acceptodds
Under review as a conference paper at ICLR 2027

DualGround-TS: Addressing Hollowness in Event-Conditioned Time-Series Prediction

96 views

Abstract

Event-conditioned time-series prediction requires models to integrate historical dynamics with textual events, yet correct predictions can arise without dependence on the intended event–history relationship, while plausible explanations can cite evidence that does not support the prediction. We characterize this failure as Hollowness: apparent evidence use does not guarantee behaviorally valid predictive dependence. We introduce DualGround-TS, which addresses prediction and explanation hollowness through two complementary grounding mechanisms. Future-consistent cross-modal grounding uses realized futures to distinguish matched from mismatched event–history states, while behavioral grounding requires explanation-associated temporal and textual evidence to receive corresponding support from prediction behavior. We evaluate DualGround-TS across three complementary regimes: controlled grounding on TimeLitmus, zero-shot contextual forecasting on Context is Key (CiK), and native supervised multimodal forecasting on Time-MMD. On TimeLitmus, DualGround-TS improves all reported metrics in both Finance and Traffic, with its largest gains concentrated on grounding-sensitive diagnostics: relative to Task FT, Event CF-PC improves by 29.6 points, Evidence F1 by 28.9 points, and Avg. CSDR by 15.5 points. On CiK, it reduces zero-shot RCRPS by 14.4% and improves all five context categories. On Time-MMD, it reduces average normalized MSE by 42.6% and achieves lower error in all 9/9 domains. Mechanistic analyses further show stronger selective dependence on predictive evidence and greater matched–mismatched future separation. Together, these results show that DualGround-TS substantially mitigates Hollowness by strengthening both event–history predictive dependence and the behavioral support of explanation-associated evidence, establishing grounded predictive dependence beyond output correctness alone.

Primary area:
learning on time series and dynamical systems
Keywords:
Event-Conditioned Time-Series Prediction, Large Language Models, Multimodal Time Series

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

This paper proposes DualGround-TS for event-conditioned time-series forecasting. It addresses two issues: forecasts may fail to use the relationship between events and historical series, and explanations may cite evidence that does not affect the forecasts. The method uses observed future outcomes to constrain cross-modal representations and tests explanations by retaining or removing cited evidence. It is evaluated on TimeLitmus, CiK, and Time-MMD.

Critical strengths
  1. Compared with standard task fine-tuning of the same backbone, DualGround-TS improves all reported TimeLitmus metrics and reduces forecasting error across all nine Time-MMD domains.

  2. Testing whether forecasts use the relationship between events and historical series, and whether explanations are supported by model behavior, has practical value.

Critical weaknesses
  1. The paper introduces several new terms. In particular, “hollowness” appears to combine shortcut learning and unfaithful explanations without explaining how it differs from these established concepts. This makes the paper harder to follow.

  2. The diagnostic experiments cover only two TimeLitmus domains and one backbone model. CiK and Time-MMD report only forecasting error, leaving it unclear whether the diagnostic findings extend to other tasks or models.

  3. The paper suggests that prior methods focus mainly on forecasting accuracy or the surface plausibility of explanations. However, TimeX and TimePNS have used interventions to assess explanation faithfulness. The paper also omits TESS, which studies whether forecasting models effectively use textual information.

Questions for authors
  1. What does “hollowness” capture beyond shortcut learning and unfaithful explanations?

  2. Do the diagnostic findings hold in other domains and with other backbone models?

  3. How does the paper account for intervention-based faithfulness studies and TESS’s analysis of text use in forecasting?

7 comments
AuthorsOpenReview ↗

Response to Reviewer qiPC [Part 1]

Thank you for your thoughtful and constructive review. Below, we address each concern and summarize the additional evidence and manuscript changes.

W1: The paper introduces several new terms. In particular, “hollowness” appears to combine shortcut learning and unfaithful explanations without explaining how it differs from these established concepts.

We sincerely appreciate your thoughtful comments, especially your recognition that testing whether forecasts truly use the event–history relationship and whether their explanations are supported by model behavior has practical value. Hollowness reveals a shared failure pattern behind shortcut-dependent predictions and behaviorally unsupported explanations: apparent evidence use does not guarantee behaviorally valid predictive dependence. Addressing either failure alone is insufficient for reliable prediction and explanation. Hollowness thus provides a unified reliability goal with two complementary, testable requirements. We have made this contribution explicit in the revised manuscript.

(1) Why do these two failures need to be studied together?

The shared failure pattern manifests in two distinct ways:

  • Prediction Hollowness: A model may predict correctly by relying on event sentiment or a historical trend alone, without considering how the event and history work together.
  • Explanation Hollowness: A model may give a convincing explanation, even though the evidence it cites does not support the predictive role it claims.

A faithful explanation may describe a shortcut-based prediction, while a correctly grounded prediction may still have an unsupported explanation. Neither case ensures reliable prediction and explanation. This motivates two necessary checks: whether predictions depend on the intended event–history evidence, and whether explanations reflect the evidence supporting those predictions.

Figure 1 illustrates both failures: 73.6% of baseline-correct controlled pairs fail the full counterfactual diagnostics, while 81.0% of cited-factor interventions lack the fully expected predictive transition in their respective tests. We have strengthened this explanation in Sec. 1, Sec. 3, and Figure 1‾\underline{\text{Sec. 1, Sec. 3, and Figure 1}}.

