acceptodds
Under review as a conference paper at ICLR 2027

Scan-Parallel Variational Inference for Event-Driven Latent Jump Models

Abstract

Continuous-time latent state-space models are widely used for irregularly observed time series, typically representing uncertainty through a random initial state followed by smooth ordinary differential equation (ODE) dynamics, or Brownian perturbations accumulated by a stochastic differential equation (SDE). While these constructions are natural for smooth or diffusive systems, they do not directly encode a finite set of unobserved shocks whose timing, magnitude, and effect on memory are themselves uncertain. We instead introduce latent jump models (LJMs): the latent state follows closed-form deterministic dynamics between Poisson-driven event times and undergoes a learned instantaneous jump at each event, with state-dependent intensity and mark distributions. This defines a time-homogeneous Markov process whose jumps are intrinsic state transitions rather than numerical integration steps. For training, we derive a tractable event-augmented path-space evidence lower bound (ELBO) and introduce _Scan-Parallel Variational Inference (SPVI)_: an observation-conditioned variational posterior samples the complete latent event configuration before the state recurrence is evaluated using triangular affine scans. This enables fast and scalable training for LJMs in _logarithmic temporal depth_. Across synthetic stochastic systems and real irregular time series, LJMs remain competitive with continuous-time latent baselines in forecast accuracy, often with fewer latent updates than numerical solver evaluations, while training one to two orders of magnitude faster than the sequential baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
latent state space models, irregular time series modeling, marked point processes

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_A6NJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Latent Jump Models (LJMs) with Scan-Parallel Variational Inference (SPVI) for irregular time series forecasting. The framework models latent dynamics through deterministic flows and stochastic events, while using structured affine transitions to enable parallel state computation. Experiments on synthetic and real-world datasets demonstrate competitive forecasting performance and substantial computational acceleration.

Critical strengths
  1. The matched-model runtime experiments demonstrate substantial acceleration in posterior state computation.
Critical weaknesses
  1. The proposed parallelization relies on strong structural constraints, including exponential decay between events and block-triangular affine jump transitions. The experiments do not sufficiently evaluate the resulting loss of expressiveness, and the synthetic event-driven benchmarks do not directly test state-dependent event arrival mechanisms.

  2. The framework reconstructs complete historical latent-event configurations, but its advantage over directly inferring the current latent state for forecasting remains unclear. Without a controlled comparison, the practical necessity of this additional inference process is insufficiently established.

Questions for authors

Could the authors compare the proposed historical event inference with a direct terminal-state inference approach using a comparable forecasting model in the ablation study?

1 comment

We thank the reviewer for their careful assessment of our work. We address the raised points below. Tables R1–R8 are presented in the general response and Table A1 is shown below.

Weakness 1, first part: loss of expressiveness from the exponential decay and the block-triangular jump transitions. We relaxed each constraint separately and measured the change (Section 2 of the general response).

  • Block-triangular jump map (Tables R4, R5). A general jump map of the same size lowers CRPS on deterministic rotation from 0.0855 to 0.0699. On the systems with additive jumps it changes little: 0.2046 against 0.2043 on the state-dependent PDMP, and 0.2261 against 0.2322 on the Hawkes process. The general map cannot use scans and takes about 4.5 times as long to train on the state-dependent systems (Table R5).
  • Decay flow (Table R4). A closed-form flow that can rotate lowers CRPS on deterministic rotation from 0.0855 to 0.0707, at every seed, and uses fewer events (244.5 against 273.4). Each block's recurrence stays affine, so the scans still apply.

The two constraints therefore cost about 18% in CRPS on deterministic rotation, and the rotating flow recovers this while keeping the scans. On the additive-jump systems, the block-triangular map costs at most 2.6% or is better.

Weakness 1, second part: state-dependent event arrival mechanisms. Tables R1 and R2 add two systems in which event arrivals depend on the latent state. In the state-dependent PDMP, the event intensity and the mark distribution are functions of the latent state. In the Hawkes process, each event raises the intensity of later events through an excitation state that is not observed. With every baseline at its best setting, LJM has the lowest CRPS and the highest LPD on both systems: 0.2043 against 0.2222 for the best baseline (NCDSSM) on the PDMP, and 0.2322 against 0.2466 for the best baseline (Latent SDE) on the Hawkes process.

We also tested whether the state dependence of LJM's own prior matters (Table R6). A constant prior rate raises CRPS by 1.9% and 1.5% and lowers LPD from −0.798 to −1.052 and from −2.376 to −3.397. A constant prior mark law raises CRPS by 10% and 13%, and both together by 38% and 67%. The held-out ELBO falls in every case.

Weakness 2 and Question: historical event inference versus direct terminal-state inference. Under our Markov model, future predictions depend on the context through the posterior distribution of the terminal state. We therefore compared LJM with direct terminal-state inference on REFIT (Table A1). The direct baseline uses the same encoder backbone, dynamics and decoder architectures, and prior forecast simulation procedure, but outputs a distribution over the terminal state without inferring historical events. LJM uses a 512-minute context, whereas the fixed-context direct baseline uses 482 minutes to match the training budget.

The direct baseline is trained with a forecast-window ELBO: the expected log-likelihood of the next 30 minutes, minus the terminal-state KL to a reference prior N(0,I)\mathcal{N}(0,I) and the event KL over the forecast window. We also train a variant with the context length sampled uniformly between 30 and 482 minutes.

LJM achieves lower CRPS and RMSE than both direct-state baselines under these protocols. These results support the effectiveness of our path-based learning approach relative to the tested direct-state baselines. Because the context lengths and training objectives differ, this is an end-to-end comparison and does not isolate the contribution of reconstructing historical events.

Table A1. REFIT: LJM against direct inference of the terminal state (test windows; each direct model at its validation-selected learning rate).

