acceptodds
Withdrawn from ICLR 2027

From Temporal Representation to Point Prediction: Auditing Sensitivity to Event–Time Reassignment Across Prediction Stages

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Temporal point process (TPP) models are evaluated through predictive distributions, decoded point forecasts, and information accessible from hidden representations, but sensitivity at these stages need not agree. We introduce a same-target audit that reassigns historical inter-event times while preserving event identities and order, the interval multiset, total duration, and the future target. The audit separates raw-history predictability, representation accessibility, native distributional sensitivity, prediction movement, and decoded point-error response. Across five TPP architectures on three cascade datasets, SAHP and AttNHP show strong representation-access and positive native-likelihood responses, yet their operational point-MAE effects are weak or negative. Matched fixed-rate re-decoding shows that the endpoint depends strongly on the selected point functional: on Retweet, conditional-median and conditional-mean readouts restore positive responses, whereas on Cyberbullying and YouTube the conditional median remains negative while the conditional mean is positive. In contrast, S2P2 retains positive point-error responses under both its operational conditional-mean readout and a conditional-median control across all three datasets. These results show that temporal sensitivity can persist in learned representations and predictive distributions while being attenuated, reversed, or preserved at the final point readout. The effect of the point functional is itself model- and dataset-dependent, so temporal reliance should be audited across prediction stages rather than inferred from any single representation, likelihood, or point-error measure.

Primary area:
learning on time series and dynamical systems
Keywords:
temporal point processes, temporal representation learning, event-time reassignment, model auditing, point prediction, distributional prediction

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_EJFqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how sensitivity to historical event-time assignment propagates through a trained TPP's prediction pathway. The authors introduce a same-target intervention audit: historical inter-event intervals are permuted among event positions while preserving event identities and order, the interval multiset, total duration, and the (log-transformed) future target. Sensitivity is measured at four stages: raw-history predictability (S_raw), representation access via a standardized 65-parameter linear probe (S_access), native distributional response (S_NLL), and decoded point-MAE response (S_MAE), plus a prediction-drift measure D_point. Experiments cover five architectures (THP, NHP, SAHP, AttNHP, S2P2) on three retweet-cascade datasets (Retweet, Cyberbullying, YouTube), with a MIMIC-IV cross-domain check. Main findings: (i) SAHP/AttNHP show strong access (0.91/0.83) and positive native-NLL responses yet near-zero or negative operational point-MAE effects (-0.010/-0.017), while THP, NHP, and S2P2 retain positive effects at all stages; (ii) re-decoding the same fitted Exponential rates with the conditional median or mean reverses the Retweet endpoint (+0.35/+0.86), but the median stays negative on Cyberbullying/YouTube, so readout dependence is model- and dataset-dependent; (iii) a THP/AttNHP × Gaussian/Exponential factorial associates the endpoint contrast with the joint native pathway; (iv) the controls (graded reassignment, same-mark reassignment, frozen-representation and distillation controls, per-example cancellation analysis) support the stage-wise interpretation.