(2) What does this unifying view let us do?

It turns a general concern about reliability into two clear requirements that guide the same model:

  • Future Grounding trains the prediction to use the relationship between an event and its historical conditions, using what actually happened afterward as supervision.
  • Behavioral Grounding trains the evidence associated with a generated explanation to have the expected effect on prediction when kept, removed, or accompanied by irrelevant context.

Both operate on the shared representation used for prediction and explanation. The result is a practical way to test where the evidence chain fails and train both parts of it. We clarify these requirements in Secs. 4.3–4.4 and Figure 2‾\underline{\text{Secs. 4.3–4.4 and Figure 2}}.

(3) Do the two requirements improve different behaviors?

We evaluated four variants with the same architecture and 61.74M trainable parameters, changing only the objectives. Complete results appear in Appendix J.1‾\underline{\text{Appendix J.1}}.

VariantEvent CF-PC (%) ↑Sufficiency (%) ↑Comprehensiveness (%) ↑
Task-only29.2266.4123.29
Future-only46.6065.6924.51
Behavioral-only34.4073.4536.89
Full53.7574.5737.93

Findings.

(1) Future-only improves Event CF-PC by 17.38 points, showing better responses to event changes.

(2) Behavioral-only improves Sufficiency by 7.04 points and Comprehensiveness by 13.60 points, showing stronger behavioral support for selected evidence.

(3) Full improves both within the same architecture.

Overall, Hollowness identifies a shared evidence-use failure behind shortcut-dependent predictions and unsupported explanations. This unified view provides two complementary requirements for designing, training, and evaluating more reliable event-conditioned prediction systems.

References

[1] Queen et al. Encoding Time-Series Explanations through Self-Supervised Model Behavior Consistency. NeurIPS, 2023.

[2] Ma et al. Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity. arXiv:2607.21573, 2026.

AuthorsOpenReview ↗

Response to Reviewer qiPC [Part 2]

W2: The diagnostic experiments cover only two TimeLitmus domains and one backbone model. CiK and Time-MMD report only forecasting error, leaving it unclear whether the diagnostic findings extend to other tasks or models.

We sincerely appreciate your recognition that DualGround-TS improves all reported TimeLitmus metrics and reduces forecasting error across all nine Time-MMD domains compared with standard task fine-tuning. These results establish the effectiveness of our framework in the tested settings. The next question is how broadly the evidence-use improvements transfer. We agree that forecasting accuracy and direct grounding behavior require different tests. Our TimeLitmus and held-out interventions directly measure evidence use, and new Qwen3.5-4B and Llama-3.1-8B experiments test it across backbone sizes and families, while CiK and Time-MMD assess predictive usefulness under different tasks.

(1) What our evaluation covers. We evaluate the full TimeLitmus diagnostic benchmark across its Finance and Traffic domains. It directly tests how predictions and explanation-associated evidence respond to controlled changes. CiK covers five contextual forecasting categories under zero-shot evaluation, and Time-MMD covers nine real-world domains under native supervised evaluation. The two external benchmarks test whether the model also forecasts well on independent tasks; they do not directly use TimeLitmus's faithfulness measures. We have clarified this distinction in Secs. 5.1–5.2‾\underline{\text{Secs. 5.1–5.2}}.

(2) Matched comparisons on independent benchmarks. Complete results appear in Appendices J.7–J.8‾\underline{\text{Appendices J.7–J.8}}, with key comparisons in Sec. 5.2‾\underline{\text{Sec. 5.2}}.

VariantCiK Average RCRPS ↓Time-MMD Average Normalized MSE ↓
Task FT0.2853.8067
Architecture-matched Task-only0.2833.4385
Future-only0.2672.8127
Behavioral-only0.2742.9838
Full0.2442.1842

(3) New intervention tests. We also evaluated 200 instances with unseen event paraphrases, irrelevant metadata, lexical distractors, and alternative evidence masks. The detailed procedure and results appear in Appendix J.6‾\underline{\text{Appendix J.6}}.

VariantPrediction Invariance (%) ↑Evidence Stability (%) ↑Explanation Consistency (%) ↑
Task-only85.572.073.0
Full92.085.088.0

Findings.

(1) Forecasting improves on both independent benchmarks. Full reduces CiK average RCRPS by 14.4% versus Task FT, with improvements across five context categories, and Time-MMD average normalized MSE by 42.6%, with lower error in all nine domains.

(2) The model also becomes more stable under new interventions. Compared with Task-only, prediction invariance improves by 6.5 points, evidence stability by 13.0, and explanation consistency by 15.0.

(3) Grounding also improves on additional backbones. We compare Task-only and Full on Qwen3.5-4B and Llama-3.1-8B, using both Finance and Traffic and three seeds per configuration (Appendix J.10‾\underline{\text{Appendix J.10}}). Full improves Sufficiency and Comprehensiveness in all four backbone–domain settings, with all eight paired 95% CIs excluding zero.

Overall, the three benchmarks provide complementary evidence: TimeLitmus directly measures evidence use, while CiK and Time-MMD show useful forecasting benefits on independent tasks and domains. The additional interventions and cross-backbone results strengthen the grounding analysis beyond the original perturbation types and backbone.