ModelCRPS/shs_h ↓RMSE (W) ↓90% coverage
LJM (events inferred over the 512-minute context)0.0495 ±\pm 0.0004551.7 ±\pm 2.90.925
Direct terminal state, context of 482 minutes0.0779 ±\pm 0.0000664.1 ±\pm 0.30.875
Direct terminal state, random context length0.0750 ±\pm 0.0051652.3 ±\pm 20.40.851
Reviewer_gztYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a novel continuous-time latent variable model for time series: a latent sequence of event τn\tau_n, each paired with a "mark" UnU_n, drives jumps in an otherwise deterministically evolving latent state ZZ. At each event time, jumps from Zτn−Z_{\tau_n^-} to ZτnZ_{\tau_n} are determined by a map F(Zτn−,Un)F(Z_{\tau_n^-}, U_n), where UnU_n itself is sampled from a Gaussian distribution conditioned on Zτn−Z_{\tau_n^-}. In between events, the latent state decays exponentially to zero.

In order to make learning tractable, the authors use amortized variational inference. They propose to parameterize directly the rate of the Poisson point process generating the events, as well as the mark distribution. The rates and the mark distribution are optimized for each interval in-between observations. The exponential decay of the latent states, paired with a block-autoregressive parameterization of the map FF, allow for an efficient blockwise scan updating scheme.

In the experiments, the model is compared against other continuous-time models (latent SDEs, NSFs, etc.) and demonstrates good performance across the board.

Critical strengths
  • Latent event models are an underexplored area, and showing that they can match or surpass the performance of continuous-time methods like latent ODEs/SDEs is an interesting finding.
  • The model admits parallel scan updates of its latent states, leading to efficient inference of the latent states. Building an expressive model while respecting these efficiency constraints is a substantial contribution.
Critical weaknesses

Methodology

  • The latent dynamics induced by the method are quite simple, and the paper does not investigate when and where the model breaks. When are these latent events needed? What advantage do they provide, and at what scale? The method is only tested on relatively low-dimensional datasets, so these questions can't be answered by the current experiments.
  • Having uniform intensity in the variational posterior within each inter-observation interval seems quite limiting. I can see how it may work if samples are dense in time, but I would imagine that performance may degrade quickly if observations are sparse in time / noisier. This limitation hinders scalability to scenarios that go beyond the ones considered in this paper.

Experiments

  • Numbers are great, but figures can be a great way to qualitatively assess whether the baselines are well tuned, what specific bias does each method exhibit, and just getting insights into the latent dynamics learned by the model. Other than Fig.1, the paper has no visual analysis of the forecasts or the latent dynamics. The lack of figures for the real-world dataset is unfortunate. This is particularly relevant considering that this method relies on learning latent events, which presumably should be somewhat reflected in the structure of the time series.
  • The experiments do not test long-horizon forecasts at all. The horizons are all quite shorts, and comparisons between models at longer horizons are absent.
  • A numerical solver for ODEs/SDEs can easily break if it's ran with a suboptimal tolerance wrt the stiffness of the dynamics. Figure 1 and Table 1 say that results are reported "at the closest budget". This is not a fair comparison. Baseline should be reported at their best performance, even if they are much more expensive. Where LJM sits in the performance / speed Pareto frontier wrt the models ran at coarser resolutions is a different question, answered in part by Figure 3.

Minor

  • Gradients of the ELBO are biased, and this is not discussed or mentioned in the paper.
  • The paper is not citing Gumbel-Softmax (Jang et al, Maddison et al) in sec 4.4.
Questions for authors
  1. Can the authors provide scaling experiments and figures displaying the qualitative performance of the model?
  2. Can the authors test each model on a sufficiently complex long horizon forecasting task?
4 comments

We thank the reviewer for their careful assessment of our work. We address the raised points below. Tables R1–R8 are presented in the general response and Tables G1–G4 are shown below.

Baselines at their best performance. Table 1 of the revised paper now shows both comparisons: each baseline at its best setting, and each baseline at a setting matched to LJM's budget as closely as possible. We also report every baseline at its best setting below, with its cost reported separately.

  • New systems (Tables R1, R2). We include two new systems to address the reviewer's concerns. The Hawkes process has self-exciting arrivals driven by an unobserved excitation state, and the SD-PDMP (state-dependent) has arrivals and marks that depend on the latent state. LJM has the lowest CRPS and the highest LPD on both systems: 0.2043 against 0.2222 for the best baseline (NCDSSM) on SD-PDMP, and 0.2322 against 0.2466 for the best baseline (Latent SDE) on the Hawkes process.
  • The synthetic systems of the paper (Table G1). Among LJM and the baselines of the paper, LJM has the lowest CRPS on three of the four systems. The Latent SDE is lower on the regular PDMP, by 0.0035. LJM has the highest LPD on all the systems. LJM uses 1.7 to 14 times fewer dynamics updates per context than the best baseline.
  • We will follow this protocol in Table 1 and Figure 1 of the revised paper.

Table G1. The synthetic systems of the paper: LJM against the best baseline, each family at its best setting.

SystemCRPS ↓, LJMCRPS ↓, best baselineBest baselineNFE, LJM / best baseline
Cyclic jumps0.0677 ±\pm 0.00230.0698 ±\pm 0.0041Latent SDE89 / 147
Regular PDMP0.2427 ±\pm 0.00020.2392 ±\pm 0.0040Latent SDE18 / 220
Irregular PDMP0.2323 ±\pm 0.00120.2406 ±\pm 0.0018Latent SDE19 / 63
State-dependent PDMP0.2043 ±\pm 0.00130.2222 ±\pm 0.0013NCDSSM19 / 176
Hawkes process0.2322 ±\pm 0.00090.2466 ±\pm 0.0078Latent SDE16 / 220
Deterministic rotation0.0707 ±\pm 0.0019 †^\dagger0.0715 ±\pm 0.0100NCDSSM245 / 880

†^\dagger LJM with the rotating flow of Table R4. With the decay flow of the paper, LJM scores 0.0855 ±\pm 0.0050 with 273 events.