Critical strengths
  • A well-controlled, multi-stage audit design that cleanly separates representation access, distributional response, and point-error response. This matters because the paper's central negative result (that MAE-based endpoints can report "no temporal reliance" while representations and likelihoods remain sensitive) is only credible under exactly this kind of paired, same-target, multi-stage protocol with matched decoders. The fixed-rate re-decoding experiment (Table 2) is the decisive piece of evidence, and it is done properly.
  • A practically important warning for TPP evaluation. The paper shows that the widely used SAHP/AttNHP operational readouts produce near-constant point predictions, and that switching to the conditional mean or median of the same fitted rates changes both accuracy (2.47 to 1.04 clean MAE) and measured sensitivity. This is a concrete, actionable finding for anyone benchmarking TPPs on point-error metrics.
Critical weaknesses
  1. The contribution is an audit of existing models, and part of the headline phenomenon admits a more mundane explanation. Once it is documented that the SAHP/AttNHP operational decoder outputs nearly constant predictions (std ratio 0.18) with 2.4x worse clean MAE than THP, "no MAE response to input perturbation" is close to a tautology for that decoder. The paper's own decoder-swap analysis shows the interesting structure lives upstream; the framing would be stronger if it led with the calibration/pathology finding rather than the stage-dissociation framing.
  2. Limited scope of conclusions. The main evidence rests on three same-family cascade datasets plus one clinical construction; the cross-dataset inconsistency of the conditional-median result is reported but not explained (calibration vs. misspecification vs. dynamics is left open). This caps how much the paper can teach beyond the specific models and datasets audited.
Questions for authors
  1. For SAHP/AttNHP under the operational readout, what fraction of the near-zero S_MAE is attributable to the near-constant prediction range vs. genuine sign cancellation? The per-example analysis (Appendix Table 14) seems to bear on this; can you decompose it quantitatively?
  2. The operational decoder log(1+1/lambda) is neither the mean nor the median of the fitted Exponential in y-space. Is there any training-objective sense in which it is optimal, or is it simply a heuristic inherited from prior codebases? This affects whether the finding is "models are insensitive" or "a common decoder is miscalibrated."
  3. Can you add the prospective conditional-median decoder for NHP (noted as not evaluable) or otherwise complete the decoder-family grid, so that the cross-model readout comparison is symmetric?
  4. A brief discussion of whether the audit's S quantities relate to proper scoring rules (e.g., is S_MAE under the conditional mean the "right" pairing for MSE-optimal readouts?) would connect the framework to forecast-evaluation theory more tightly.
Reviewer_1eguOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether sensitivity to historical event timing is preserved across the stages of a temporal point-process (TPP) prediction pipeline. It introduces an intervention that permutes historical inter-event intervals while preserving event identities and order, the interval multiset, total history duration, and the future target. The proposed audit separately measures (i) predictability from raw histories, (ii) same-target accessibility through a standardized linear head on frozen representations, (iii) response of the native predictive likelihood, (iv) movement of decoded predictions, and (v) change in point-prediction MAE.

The empirical study covers THP, NHP, SAHP, AttNHP, and S2P2 on Retweet, Cyberbullying, and YouTube cascade data, with additional controls and a construction-specific MIMIC-IV check. The main finding is a stage-wise dissociation: SAHP and AttNHP show substantial representation-access and native-likelihood responses to event-time reassignment, but weak or negative MAE responses under their operational point decoders. On Retweet, re-decoding the same fitted exponential rates with the conditional median or mean restores positive MAE responses, with the mean producing the largest effect. This decoder result is not universal: on Cyberbullying and YouTube the median response remains negative, while S2P2 and NHP retain positive sensitivity under alternative readouts. Factorial, same-mark, graded-intervention, metric-space, and training controls support the conclusion that representation access, distributional sensitivity, prediction movement, and aggregate point-error response should not be treated as interchangeable evidence of temporal reliance.

Critical strengths
  1. The paper makes an important conceptual distinction between several quantities that are often conflated. The same paired intervention is propagated through representation access, native likelihood, prediction movement, and decoded MAE, and the results demonstrate that these stages can disagree sharply. This is consequential because conclusions about whether a TPP “uses time” can otherwise depend almost entirely on the chosen endpoint.

  2. The empirical controls are unusually thorough. In particular, the matched fixed-rate re-decoding isolates the point functional without retraining; the end-to-end backbone/head factorial, same-mark reassignment, graded interventions, alternative metric spaces, S2P2 comparison, prospective NHP checks, multiple datasets, and five-seed intervals all probe plausible alternative explanations. The paper also appropriately limits its causal claims and reports cases that do not fit a simple universal decoder story.

Critical weaknesses
  1. The intervention can create histories that are substantially out of distribution, so a response to reassignment is not by itself evidence about naturally occurring temporal reliance. Preserving the interval multiset and total duration is useful, but the permutation can disrupt mark-time relationships and local temporal dynamics. The same-mark and blockwise controls help, yet the paper does not quantify intervention plausibility under the fitted data distribution or relate sensitivity to in-distribution perturbations. This limits how far the central audit can be interpreted beyond robustness to the specified reassignment.