AuthorsOpenReview ↗

Response to Reviewer qiPC [Part 3]

W3: The paper suggests that prior methods focus mainly on forecasting accuracy or the surface plausibility of explanations. However, TimeX and TimePNS have used interventions to assess explanation faithfulness. The paper also omits TESS, which studies whether forecasting models effectively use textual information.

We sincerely appreciate your constructive comments and your recognition of the practical value of checking whether predictions use event–history evidence and whether explanations are supported by model behavior. Our contribution is to study these requirements together in the same prediction-and-explanation system. Better text use alone need not rule out a shortcut; behaviorally faithful explanations can still faithfully describe a shortcut-based predictor. We agree that intervention-based explanation methods and prior work on text use should be represented accurately. We have therefore expanded the comparisons with TimeX, TimePNS, and TESS and clarified how our two training objectives address the linked requirements.

(1) What the related methods already contribute.

TimeX [1] trains an interpretable surrogate to preserve important aspects of a pretrained time-series model's behavior. It establishes why an explanation should reflect what the prediction model actually does.

TimePNS [2] uses counterfactual necessity to identify temporal evidence that matters for a prediction. It advances learning and evaluating explanations through interventions.

TESS [3] converts textual descriptions into meaningful temporal signals, helping forecasting models use information such as changes in magnitude, shape, volatility, and lag.

These are important foundations for our work. Our contribution is not the first use of textual context, contrastive learning, or intervention-based explanation methods.

(2) What remains important when these directions are considered together?

Using text effectively does not necessarily mean that a model learns how the same event can have different implications under different historical conditions. And identifying influential evidence does not necessarily mean that a generated explanation correctly describes its role.

More importantly, a faithful explanation may faithfully describe a shortcut, while a well-grounded prediction may still have an unsupported explanation. This is the shared reliability problem captured by Hollowness. It explains why improving either capability on its own is not enough for an event-conditioned prediction-and-explanation system.

(3) What DualGround-TS adds.

Future Grounding keeps the history and observed future fixed while changing the event. The model learns which event–history combination better fits what actually happened afterward, rather than relying on input matching alone.

Behavioral Grounding works within the same model that generates the prediction, natural-language explanation, and selected temporal and textual evidence. It trains the evidence attached to the explanation using the prediction's response to retaining, removing, or adding irrelevant information.

The two objectives are applied to a shared representation, and the architecture-matched ablation in Appendix J.1‾\underline{\text{Appendix J.1}} shows different benefits: Future-only improves Event CF-PC by 17.38 points, while Behavioral-only improves Sufficiency by 7.04 points and Comprehensiveness by 13.60 points.

Findings.

(1) Prior work provides strong foundations for both text use and behavior-based explanation.

(2) Hollowness reveals why neither addresses the whole evidence-use problem in a prediction-and-explanation system on its own.

(3) Our two objectives train the two necessary links together, with separately measurable improvements.

Overall, DualGround-TS connects these research directions through one practical reliability principle: predictions should depend on the intended event–history evidence, and explanations should accurately reflect that dependence. This provides a systematic way to define reliable evidence use, train for it, and test it in event-conditioned forecasting.

References

[1] Queen et al. Encoding Time-Series Explanations through Self-Supervised Model Behavior Consistency. NeurIPS, 2023.

[2] Ma et al. Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity. arXiv:2607.21573, 2026.

[3] Li et al. From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space. ICML, 2026.

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

This paper investigates the problem of shortcut learning and unfaithful explanations in event-conditioned time-series prediction, which the authors term "Hollowness" (specifically, prediction hollowness where models rely on unimodal shortcuts rather than cross-modal interaction, and explanation hollowness where cited evidence does not behaviorally drive the prediction). To address this, the authors propose DualGround-TS, incorporating two grounding mechanisms: (1) future-consistent cross-modal grounding, which contrasts matched and mismatched event-history states against realized futures; and (2) behavioral grounding, which enforces consistency between explanation-cited evidence and actual prediction shifts under perturbation. The framework is evaluated across TimeLitmus, Context is Key (CiK), and Time-MMD.