Methodology, first point; Experiments, first and second points; Questions 1 and 2: dimension, long horizons and figures.

  • Long horizon on the simulated systems (Table G2). With a 44-observation forecast, LJM has the lowest CRPS of all families in every block on the PDMP. On the Hawkes process it is the lowest over observations 1–16, level with the Latent SDE over 17–32, and behind the Latent SDE and the Latent NSF over 33–44 (0.645 against 0.610 and 0.631). Extended to all eight simulated systems, LJM has the lowest or second lowest CRPS over the whole 44-observation horizon on seven systems, among LJM and the five baselines. Its forecasts stay on the scale of the data on every system and seed. In contrast, the Latent SDE (in every seed) and the Latent ODE (in two of three seeds) degrade sharply after the first eight observations on the two cyclic-jump systems, and the Latent NSF diverges on the regular PDMP.
  • Long horizon on REFIT (Tables G3, G4). We extended the REFIT forecast from 30 to 240 minutes and added Neural MJD. LJM has the lowest CRPS at every context length on both horizons. Over 240 minutes, its CRPS/shs_h is 0.0630 to 0.0636, against 0.0639 to 0.0694 for Neural MJD and 0.0709 to 0.0726 for SDE Matching. The Latent NSF is unstable. At 240 minutes, its forecasts are non-finite at contexts 1,024 and 2,048, including the context at which all of its seeds are finite at 30 minutes.
  • Video (Table R3). The Box video benchmark has 1,024 observed dimensions and a 40-frame forecast from a 20-frame context. The results are in Table R3 of the general response.
  • Figures. For REFIT and USHCN, the revised paper shows the forecasts of every model, together with LJM's posterior event rate in Figures 4 and 5.
  • Scaling. Appendix F measures how the posterior computation scales with the number of events: at 4,096 proposals, the scan is 134 times faster than the sequential loop for the same model.

Table G2. Forecast of 44 observations from a 44-observation context: CRPS ↓ by block of forecast observations (mean over three seeds).

ModelPDMP, 1–8PDMP, 9–16PDMP, 17–32PDMP, 33–44Hawkes, 1–8Hawkes, 9–16Hawkes, 17–32Hawkes, 33–44
LJM (ours)0.1940.3220.4190.4730.2380.4410.5790.645
Latent SDE0.2220.3440.4350.5310.2540.4610.5730.610
NCDSSM0.2140.3480.4510.5400.2940.5420.7550.944
Latent NSF0.2360.3530.4480.5060.2690.4660.5850.631
SDE Matching0.2420.3560.4420.4990.2680.4950.6570.761
Latent ODE0.3930.5120.6400.7320.4500.6470.7860.836

Table G3. REFIT, 30-minute forecast: CRPS/shs_h ↓ (mean ±\pm standard deviation over three seeds).

ModelContext 512Context 1,024Context 2,048
LJM (ours)0.0495 ±\pm 0.00040.0492 ±\pm 0.00040.0495 ±\pm 0.0005
Neural MJD0.0520 ±\pm 0.00010.0528 ±\pm 0.00020.0529 ±\pm 0.0002
SDE Matching0.0544 ±\pm 0.00110.0568 ±\pm 0.00060.0555 ±\pm 0.0008
Latent NSF0.0588 ±\pm 0.00180.0599 ±\pm 0.00150.0590 ±\pm 0.0014

Table G4. REFIT, 240-minute forecast, whole horizon (mean ±\pm standard deviation over three seeds).

MetricModelContext 512Context 1,024Context 2,048
CRPS/shs_h ↓LJM (ours)0.0636 ±\pm 0.00080.0630 ±\pm 0.00010.0635 ±\pm 0.0010
CRPS/shs_h ↓SDE Matching0.0709 ±\pm 0.00090.0726 ±\pm 0.00060.0709 ±\pm 0.0012
CRPS/shs_h ↓Latent NSF[0.0747]†non-finite†non-finite†
CRPS/shs_h ↓Neural MJD0.0694 ±\pm 0.00040.0671 ±\pm 0.00070.0639 ±\pm 0.0005
RMSE (W) ↓LJM (ours)621.0 ±\pm 3.5620.0 ±\pm 1.4620.8 ±\pm 2.7
RMSE (W) ↓SDE Matching642.2 ±\pm 2.2647.9 ±\pm 0.2641.8 ±\pm 3.7
RMSE (W) ↓Latent NSF[660.8]†non-finite†non-finite†
RMSE (W) ↓Neural MJD648.5 ±\pm 1.0640.9 ±\pm 2.2629.2 ±\pm 2.2
90% coverageLJM (ours)0.906 ±\pm 0.0130.902 ±\pm 0.0070.905 ±\pm 0.008
90% coverageSDE Matching0.866 ±\pm 0.0080.871 ±\pm 0.0080.911 ±\pm 0.011
90% coverageLatent NSF[0.849]†[0.835]†[0.801]†
90% coverageNeural MJD0.774 ±\pm 0.0050.772 ±\pm 0.0030.773 ±\pm 0.003

† At least one of the three seeds diverges. A value in brackets is the median over the three seeds.

Methodology, first point: when the latent events help and where the model is weakest.