  2. The strongest explanatory claim remains incomplete across datasets. Decoder choice explains much of the Retweet attenuation, but the conditional median remains negative on Cyberbullying and YouTube, and the paper does not distinguish calibration error, density misspecification, target geometry, or dataset dynamics as the cause. Moreover, the access head is an auxiliary predictor trained on clean representations; its response establishes accessible target information under distribution shift, not that the native model causally uses that information. These caveats do not invalidate the measurements, but they make the mechanistic interpretation less settled than the extensive analysis may initially suggest.

Questions for authors
  1. Can you quantify how out of distribution the global, blockwise, and same-mark reassigned histories are under an independent density model or a held-out likelihood diagnostic? Would the stage-wise conclusions persist for perturbations calibrated to remain within a high-density region of the clean history distribution?

  2. For Cyberbullying and YouTube, can you provide calibration or conditional-error analyses that discriminate among density misspecification, scale/range compression, and dataset-specific dynamics as explanations for the negative median but positive mean response? This would materially affect how general the proposed decoder explanation is.

  3. How sensitive are the access results to the capacity and training distribution of the probe (for example, a nonlinear but capacity-controlled probe, or a probe trained on a mixture of clean and reassigned representations)? The answer would clarify whether the reported access/native gap reflects representation accessibility or probe extrapolation under the intervention.

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

This paper studies how temporal point process models use historical time information. The authors shuffle past time intervals while keeping the event order and prediction target unchanged. They compare changes in hidden representations, likelihood, and prediction error. The main finding is that these measures can give different results, and the final error depends strongly on how the model gives a point prediction.

Critical strengths
  1. The paper has a clear structure and presents the main results well. The authors also discuss several limits of the experiments.

  2. Comparing different readouts on the same model is a useful check. It directly shows that the readout can change the evaluation result.

Critical weaknesses
  1. The main result depends strongly on a decoder choice that is not well justified. The Exponential head is trained on log-transformed time, but the operational readout applies another log transform. In Table 2, changing only this readout greatly improves clean MAE and changes the response to shuffling. The authors should explain where this decoder comes from and whether it follows the original model implementation.

Meanwhile, the Exponential mean and median only differ by a fixed factor, log⁡2\log 2. Their different MAE responses may come from scaling and poor distribution fit. It is better to provide calibration results or test a more flexible distribution to check this explanation.

  1. The contribution and experimental scope are limited. Earlier work has discussed the match between point forecasts and error measures (Gneiting, 2011), the gap between information in representations and its actual use (Ravichander et al., 2021), and the effects of TPP encoders and decoders (Bosser and Ben Taieb, 2023). The paper cites these studies, but the additional insight is still not clear enough. The main experiments also use three datasets from the same cascade family. The MIMIC experiment does not include the same alternative-readout checks, which limits the support for a broader conclusion.
Questions for authors

Please refer to the questions in the Weaknesses part.

Reviewer_a7ejOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how TPP models respond to historical timing perturbations across different prediction stages.

Critical strengths

Separating representation probing, likelihood, and point MAE under the exact same time-shuffling intervention is a clean way to check where timing information gets lost.

Critical weaknesses

1.The main gap between likelihood and MAE for SAHP and AttNHP seems mostly caused by a flawed default point readout rather than an inherent issue with the learned representations.

2.It is hard to tell if the measured changes reflect how the model normally uses time or just how the probe and decoder react to out-of-distribution inputs.

Questions for authors

On Cyberbullying and YouTube, why does the conditional mean give a positive MAE response while the conditional median stays negative? Could you provide some calibration results or error distributions to show whether this is caused by poor Exponential distribution fit or heavy-tailed targets?

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

Related papers

Discussion (0)

Sign in to comment.