Critical strengths
  1. Relevant and Practical Problem Formulation: Identifying shortcut learning and unfaithful rationales in multimodal time-series forecasting is well-motivated. Ensuring that models genuinely integrate textual events with temporal dynamics rather than relying on unimodal priors is critical for high-stakes applications.
  2. Strong Empirical Gains Across Diverse Regimes: The proposed method demonstrates consistent and substantial improvements across controlled diagnostic benchmarks (TimeLitmus), zero-shot contextual forecasting (CiK), and supervised multimodal forecasting (Time-MMD).
Critical weaknesses
  1. Over-Complicated Terminology and Conceptual Packaging: Although the paper introduces "Hollowness" as a central problem, the two identified failure modes are closely related to existing concepts of shortcut learning and explanation unfaithfulness. Similarly, future-consistent grounding appears conceptually similar to cross-modal contrastive learning, while behavioral grounding resembles perturbation-based faithfulness regularization. The paper should more clearly explain what is fundamentally different from these existing approaches. At present, the novelty of the proposed framework is not entirely convincing.
  2. Heavy Reliance on Synthetic Perturbations and Counterfactual Priors: Both the training of behavioral grounding and the evaluation on TimeLitmus rely heavily on identifying "evidence spans" and constructing matched/mismatched pairs or counterfactual interventions. In real-world noisy time series and unstructured text, cleanly isolating causal evidence and constructing valid counterfactuals without introducing out-of-distribution artifacts is notoriously difficult. It is unclear how DualGround-TS defines and perturbs these evidence states when explicit ground-truth rationale annotations are unavailable (e.g., on standard datasets like Time-MMD).
  3. Potential Over-Alignment with the Diagnostic Benchmark (TimeLitmus): The training objectives of DualGround-TS (specifically optimizing for matched vs. mismatched separation and intervention sensitivity) appear to directly mirror the exact diagnostic metrics used in TimeLitmus (such as Event CF-PC and CSDR). Training a model directly on counterfactual/behavioral objectives and then evaluating it on a counterfactual diagnostic benchmark naturally guarantees large metric gains (+29.6 points) compared to standard task fine-tuning (Task FT), raising concerns about whether the comparison is fair or simply reflects metric-targeted optimization.
  4. Unclear Trade-off Between Grounding Constraints and Computational/Backbone Overhead: Enforcing dual grounding, especially behavioral grounding, which typically requires multiple forward passes over perturbed inputs or counterfactual candidates, likely introduces substantial training and inference overhead. The paper lacks a clear analysis of the computational cost, training complexity, and sensitivity to the underlying backbone model size.
  5. Subpar Presentation and Readability: From a presentation standpoint, the paper is not easy to follow. The narrative is overly dense and abstract, and Figure 2 is visually cluttered and confusing rather than intuitive, failing to clearly illustrate the concrete mathematical pipeline of the proposed method.
Questions for authors

Please refer to the points raised in the Weaknesses above.

6 comments
AuthorsOpenReview ↗

Response to Reviewer RjuU [Part 1]

Thank you for your detailed and constructive review. Below, we respond to each concern and summarize the changes and additional evidence.

W1: The paper should more clearly explain what is fundamentally different from these existing approaches. At present, the novelty of the proposed framework is not entirely convincing.

We sincerely appreciate your recognition of the practical importance of preventing shortcut-based predictions and unsupported explanations, as well as the strong results across different forecasting settings. Our central insight is that these failures belong to the same evidence-to-prediction-to-explanation chain. A faithful explanation can describe a shortcut, and a well-grounded prediction can still be explained incorrectly. Solving only one failure therefore does not make the entire chain reliable. Hollowness defines one reliability goal with two necessary checks, which Future Grounding and Behavioral Grounding train through distinct supervision rather than a simple combination of losses.

(1) Why is Hollowness a useful unifying concept?

Consider a model predicting a price increase after positive news. It may be correct because it follows the positive wording, without considering whether the news was already reflected in historical prices. It may then cite a historical trend in its explanation, even though removing that trend barely changes its prediction.

Prediction accuracy cannot reveal the first failure, and a plausible explanation cannot reveal the second. Crucially, a faithful explanation could faithfully describe a shortcut, while a prediction that uses both modalities could still have an unsupported explanation. Fixing either issue alone would leave the other unchecked. Hollowness identifies the shared, testable gap: apparent evidence use does not guarantee actual predictive dependence. Figure 1 makes the gap concrete: 73.6% of baseline-correct controlled pairs fail the full counterfactual diagnostics, while 81.0% of cited-factor interventions lack the fully expected predictive transition.

This unifying view matters because it gives us one reliability goal with two necessary checks: does the prediction depend on the event–history relationship, and does the explanation reflect evidence that supports that prediction? It also makes clear which part of evidence use a method needs to improve. We have emphasized this contribution in Sec. 1, Sec. 3, and Figure 1‾\underline{\text{Sec. 1, Sec. 3, and Figure 1}}.

(2) How does this lead to a different training design?

Future Grounding goes beyond matching input representations. Input-level cross-modal contrastive learning can teach a model whether text and time series fit together. We instead ask: Which event–history combination better fits what actually happened afterward? Keeping the history and observed future fixed, we change the event and compare both combinations with that same future across several horizons. This trains the model to consider what the event means under the particular historical conditions. In the separate objective-level study, Conditional Latent yields a +14.08-point Event CF-PC gain, versus +8.71 for Future Reconstruction and +12.29 for Series Latent (Sec. 4.3; Appendix J.2‾\underline{\text{Sec. 4.3; Appendix J.2}}).

Behavioral Grounding goes beyond selecting important spans. TimeX [1] and TimePNS [2] have established the value of behavior-based explanation tests. Our model jointly generates the prediction, natural-language explanation, and associated temporal and textual evidence from one shared state. We train the evidence associated with that explanation to preserve the prediction when retained, affect it when removed, and leave it stable under irrelevant additions. Thus, the interventions train the link between the generated explanation and the prediction, alongside Future Grounding (Sec. 4.4‾\underline{\text{Sec. 4.4}}).

(3) Do the two requirements produce different improvements?

We compared four variants with the same architecture and 61.74M trainable parameters (Appendix J.1‾\underline{\text{Appendix J.1}}).

VariantEvent CF-PC (%) ↑Sufficiency (%) ↑Comprehensiveness (%) ↑
Task-only29.2266.4123.29
Future-only46.6065.6924.51
Behavioral-only34.4073.4536.89
Full53.7574.5737.93

Findings.

(1) Future-only improves Event CF-PC by 17.38 points.

(2) Behavioral-only improves Sufficiency by 7.04 points and Comprehensiveness by 13.60 points.

(3) Full achieves the highest values on all three measures, improving both behaviors together.