  • Where they help. LJM is strongest when the signal changes through jumps at unobserved times. Removing the latent event times (one event at each observation time, Table R7) lowers LPD and coverage on the three systems whose events fall between observations, and raises CRPS by 6.7% on the state-dependent PDMP. It costs nothing on cyclic jumps, whose events lie on the observation grid.
  • What the events capture. On the two new simulated systems, the posterior event rate is 7.6 and 7.9 times higher on intervals that contain a true event (AUC 0.99 and 1.00), and the learned prior intensity ranks with the true intensity (Spearman correlation 0.90 and 0.87). The model does not recover the number of events: it uses about 18 posterior events per context where the simulators place 2.2 to 2.7. On REFIT, the posterior rate is 1.26 times higher on the 5% of minutes with the largest power change (95% interval 1.17 to 1.36).
  • Competitive on smooth systems. In systems with no true events, LJMs approximate the smooth dynamics with many small jumps. On the rotations it uses 200 to 270 events per context (Table G1). On deterministic rotation, a flow that can rotate lowers its CRPS from 0.0855 to 0.0707 (Table R4). On the videos the ball moves at constant velocity between bounces. LJM represents this motion with about 85 posterior events per 20-frame context, against about 2.5 true bounces. However, even on smooth systems, LJMs maintain competitive accuracy of sequential baselines like Latent SDE and NCDSSM while training an order of magnitude faster.

Methodology, second point: posterior intensity that is uniform within each inter-observation interval. We tested this in two ways.

  • A posterior without this restriction (Table R5). The rate and mark law of the state-dependent posterior read the current latent state, so the rate varies within an interval as the state evolves. Trained with this posterior, LJMs achieve lower CRPS only by 1.2% on the SD-PDMP and by 3.3% on the Hawkes process. Its training takes about 15 times as long on the simulated systems.
  • Sparser and noisier observations (Table R8). When only every 2nd or every 4th observation is kept, LJM's CRPS changes by at most 2.2% (0.2043 to 0.2087 on the SD-PDMP, and 0.2322 to 0.2330 on the Hawkes process). The difference to the state-dependent posterior does not grow (0.0025 to 0.0019 on the PDMP, and 0.0077 to 0.0019 on the Hawkes process), and LJM stays ahead of the Latent SDE at both spacings on both systems.

Other points: gradient bias and citations. The reviewer is right. Eq. (9) is the straight-through estimator applied to a binary Concrete (Gumbel-Softmax) relaxation. The forward pass uses hard events, so the estimate of the objective is unbiased, but the gradient with respect to the posterior rates is a biased surrogate. The paper now states this in Section 4.4 and adds the citations to Bengio et al. (2013), Maddison et al. (2017) and Jang et al. (2017).

I thank the Authors for their response. Despite the substantial additional experiments, many of the new experiments are not included in the revised manuscript. The new experiments are informative, but the motivation for the proposed latent event formulation is still not clearly investigated or discussed, which in my opinion is the most critical aspect of this work. I therefore maintain my original score.

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

This paper proposes Latent Jump Models (LJMs) for irregular time-series modeling, where latent states evolve through deterministic dynamics between stochastic events and undergo learned jumps at event times. To enable efficient training, the authors introduce Scan-Parallel Variational Inference (SPVI), which combines an observation-conditioned event posterior with block-triangular affine transitions, reducing the temporal computational depth from linear to logarithmic. The paper also derives an event-augmented path-space ELBO for variational learning. Experiments on synthetic and real-world datasets demonstrate competitive forecasting performance and substantial training speedups compared to continuous-time latent baselines.

Critical strengths
  1. The most significant contribution of this paper is the integration of structured jump mappings with associative scan operations, enabling parallel computation of state recurrences that would otherwise require sequential processing of events. By partitioning the latent state into ordered blocks and maintaining an affine structure with respect to each block's own state, the authors decompose complex nonlinear state evolution into multiple layers of parallelizable computations. As established in Proposition 1, given a fixed event sequence, the proposed method exactly reproduces the latent state trajectories obtained through sequential computation without introducing additional temporal discretization errors. This design is not only mathematically well-founded but also provides a practical solution to the sequential computation bottleneck commonly encountered in event-driven models.

  2. Compared with models based on smooth continuous dynamics, LJM explicitly models abrupt state changes through latent event times and random marks, making its modeling assumptions more suitable for certain real-world dynamic systems. The experimental results support this modeling approach. For example, on the REFIT dataset, LJM achieves better CRPS, RMSE, and energy forecasting performance across all three context lengths. With a context length of 512, LJM achieves a normalized CRPS of 0.0495, outperforming SDE Matching (0.0544) and Latent NSF (0.0588). These results demonstrate that the proposed method offers not only theoretical computational advantages but also practical value for event-driven time-series modeling tasks.

Critical weaknesses
  • To enable scan-parallel computation, LJM restricts the between-event dynamics to coordinate-wise exponential decay, imposes a block-triangular affine structure on jump mappings, and assumes that the variational event distribution is independent of sampled latent states. While these constraints facilitate efficient computation, they may limit the model's ability to capture complex continuous dynamics and state-dependent posterior dependencies. This concern is supported by the experiments on smooth systems. As shown in Appendix Table 6, LJM generates an average of 44.3 future events per trajectory on the deterministic rotation task without event-count regularization, despite the absence of true jumps. This suggests that the model may rely on frequent artificial jumps to approximate smooth dynamics. Moreover, in Table 1, LJM achieves a CRPS of 0.1012 on Damped Rotation, compared to 0.0743 for the best baseline. These results raise concerns about the trade-off between computational efficiency and modeling expressiveness.

  • The proposed generative prior assumes a bounded state-dependent event intensity and Gaussian event marks. However, real-world abrupt events may be driven by external factors, periodic patterns, or more complex dependencies. Moreover, the event sampling procedure operates over a predefined time horizon, while the paper does not sufficiently discuss its applicability to online settings where observation or prediction intervals may not be known in advance. These assumptions raise concerns about the flexibility and applicability of the proposed event prior to real-world systems.

