively. Table 3 also reports 95% hierarchical bootstrap intervals.
  • We replaced the former residual Neural-qq with an independently estimated routing budget by networks and it performs close to CoBiST without significant difference (Table 3), supporting the geometric mean as a simple, semantically interpretable choice with competitive accuracy.
  • Raw control scores are reported in Table 9, other metrics in Tables 6–8 and 10–12. We also report the parameter counts in Table 13.

Revised intervention study

  • We repeated the interpretability analysis on the four formal training seeds 2027–2030. The result is stable.
  • Instead of using frozen gate, Figure 2 (d) now measures final-sample PCRPS change with dynamic gate after ordered removal.
  • Table 17 gives dataset-macro hierarchical intervals for the fixed-state statistics.

New theory and experiment: Sensitivity study

  • We previously derive the frozen-gate theory because we didn't add assumption on the gate. We realized that the frozen-gate theory is less applicable, so we have replaced it with a stronger conditional result with Lipschitz assumptions.
  • We test whether CoBiST attenuates perturbations when its gates adapt to the perturbed input, within a single transport operation and through complete sampling. We first perturb the embedded horizon representation by 10−510^{-5} to 10−110^{-1} of its norm and recompute all gates. All dataset-wise median transport response ratios remain below 0.90.9 and vary little with perturbation size (Figure 3 (a)--(c)). Next, we perturb the initial Flow or Diffusion sampling state by 1%1\% and track the distance between paired runs, reported in Figure 3 (d) and (e). Generally, these results support CoBiST's ability to combine adaptive cross-horizon routing with attenuation of the tested perturbations across all three generator families, and most evaluated Flow and Diffusion samplers also attenuate initial-state perturbations over complete sampling (except amplification for Weather). We also examine the contribution of gate feedback through reference-gate replay controls in Appendix E.4.

New experiment: Computational study

  • We examine the computational efficiency of CoBiST on simulated AR(1)AR(1) data. CoBiST's training time grows more slowly than Dense's: although CoBiST is slower than both Dense and Bi-LSTM at shorter horizons, it becomes faster than Dense and approaches Bi-LSTM as the horizon increases. For inference time, Dense is the slowest, followed by CoBiST and then Bi-LSTM. Peak memory is substantially higher for Bi-LSTM than for the other two methods, while the slower growth rate of CoBiST makes it markedly more memory-efficient than Dense at larger horizons. We plot the results in Figure 8 in Appendix E.5.
AuthorsOpenReview ↗

Summary of Discussion

Since our contributions focus on applying generative models to time series, our model emphasizes statistical inference and interpretability, which distinguishes it from standard generative learning tasks. Therefore, we would like to clarify our novelty, contributions, and the future potential of CoBiST.

Novelty

Using a mixer for latent or intermediate representations across the generated forecast horizon is itself a novel aspect of our work. As mentioned in the introduction, most generative models, such as those used for image generation, do not focus on the internal dependencies within the forecast horizon.

In fact, our implementations of Bi-LSTM and Transformer differ from their standard architectures, as we modified them to act as horizon mixers like CoBiST for the ablation study. These architectures are otherwise seldom used for this purpose, as they cannot be directly inserted at this position without adaptation.

For model performance, CoBiST can be inserted into multiple kinds of generators without changing their general architecture. CoBiST-Flow and CoBiST-Diffusion achieve the best average ranks among all evaluated modern probabilistic time-series forecasters. While the performance margins over Bi-LSTM and Transformer appear relatively modest compared to the overall gains against other baselines, This is because some of our contributions have been absorbed into the mixing mechanism within the forecast horizon. Even so, CoBiST outperforms Bi-LSTM on 6 of 8 datasets for both Flow and Diffusion, and 7 of 8 for GAN. Against Transformer, the corresponding counts are 5 of 8 for both Flow and Diffusion, and 6 of 8 for GAN, while offering significant advantages in time and memory efficiency. Therefore, CoBiST holds strong potential for future applications, with interpretability as a key advantage.

Key Contribution: Interpretability and Its Importance

Interpretability is our core contribution and is particularly valuable in time-series forecasting. Unlike mainstream generative tasks such as image and video generation, time-series forecasting often places greater emphasis on statistical inference. For instance, it is difficult to interpret individual parameters in an image-generation model, whereas parameters in time-series models can carry concrete meanings for subsequent diagnostics and statistical inference.

CoBiST provides these capabilities. Its internal parameters have clear temporal meanings, and as demonstrated in our experiments, we can even verify how well they align with time-series features through visualization. In contrast, for complex Transformer networks, it is nearly impossible to identify the temporal meaning of the parameters in Q, K, and V.

Interpretability also enables the following practical applications:

  • Prediction auditing: By examining the retention, past-to-future, and future-to-past weights at each prediction step, we can determine whether the model relies primarily on long-term history, local information, or a specific transmission direction.

  • Monitoring: We can empirically estimate the statistical ranges of relevant quantities. After deployment, practitioners can diagnose whether serious anomalies have occurred by monitoring these quantities.

  • Error diagnostics: When predictions deteriorate over a specific period, users can investigate whether the issue stems from excessively low retention, an imbalance between transmission directions, or an insufficient local budget. Such diagnostics help identify whether the problem arises from historical dependence, directional selection, or short-term variations.

  • Model and configuration selection: The associated interpretable parameters also help determine whether bidirectional transmission, stronger historical memory, or greater local correction capability is needed for a given task.

Finally, as the discussion period is ending, we would like to once again express our sincere gratitude to the ACs and reviewers for helping us improve our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.