Overall, Hollowness makes reliable evidence use a two-part, testable requirement: predictions should depend on the event–history relationship, and explanations should accurately reflect the evidence supporting those predictions. DualGround-TS trains both requirements within one model, offering a reusable principle for designing and assessing more reliable event-conditioned forecasting systems.

AuthorsOpenReview ↗

Response to Reviewer RjuU [Part 2]

W2: It is unclear how DualGround-TS defines and perturbs these evidence states when explicit ground-truth rationale annotations are unavailable (e.g., on standard datasets like Time-MMD).

We sincerely appreciate your thoughtful comments, especially your recognition that avoiding shortcuts and unsupported explanations is important for practical multimodal forecasting. Our approach does not require hand-labeled causal factors or knowledge of what would actually have happened under an alternative event. Real future trajectories supervise factual event–history combinations; synthetic mismatches serve as contrastive training examples, not causal ground truth. Behavioral Grounding learns from the model's prediction response when selected evidence is retained or removed, without needing human rationale labels. We agree that artificial contrasts can introduce shortcuts, so we test harder event changes and unseen perturbations, while evaluating forecasting separately on Time-MMD.

Our counterfactuals have two different roles. Future Grounding compares factual event–history data against the real observed future, using mismatched events as negative training contrasts. These negatives are not claimed to represent verified alternative causal outcomes. Behavioral Grounding tests the model’s own prediction changes under evidence interventions; it does not require a known causal explanation or hand-annotated rationale for each example. We evaluate more difficult contrasts separately to check whether simple mismatch artifacts account for the gains.

(1) How the method works without evidence annotations. The model selects evidence at two levels: time-step–variable cells from the history and tokens from the event text. It then creates three modified inputs: one retaining the selected evidence, one removing it, and one adding irrelevant context. The model's prediction responses provide the training signal. If human evidence labels exist, they can provide extra supervision, but the behavioral objective itself does not require them. The complete procedure is clarified in Sec. 4.4 and Appendix B.6‾\underline{\text{Sec. 4.4 and Appendix B.6}}.

(2) Real-world forecasting without TimeLitmus evidence labels. Time-MMD follows its official supervised forecasting protocol and does not provide the same controlled rationale labels. We evaluated the architecture-matched variants across its nine real-world domains, with detailed results in Appendix J.8‾\underline{\text{Appendix J.8}} and the main results in Sec. 5.2‾\underline{\text{Sec. 5.2}}.

VariantTime-MMD Average Normalized MSE ↓
Task FT3.8067
Architecture-matched Task-only3.4385
Future-only2.8127
Behavioral-only2.9838
Full2.1842

(3) Independent explanation tests. Two blinded annotators evaluated 100 Finance/Traffic cases. Full achieves 82.7% Human Evidence F1 and 76.0% Human Explanation Support, versus 81.5% and 61.0% for Task-only. The 15.0-point improvement in Explanation Support has a paired 95% CI of [+5, +25]. We also tested 200 previously unseen perturbations, with full results in Appendices J.5–J.6‾\underline{\text{Appendices J.5–J.6}}.

VariantPrediction Invariance (%) ↑Evidence Stability (%) ↑Explanation Consistency (%) ↑
Task-only85.572.073.0
Full92.085.088.0

Findings.

(1) Behavioral Grounding can learn from prediction responses without ground-truth rationale spans.

(2) Full improves forecasting across all nine Time-MMD domains, reducing average normalized MSE by 42.6% relative to Task FT.

(3) Human judgments and unseen-intervention tests also support more reliable use of selected evidence and more consistent generated explanations.

Overall, the method can be trained without manual evidence labels, and its predictive benefits extend to real-world multimodal forecasting data. The independent explanation tests provide further evidence that the model's selected evidence and explanations are meaningfully connected to prediction behavior.

AuthorsOpenReview ↗

Response to Reviewer RjuU [Part 3]

W3: Training a model directly on counterfactual/behavioral objectives and then evaluating it on a counterfactual diagnostic benchmark naturally guarantees large metric gains (+29.6 points) compared to standard task fine-tuning (Task FT), raising concerns about whether the comparison is fair or simply reflects metric-targeted optimization.

We sincerely appreciate your recognition of the strong empirical improvements across TimeLitmus, zero-shot CiK, and supervised Time-MMD. The scientific goal is to learn evidence-dependent prediction and explanation, so training for these behaviors and measuring them with TimeLitmus is appropriate. What would undermine the result is learning quirks of its particular tests, rather than a reusable behavior. We agree that this distinction must be checked independently. We therefore add hard counterfactuals, unseen interventions, blinded human evaluation, and architecture-matched comparisons, alongside the external forecasting benchmarks you highlighted.

(1) Training for a capability and measuring that capability is a valid design. Our goal is to train a model to use the event in the context of the history and to produce explanations supported by evidence. TimeLitmus was designed to measure these behaviors. It is therefore natural—and intentional—that our objectives target the abilities measured by its diagnostics, just as a forecasting model is trained to reduce forecasting error and then evaluated using forecasting metrics.

The important distinction is between learning an ability and learning the quirks of a specific test. Sharing a scientific target with an evaluation metric does not by itself guarantee success on held-out instances or under new interventions. We address that distinction in two ways: source-separated training and evaluation, followed by different intervention types, independent human judgments, and independent forecasting tasks. We also include a same-architecture Task-only control so the gains are not attributed to extra parameters. We have clarified this evaluation logic in Secs. 5.1–5.2‾\underline{\text{Secs. 5.1–5.2}}.