  • Although the paper discusses several related jump-based models, such as Neural Jump SDEs, Neural Jump ODEs, TriTPP, and NeuralMJP, the experimental comparisons mainly focus on ODE-, SDE-, and flow-based approaches. While these methods do not necessarily share the same problem setting, the lack of comparisons with jump-based alternatives makes it difficult to assess whether the observed improvements arise from explicit latent jump modeling or the proposed model architecture.

  • The paper frames finite, unobserved jumps as its central modeling problem, but the experiments do not validate the corresponding modeling claim. The results show that the proposed scan-parallel architecture can forecast competitively, but not that latent jump modeling is responsible for these gains.

  • The paper is difficult to follow even after careful reading. Its notation and experimental descriptions are inconsistent, and undefined or shifting symbols appear throughout Section 4.

Questions for authors
  1. To enable parallel inference, SPVI assumes that the variational event intensity and mark distribution depend only on observation embeddings rather than sampled latent states. Could the authors compare this approach with a sequential variational posterior that incorporates state dependence, particularly under strong state-dependent event dynamics or sparse observations? Such a comparison would help clarify the trade-off between posterior expressiveness and computational efficiency.

  2. As shown in Appendix Table 6, LJM generates many future events on the deterministic and damped rotation tasks, despite the absence of true jumps. Could the authors investigate whether these events compensate for the limited expressiveness of the coordinate-wise exponential decay dynamics? Would more expressive closed-form between-event dynamics reduce the number of required events while improving predictive performance?

  3. In the REFIT NSF experiments, does the model retain access to the full 512-2048-point conditioning history when training on length-50 subsequences? If not, how should the long-context comparison in Table 2 be interpreted?

2 comments

We thank the reviewer for their careful assessment of our work. We address the raised points below. Tables R1–R8 are presented in the general response and Table J1 is shown below.

Testing LJMs under more complex event dependencies. We include two new systems to address the reviewer's concerns. The Hawkes process has self-exciting arrivals driven by an unobserved excitation state, and the SD-PDMP (state-dependent) has arrivals and marks that depend on the latent state. LJM has the lowest CRPS and the highest LPD on both systems (Tables R1, R2).

Clarifications about the model. The Gaussian marks do not make the jumps Gaussian: a mark is a noise input to the learned jump map, so the jump distribution is a Gaussian pushed through a neural network. The bound on the intensity is a standard assumption used to obtain exact simulation by thinning, and Table R2 tests the bounded model on a Hawkes process, whose intensity is not bounded. We also note that observed external factors can enter the intensity and mark networks as covariates, which is an extension of the present model. Lastly, forecasts need no horizon fixed in advance, because thinning proceeds sequentially and can be continued to any time.

Weakness 1: trade-off between computational efficiency and expressiveness. We agree with the reviewer that the structural restriction of the jump map and state independent posteriors can cost expressiveness. To address these concerns, we relax each of the three constraints separately and directly test them on systems where events depend on the state, isolating the error contributions of each restriction. Please see Section 2 of the general response for implementation details and result tables.

  • Fully general jump map (Tables R4, R5). On the rotation systems, a general jump map lowers CRPS from 0.0855 to 0.0699 on deterministic rotation and from 0.1007 to 0.0857 on damped rotation (Table R4). On the systems with additive jumps it changes only marginally: 0.2046 against 0.2043 on the SD-PDMP, and 0.2261 against 0.2322 on the Hawkes process. The general map cannot use scans and takes about 4.5 times as longer to train (Table R5).
  • State-dependent posterior (Table R5). Our observed improvements with the state-dependent posterior are also only incremental on the 8-step forecast (1.2% on SD-PDMP, 3.3% on Hawkes).
  • Both relaxations together (Table R5). Even with the general jump map trained with the state-dependent posterior, we obtain only a marginally better (2.0% to 3.6%) CRPS of 0.2002 on the SD-PDMP and 0.2238 on the Hawkes process, at about 12 times the training time of LJM.
  • More expressive closed-form flow (Table R4). When we replaced the decay by a closed-form flow that can rotate, CRPS improves from 0.0855 to 0.0707 on deterministic rotation (Table R4). Please see our detailed response to Question 2 below.

In summary, the block-triangular map and the decay flow cost 15% to 18% in CRPS on the rotations, however the rotating flow recovers this while keeping the scans. On the additive-jump systems, the block-triangular map costs at most 2.6% or is better. The observation-only posterior costs only 1.2% to 3.3% in CRPS on the simulated systems, in exchange for training that is about 15 times faster. With both relaxations together, CRPS is 2.0% to 3.6% lower than for the paper's model, at about 12 times the training time.

Question 1: comparison with a sequential, state-dependent posterior. We implemented this posterior (Section 2 of the general response) and compared it with SPVI on the two systems with state-dependent event dynamics, and under sparse and noisy observations.

  • Table R5. When all parameters are trained with the state-dependent posterior, CRPS falls from 0.2043 to 0.2018 on the PDMP (1.2%), and from 0.2322 to 0.2245 on the Hawkes process (3.3%).
  • Sparse and noisy observations (Table R8). Any stark CRPS improvements do not appear due to the state-dependent posterior (over our observation-only posterior) even when the observations become sparser. When every observation, every 2nd and every 4th is kept, the CRPS difference between the two posteriors is 0.0025, 0.0022 and 0.0019 on the SD-PDMP, and 0.0077, 0.0041 and 0.0019 on the Hawkes process.
  • Cost. Training with the state-dependent posterior sacrifices the parallel scans, taking about 15 times longer than SPVI on the simulated systems (83 to 86 against 5.5 to 5.6 minutes, Table R5), because events are sampled and states are computed one at a time. The optimization is also less stable, with large spikes in the gradient norm on the SD-PDMP.