(2) We test the learned behavior in ways the original training tests do not directly reward. We introduced three complementary evaluations:

  • Hard event–history contrasts (E4). Priced-in versus surprise events, historical-regime changes, and same-polarity strength changes check whether the model handles difficult relationships rather than recognizing obvious event mismatches. Full improves Hard-Pair Correctness from 18.89% to 25.18% over matched Task-only (+6.30 points, paired 95% CI: [+4.63, +8.06]).
  • Unseen interventions (E6). We tested 200 instances with event paraphrases, irrelevant metadata, lexical distractors, and alternative evidence masks that were not used in the original training interventions.
  • Independent human evaluation (E5). Two blinded annotators evaluated the selected evidence and explanations. Human-rated Explanation Support increases from 61.0% to 76.0% (paired 95% CI: [+5, +25]). These judgments are not the model's own training scores.
VariantPrediction Invariance (%) ↑Evidence Stability (%) ↑Explanation Consistency (%) ↑
Task-only85.572.073.0
Full92.085.088.0

The complete protocols and results appear in Appendices J.3, J.5, and J.6‾\underline{\text{Appendices J.3, J.5, and J.6}}, with the key findings in Secs. 5.3–5.4‾\underline{\text{Secs. 5.3–5.4}}.

(3) The objectives also improve independent forecasting tasks. CiK and Time-MMD use forecasting-error metrics rather than TimeLitmus's grounding scores. We evaluated all four architecture-matched variants under their respective zero-shot and native supervised protocols.

VariantCiK Average RCRPS ↓Time-MMD Average Normalized MSE ↓
Task FT0.2853.8067
Architecture-matched Task-only0.2833.4385
Future-only0.2672.8127
Behavioral-only0.2742.9838
Full0.2442.1842

Full reduces CiK RCRPS by 14.4% and Time-MMD normalized MSE by 42.6% relative to Task FT; on Time-MMD it improves all nine domains. Full external results are in Appendices J.7–J.8‾\underline{\text{Appendices J.7–J.8}}.

Findings.

(1) The gains extend to harder and previously unseen interventions, where success requires more than passing the original diagnostic cases.

(2) Independent people also judge the explanations to be better supported, providing a check outside the training loss.

(3) The same objectives improve independent forecasting tasks that do not use TimeLitmus's grounding metrics.

Overall, we deliberately train for the evidence-use abilities that TimeLitmus measures. The hard counterfactuals, unseen interventions, independent human judgments, and external forecasting results show that these benefits extend beyond a single diagnostic test. This is the distinction between learning evidence-dependent behavior and merely optimizing a benchmark score.

Reviewer_wjx6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper defines “hollowness” as correct predictions or plausible explanations that do not behaviorally depend on the intended temporal and textual evidence. DualGround-TS builds a shared temporal-language state, trains it to match realized futures better for matched than mismatched events, and applies sufficiency, comprehensiveness, and irrelevant-context constraints to model-selected evidence. It is evaluated on the controlled TimeLitmus benchmark, zero-shot Context is Key, and supervised Time-MMD

Critical strengths
  • The paper identifies an important failure mode. Evaluating whether a model responds to controlled changes in both event and temporal evidence is more informative than reporting predictive accuracy or free-text explanations alone.

  • The evaluation attempts to go beyond one benchmark, covering controlled diagnostics, zero-shot transfer, and supervised multimodal forecasting. The intervention protocol and mismatch construction are described in more detail than is typical for this area.

Critical weaknesses
  • The main comparison does not isolate grounding from model capacity and architecture. Task FT trains 29.10M task-specific parameters using the language backbone, whereas DualGround-TS trains 61.74M and adds a temporal encoder plus temporal-textual pathway. Its advantage can therefore come from access to a substantially stronger multimodal architecture rather than the future-consistent or behavioral objectives. The paper needs architecture-matched ablations for each objective, not only a task-fine-tuning baseline with fewer parameters

  • The proposed losses do not establish the claimed causal event-history dependence or faithful explanation. A mismatched event is selected from a different mechanism category or a fixed derangement and is trained against the factual future; winning this contrast can be achieved by detecting incompatibility or surface event identity, without learning the intended interaction. Likewise, the model produces the evidence selector and is optimized to make its own selected evidence sufficient and its complement consequential. This can make the reported behavioral scores self-fulfilling. There is no independent human evidence annotation, adversarially matched event control, or external faithfulness test that rules this out

Questions for authors
  1. Please compare against an architecture- and parameter-matched model trained only with the task loss, then add the future and behavioral losses separately and jointly. Do the reported gains survive this comparison?

  2. Can the authors construct hard mismatches matched for event mechanism, length, sentiment, entity type, and lexical overlap, and show that the model responds to the correct event-history interaction rather than merely identifying an incompatible event?

  3. Please validate explanation evidence against independently annotated evidence or held-out interventions not used in training or selector construction. How often do selected time cells and tokens agree with external judgments?

5 comments
AuthorsOpenReview ↗

Response to Reviewer wjx6 [Part 1]

Thank you for your constructive and detailed review. Below, we address your concerns with new controlled experiments and independent validation.

W1: The main comparison does not isolate grounding from model capacity and architecture.

We sincerely appreciate your thoughtful review, especially your recognition that testing how predictions respond to controlled temporal and event changes reveals more than accuracy alone, and that our evaluation covers multiple forecasting settings. Our contribution is to train precisely this evidence-use behavior. Architecture size, however, could explain part of the original improvement. We agree that the learning objectives should therefore be tested within a fixed model. We perform a four-way, architecture- and parameter-matched comparison in which only Future Grounding and Behavioral Grounding are switched on or off.

(1) Four models with the same architecture. The original Task FT baseline trains 29.10M parameters, while DualGround-TS trains 61.74M and adds a temporal–textual pathway. To remove this architectural difference, we trained four variants that all use the same DualGround architecture, 61.74M trainable parameters, and matched training conditions:

  • Task-only: task loss only.
  • Future-only: task loss plus Future Grounding.
  • Behavioral-only: task loss plus Behavioral Grounding.
  • Full: task loss plus both grounding objectives.

We have added the complete comparison to Appendix J.1 of the revised manuscript‾\underline{\text{Appendix J.1 of the revised manuscript}}, with the main results in Sec. 5.3‾\underline{\text{Sec. 5.3}}. The new controlled study is labeled separately from the original Task FT comparison, with its evaluation and aggregation protocol specified in the appendix.

Prediction results (%)

VariantNatural Acc. ↑Series CF-PC ↑Event CF-PC ↑HPC ↑
Task-only46.3812.8229.2219.72
Future-only47.9113.9246.6020.93
Behavioral-only46.8411.4734.4021.23
Full48.5012.9553.7521.15

Explanation results (%)

VariantEvidence F1 ↑Sufficiency ↑Comprehensiveness ↑Invariance ↑
Task-only89.3566.4123.2999.41
Future-only89.3465.6924.5198.68
Behavioral-only89.0873.4536.8999.75
Full89.5574.5737.9399.51

Findings.

(1) The improvements remain when model capacity is fixed. Full improves Natural Accuracy by 2.12 points, Series CF-PC by 0.13, Event CF-PC by 24.53, and HPC by 1.43 over Task-only.

(2) The two objectives help in different ways. Future-only improves Event CF-PC by 17.38 points. Behavioral-only improves Sufficiency by 7.04 points and Comprehensiveness by 13.60 points.

(3) The joint model benefits from both. Full has the highest Event CF-PC, Sufficiency, and Comprehensiveness among the matched variants; Future-only leads Series CF-PC, and Behavioral-only leads HPC and Invariance. Evidence F1 remains near 89%.

Overall, the additional experiment isolates what our grounding objectives contribute under exactly the same architecture. The observed improvements come from the learning objectives beyond the extra capacity of the DualGround model.

AuthorsOpenReview ↗

Response to Reviewer wjx6 [Part 2]

W2: The proposed losses do not establish the claimed causal event-history dependence or faithful explanation. ... This can make the reported behavioral scores self-fulfilling.

We sincerely appreciate your detailed comments, especially your recognition of the importance of measuring whether predictions truly use temporal and textual evidence, rather than relying on accuracy or plausible explanations alone. Our claim is behavioral event–history dependence, not identification of causal effects. Future Grounding uses observed future outcomes to supervise alternative event–history combinations; it does not treat a synthetic mismatch as a known causal outcome. Behavioral Grounding tests whether explanation-associated evidence affects the model's predictions. We agree that these claims need checks beyond the training signals, so we add hard event contrasts, equal-budget evidence controls, blinded human judgments, and unseen interventions.

(1) Hard event–history comparisons. We tested priced-in versus surprise events, changes in historical regime, and same-polarity event-strength changes. We also included an Additive Fusion baseline to examine whether independently combining event and time-series signals explains the results. The full experiments are in Appendix J.3‾\underline{\text{Appendix J.3}}.

VariantTarget Accuracy (%) ↑Hard-Pair Correctness (%) ↑
Task-only57.2218.89
Future-only58.2420.09
Behavioral-only58.6124.08
Additive Fusion57.9620.56
Full59.7225.18

(2) Equal-budget evidence controls. We compared learned evidence with random selection and temporal-magnitude selection using identical evidence budgets. Complete results appear in Appendix J.4‾\underline{\text{Appendix J.4}}.

Evidence SelectionSufficiency (%) ↑Comprehensiveness (%) ↑Retain–Remove Separation (pp) ↑
Random59.6416.70−16.32
Temporal Magnitude60.0017.29−17.38
Learned Evidence73.5728.14+8.95

(3) Independent evaluation. We also checked the selected evidence against blinded human judgments, tested whether explanations accurately describe their cited evidence, and evaluated new perturbations that were not used in training. Full protocols and results are in Appendices J.5–J.6‾\underline{\text{Appendices J.5–J.6}}, with the main findings in Sec. 5.4‾\underline{\text{Sec. 5.4}}.

Findings.

(1) The model handles difficult event–history relationships better. Full improves Hard-Pair Correctness by 6.30 points over Task-only (95% CI: [+4.63, +8.06]) and 4.62 points over Additive Fusion (95% CI: [+2.50, +6.70]).

(2) Its selected evidence is more useful than simple selection rules. Compared with random selection, retained-evidence agreement increases by 13.94 points (95% CI: [+11.15, +16.72]), and the response to evidence removal improves by 11.44 points (95% CI: [+8.89, +14.06]).