Question 2: do the events compensate for the decay flow, and would a more expressive closed-form flow help? Thank you for this insightful question. Yes, a more expressive flow helps. To show this, we replace the decay by a closed-form rotation flow (Section 2 of the general response). On deterministic rotation (Table R4), CRPS falls from 0.0855 to 0.0707. On damped rotation, CRPS falls from 0.1007 to 0.0855, against 0.0736 for NCDSSM. With this flow, LJM is level with NCDSSM-NL on deterministic rotation (0.0707 against 0.0715). The SPVI scans apply even with the rotating flow, because each block's recurrence stays affine. Therefore, our models train two orders of magnitude faster than NCDSSM.

Question 3: training of the Neural Stochastic Flow on REFIT. On REFIT, the flow is trained on the same batches as LJM and SDE Matching: full context windows of 512, 1,024 or 2,048 minutes, with the loss on the context, for the same number of training iterations and same batch size. At test time also, its encoder and its initial-state posterior read the full context. The comparison in Table 2 is therefore fair in training data and in conditioning. The sentence in Appendix D.1 on sub-sequences of length 50 does not describe the REFIT runs. We apologise for the error, and we will correct the sentence in the revision.

Weakness 3: comparison with a jump-based alternative. These models do not directly apply to our setting. Neural Jump SDEs and TriTPP require observed event times, and NeuralMJP assumes a finite latent state space, so they cannot be run on real-valued series with unobserved jumps. We therefore added Neural MJD (Gao et al., NeurIPS 2025), a neural jump-diffusion forecaster (Table J1). We use the released model with its size matched to the Latent SDE, train it on the same data with the same three seeds. On all four systems whose events occur at random times, Neural MJD's CRPS (↓\downarrow) is 2.4% to 8.2% higher than LJM's. On the cyclic systems, whose jumps follow a fixed schedule on the observation grid, Neural MJD is lower by 12%. On the rotation system, LJM with the closed form rotation flow performs better than Neural MJD. This also agrees with Table R7, where fixing the events at the observation times costs nothing on cyclic jumps: latent event times help where events are random.

Table J1. LJM against Neural MJD: 8-step test CRPS ↓.

SystemEvent timesLJMNeural MJD
Regular PDMPrandom0.2427 ±\pm 0.00020.2550 ±\pm 0.0072
Irregular PDMPrandom0.2323 ±\pm 0.00120.2512 ±\pm 0.0082
State-dependent PDMPrandom0.2043 ±\pm 0.00130.2092 ±\pm 0.0058
Hawkes processrandom0.2322 ±\pm 0.00090.2436 ±\pm 0.0078
Cyclic jumpsscheduled, on the observation grid0.0677 ±\pm 0.00230.0596 ±\pm 0.0012
Deterministic rotationnone0.0707 ±\pm 0.0019 †^\dagger0.0812 ±\pm 0.0013

†^\dagger with closed-form rotation flow. With the diagonal decay flow, LJM obtains a CRPS of 0.0855 ±\pm 0.0050.

Weakness 4: are the gains due to latent jump modeling? We perform ablations that separate the event model from the architecture (Section 3 of the general response).

  • State-independent prior (Table R6). A constant prior rate raises CRPS by 1.9% on the PDMP and by 1.5% on the Hawkes process, and lowers LPD from −0.798 to −1.052 and from −2.376 to −3.397. A constant prior mark law raises CRPS by 10% and 13%. With both constant, CRPS rises by 38% and 67%.
  • Events fixed at observation times (Table R7). This variant keeps LJM's networks but has no latent event times. On the three systems whose events fall between observations, it has lower LPD and lower coverage than LJM, and its CRPS is higher by 6.7% on the state-dependent PDMP and by 1% to 2% on the other two. The difference grows with the horizon: over a 44-observation forecast, CRPS is 0.4751 against 0.3751 on the PDMP. On cyclic jumps, whose events lie on the observation grid, fixing the events costs nothing.

Both parts of the event model therefore contribute. The state-dependent event law is necessary for the advantage: with a constant mark law, LJM is no longer ahead of the best baseline (0.2256 against 0.2222 on the PDMP, and 0.2618 against 0.2466 on the Hawkes process). Latent event times improve the predictive distribution wherever events fall between observations.

Weakness 5: presentation and notation. The notation is now uniform throughout the paper. All changes in the revision are marked in red.

We thank the reviewer again for the opportunity to address their concerns.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response: Common experiments requested

We thank all reviewers for their assessment. Below are the experiments requested commonly in the reviews. We refer to Tables R1–R8 in the individual responses.

All reported metrics are computed on the independent test set with 128 predictive paths unless mentioned otherwise, and are the mean ±\pm standard deviation over three seeds. NFE is the number of dynamics evaluations.

1. New systems with state-dependent event arrivals (Tables R1, R2)

We include two new simulated systems whose event arrival depends on the latent state.

  • State-dependent PDMP (SD-PDMP). A two-dimensional latent state zz decays between events. Events arrive with intensity λ(z)=λmin⁡+(λmax⁡−λmin⁡)sigmoid(w⊤z)\lambda(z) = \lambda_{\min} + (\lambda_{\max} - \lambda_{\min}) \mathrm{sigmoid}(w^\top z), with λmin⁡=0.02\lambda_{\min} = 0.02 and λmax⁡=0.35\lambda_{\max} = 0.35. At an event, a mark U∣z∼N(m+ctanh⁡z,s2I)U \mid z \sim \mathcal N(m + c \tanh z, s^2 I) is added to the state.
  • Hawkes process. An exponential Hawkes process with intensity λ(t)=0.12+e(t)\lambda(t) = 0.12 + e(t). The excitation ee decays at rate 1 and increases by 0.5 at each event (branching ratio 0.5). Each event also adds a mark U∼N(1,0.152)U \sim \mathcal N(1, 0.15^2) to a decaying signal xx. Only xx is observed, with noise, so the excitation is latent.

Table R1. SD-PDMP, every model at its best setting.

ModelCRPS ↓RMSE ↓90% coverageLPD ↑NFETraining (min)
LJM (ours)0.2043 ±\pm 0.00130.4610 ±\pm 0.00230.854−0.798 ±\pm 0.10518.75.6
NCDSSM0.2222 ±\pm 0.00130.4688 ±\pm 0.00240.926−5.011 ±\pm 0.098176746
Latent SDE0.2310 ±\pm 0.00640.4743 ±\pm 0.00830.783−2.251 ±\pm 0.35722071
Latent NSF0.2469 ±\pm 0.00830.4976 ±\pm 0.00770.865−2.918 ±\pm 0.4424339
SDE Matching0.2532 ±\pm 0.00060.5039 ±\pm 0.00180.784−3.474 ±\pm 0.00822014
Latent ODE0.4010 ±\pm 0.02440.6494 ±\pm 0.02040.357−13.569 ±\pm 0.34988089

Table R2. Hawkes process, every model at its best setting.

ModelCRPS ↓RMSE ↓90% coverageLPD ↑NFETraining (min)
LJM (ours)0.2322 ±\pm 0.00090.5950 ±\pm 0.00210.875−2.376 ±\pm 0.41316.05.5
Latent SDE0.2466 ±\pm 0.00780.5873 ±\pm 0.00160.864−3.601 ±\pm 0.09922069
SDE Matching0.2589 ±\pm 0.00320.6079 ±\pm 0.00280.904−3.868 ±\pm 0.2014411
Latent NSF0.2621 ±\pm 0.00090.5966 ±\pm 0.00450.940−2.648 ±\pm 0.0644339
NCDSSM0.2834 ±\pm 0.00560.6343 ±\pm 0.01000.936−9.437 ±\pm 0.164352735
Latent ODE0.4422 ±\pm 0.03680.8165 ±\pm 0.05540.464−21.002 ±\pm 1.29717683

2. High-dimensional dataset (Bouncing Box)

We use the bouncing box dataset from the Pymunk benchmarks of Ansari et al. (2023). Frames are 32 × 32 (1024 dimensions). The context is 20 frames, each missing with probability 0.2, and the forecast is the next 40 frames. The score is the earth mover's distance (EMD) between predicted and true frames, averaged over 1,000 test videos, the 40 forecast frames and 16 sampled futures. CRPS is the per-pixel Brier score summed over the 1,024 pixels of a frame. LJM is trained with SPVI (restricting to the triangular jump map and observation-only posterior) with a convolutional frame encoder, a pixel decoder and a latent dimension of 64. We trained the baseline models with the same encoder and decoder and their own dynamics networks, matching the parameter count as closely as possible. LJM performs the best in terms of the (earth mover's distance) EMD (↓\downarrow, lower is better) metric over 40 future frames on 1000 test reference videos with 16 sampled forecasts. We also compare models only on longer forecast horizons and LJM achieves the lowest EMD on both longer horizons and over all frames while training roughly 2.4×2.4\times faster than the strongest baseline (Latent SDE).

Table R3. Box.

ModelParametersCRPS, frames 1–40 ↓CRPS, frames 31–40 ↓EMD, frames 1–40 ↓EMD, frames 31–40 ↓Training (min) ↓
LJM (ours)2.42M17.69 ±\pm 0.5722.98 ±\pm 0.801.055 ±\pm 0.0321.625 ±\pm 0.07464 ±\pm 1
Latent SDE2.42M21.86 ±\pm 0.7729.60 ±\pm 2.481.116 ±\pm 0.1182.074 ±\pm 0.133153 ±\pm 3
SDE Matching2.91M23.71 ±\pm 0.8923.99 ±\pm 0.142.020 ±\pm 0.0902.208 ±\pm 0.06642 ±\pm 7
Latent NSF2.56M24.03 ±\pm 0.1223.88 ±\pm 0.142.175 ±\pm 0.0312.241 ±\pm 0.028149 ±\pm 24
Latent ODE2.24M23.73 ±\pm 0.0623.82 ±\pm 0.142.037 ±\pm 0.0262.011 ±\pm 0.032179 ±\pm 10

2. Relaxing the structural constraints (Tables R4, R5)

We relax the three structural constraints of our proposed model, i.e., block-triangular jump architecture, observation-only posterior and the diagonal-decay flow, separately to measure the effect of each on accuracy and training time. Table R5 also reports the general jump map and the state-dependent posterior trained together.

  • General jump map. (R,B)=Gθ(u,z~)(R, B) = G_\theta(u, \tilde z) and z=R⊙z~+Bz = R \odot \tilde z + B, where the coefficient network reads the mark uu and the whole pre-event state z~\tilde z. Every jump map can be written in this form. The parameter count is matched to the paper's map, and the states are computed one event at a time.
  • Rotating flow. The coordinates are paired. Each pair has one decay rate γ\gamma and one learned frequency ω\omega. Over a time gap δ\delta, the flow multiplies the pair by e−γδe^{-\gamma\delta} and rotates it by the angle ωδ\omega\delta. Each block's recurrence stays affine, with 2 × 2 blocks, so the scans still apply. The jump map and the posterior are those of the paper, and the starting frequencies are chosen on the validation set.
  • State-dependent posterior. The posterior event heads also read the current latent state: λq(t,z)=Λ(ℓ(hk,z))\lambda^q(t, z) = \Lambda(\ell(h_k, z)) and qtU(⋅∣z)=N(μq(hk,z),diag⁡vq(hk,z))q^U_t(\cdot \mid z) = \mathcal N(\mu^q(h_k, z), \operatorname{diag} v^q(h_k, z)), with z=Zt−z = Z_{t^-} and t∈Ikt \in I_k. The posterior of the paper is the special case in which the heads ignore zz. Events are sampled one at a time by thinning, and the objective is the path-space ELBO of the paper with these rates and mark laws.

Table R4. Rotation systems of the paper

ModelDeterministic rotation: CRPS ↓Damped rotation: CRPS ↓
LJM (paper)0.0855 ±\pm 0.00500.1007 ±\pm 0.0020
LJM, general jump map0.0699 ±\pm 0.00880.0857 ±\pm 0.0093
LJM, rotating flow0.0707 ±\pm 0.00190.0855 ±\pm 0.0024
NCDSSM-NL0.0715 ±\pm 0.01000.0736 ±\pm 0.0011

Table R5. State-dependent systems: LJM, two relaxations, and both together.

QuantityLJM with SPVI (paper)Observation-only posterior + general jump mapState-dependent posterior + block-triangular jump mapState-dependent posterior + general jump map
State-dependent PDMP: CRPS ↓0.2043 ±\pm 0.00130.2046 ±\pm 0.00110.2018 ±\pm 0.00090.2002 ±\pm 0.0010
Hawkes process: CRPS ↓0.2322 ±\pm 0.00090.2261 ±\pm 0.00180.2245 ±\pm 0.00060.2238 ±\pm 0.0012
Posterior events per context, PDMP / Hawkes18.7 / 16.023.9 / 10.78.5 / 3.012.3 / 3.7
Training (min), PDMP / Hawkes5.6 / 5.525.4 / 24.483.1 / 85.866.5 / 68.2

3. Ablations of the event model (Tables R6, R7)

  • State-independent prior. (a) The prior intensity is a learned constant. (b) The prior mark law has a learned constant mean and variance. (c) Both.
  • Events fixed at observation times. The point process is removed: one event occurs at each observation time, with LJM's networks. The marks keep their posterior and their state-dependent prior, and a forecast applies one prior event at each forecast time.

Table R6. State-dependent against state-independent prior event law. The held-out ELBO is the plain ELBO per trajectory on the test set.

Prior event lawPDMP: CRPS ↓PDMP: LPD ↑PDMP: held-out ELBO ↑Hawkes: CRPS ↓Hawkes: LPD ↑Hawkes: held-out ELBO ↑
State-dependent rate and marks (LJM with SPVI)0.2043 ±\pm 0.0013−0.798 ±\pm 0.1059.18 ±\pm 2.220.2322 ±\pm 0.0009−2.376 ±\pm 0.4130.75 ±\pm 4.31
Constant rate0.2082 ±\pm 0.0009−1.052 ±\pm 0.155−2.98 ±\pm 3.000.2356 ±\pm 0.0019−3.397 ±\pm 0.827−6.17 ±\pm 2.43
Constant mark law0.2256 ±\pm 0.0017−1.363 ±\pm 0.098−23.31 ±\pm 11.160.2618 ±\pm 0.0049−3.201 ±\pm 0.408−17.47 ±\pm 1.04
Constant rate and mark law0.2824 ±\pm 0.0053−1.721 ±\pm 0.201−87.05 ±\pm 3.400.3878 ±\pm 0.0243−4.090 ±\pm 1.275−79.74 ±\pm 7.60

Table R7. LJM against the same networks with one event fixed at every observation time.

SystemCRPS ↓, LJMCRPS ↓, fixed eventsLPD ↑, LJMLPD ↑, fixed events90% coverage, LJM / fixed events
State-dependent PDMP0.2043 ±\pm 0.00130.2179 ±\pm 0.0003−0.798 ±\pm 0.105−1.635 ±\pm 0.0780.854 / 0.781
Hawkes process0.2322 ±\pm 0.00090.2344 ±\pm 0.0026−2.376 ±\pm 0.413−3.830 ±\pm 0.4480.875 / 0.830
Irregular PDMP0.2323 ±\pm 0.00120.2361 ±\pm 0.0017−0.642 ±\pm 0.043−0.860 ±\pm 0.0540.849 / 0.825
Cyclic jumps (events on the observation grid)0.0677 ±\pm 0.00230.0657 ±\pm 0.00400.722 ±\pm 0.0170.731 ±\pm 0.0490.872 / 0.841

4. Sparse and noisy observations (Table R8)

Sparsity: only every 2nd or every 4th observation is kept in training and in the context, and forecasts are scored on the original 8 targets. Noise: the standard deviation of the observation noise is raised from 0.1 to 0.2 or 0.4.

Table R8. Test CRPS ↓ under sparser and noisier observations.

SystemModelDense, noise 0.1Every 2nd observationEvery 4th observationNoise 0.2Noise 0.4
SD-PDMPLJM (SPVI)0.2043 ±\pm 0.00130.2060 ±\pm 0.00150.2087 ±\pm 0.00170.2428 ±\pm 0.00090.3395 ±\pm 0.0008
SD-PDMPLJM, state-dependent posterior0.2018 ±\pm 0.00090.2038 ±\pm 0.00300.2067 ±\pm 0.00100.2431 ±\pm 0.00170.3388 ±\pm 0.0012
SD-PDMPLatent SDE0.2310 ±\pm 0.00640.2510 ±\pm 0.01260.2724 ±\pm 0.05020.2742 ±\pm 0.01200.3467 ±\pm 0.0006
HawkesLJM (SPVI)0.2322 ±\pm 0.00090.2303 ±\pm 0.00160.2330 ±\pm 0.00040.2734 ±\pm 0.00210.3920 ±\pm 0.0013
HawkesLJM, state-dependent posterior0.2245 ±\pm 0.00060.2262 ±\pm 0.00080.2311 ±\pm 0.00120.2738 ±\pm 0.00110.3834 ±\pm 0.0031
HawkesLatent SDE0.2466 ±\pm 0.00780.2403 ±\pm 0.00170.2366 ±\pm 0.00120.2839 ±\pm 0.00360.3767 ±\pm 0.0016

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.