(3) Independent checks also support the explanations. Human-rated Explanation Support rises from 61.0% to 76.0%, while unseen interventions show better stability of both selected evidence and generated explanations.

Overall, the additional tests check both links in reliable evidence use: whether the prediction depends on the intended event–history information, and whether the explanation reflects evidence that actually supports the prediction. The gains are supported by hard counterfactuals, equal-budget controls, independent judgments, and new interventions.

AuthorsOpenReview ↗

Response to Reviewer wjx6 [Part 3]

Q1: Please compare against an architecture- and parameter-matched model trained only with the task loss, then add the future and behavioral losses separately and jointly. Do the reported gains survive this comparison?

Yes. We agree that this is the clearest way to separate learning-objective gains from architecture capacity. Our two objectives are intended to change what the same model learns, so we keep the architecture and all trainable parameters fixed and compare Task-only, Future-only, Behavioral-only, and Full under matched conditions. We then measure both counterfactual prediction behavior and support for explanation-associated evidence.

The four variants all have 61.74M trainable parameters. Full results appear in Appendix J.1‾\underline{\text{Appendix J.1}}, with the principal comparison in Sec. 5.3‾\underline{\text{Sec. 5.3}}.

MetricTask-onlyFullImprovement
Natural Accuracy (%) ↑46.3848.50+2.12 pp
Series CF-PC (%) ↑12.8212.95+0.13 pp
Event CF-PC (%) ↑29.2253.75+24.53 pp
HPC (%) ↑19.7221.15+1.43 pp
Sufficiency (%) ↑66.4174.57+8.16 pp
Comprehensiveness (%) ↑23.2937.93+14.64 pp

Findings.

(1) Full improves all six metrics under identical architectural conditions.

(2) Future-only raises Event CF-PC from 29.22% to 46.60%, while Behavioral-only raises Sufficiency from 66.41% to 73.45% and Comprehensiveness from 23.29% to 36.89%.

(3) Combining both gives the highest Event CF-PC and behavioral evidence-support scores among the matched variants.

Overall, the gains survive the requested architecture- and parameter-matched comparison, directly supporting the independent contributions of the proposed objectives.

Q2: Can the authors construct hard mismatches matched for event mechanism, length, sentiment, entity type, and lexical overlap, and show that the model responds to the correct event-history interaction rather than merely identifying an incompatible event?

Thank you for specifying these controls. The key ability we test is whether the same event can carry a different predictive implication under different historical conditions, rather than whether the model can spot an obviously unrelated text. We agree that easy negative examples can overstate this ability. We therefore evaluate priced-in versus surprise events, historical-regime changes, and same-polarity strength changes, and include an additive-fusion baseline. The reported comparisons support those three contrast families; their exact matching attributes are described in the experimental protocol.

(1) Hard counterfactual construction. The new evaluation includes: priced-in versus surprise events, where expectations in the history change an event's meaning; historical-regime changes, where the preceding market state matters; and same-polarity strength changes, where simple positive/negative event sentiment is not enough to distinguish the pair. We report the event attributes held fixed and changed in each family in Appendix J.3‾\underline{\text{Appendix J.3}}.

We have included the construction procedure, the attributes actually controlled by each family, and the corresponding validity checks in Appendix J.3‾\underline{\text{Appendix J.3}}. The main results are in Sec. 5.3‾\underline{\text{Sec. 5.3}}.

ModelHard-Pair Correctness (%) ↑
Task-only18.89
Future-only20.09
Behavioral-only24.08
Additive Fusion20.56
Full25.18

Findings.

(1) Full improves the two-outcome test. Pair Correctness increases by 6.30 points over Task-only (95% CI: [+4.63, +8.06]).

(2) Full also improves over Additive Fusion. The gain is 4.62 points (95% CI: [+2.50, +6.70]).

(3) All three families benefit. The gains over Task-only are +7.22 points for priced-in versus surprise events and +5.83 points for each of the other two families.

Overall, the harder contrasts provide new evidence that DualGround-TS responds to the historical context of an event, not just broad event identity or simple additive signals. The appendix reports the specific matching controls used for each family, so the evidence can be assessed against the properties you requested.

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

This paper identifies 'Hollowness", an important problem in event-conditioned time-series prediction where accurate predictions and plausible explanations do not necessarily reflect genuine dependence on relevant temporal and textual evidence.

To address this issue, the authors propose DualGround-TS, which incorporates two complementary grounding mechanisms: (1) future-consistent cross-modal grounding, which encourages meaningful interactions between historical time series and textual events using realized future trajectories, and (2) behavioral explanation grounding, which aligns explanation-associated evidence with actual prediction behavior through controlled interventions.

The authors validate their approach on three benchmarks: TimeLitmus, Context is Key, and Time-MMD. Experimental results demonstrate improvements in prediction accuracy, cross-modal grounding, explanation faithfulness, and forecasting performance compared to task-specific fine-tuning.

Critical strengths
  1. Meaningful and Important Problem Formulation. One of the key strengths of this paper is that it identifies and addresses an important yet underexplored problem in event-conditioned time-series prediction. The authors introduce the concept of Hollowness, highlighting that accurate predictions and plausible explanations do not necessarily imply genuine reliance on relevant temporal and textual evidence. By explicitly addressing both prediction and explanation hollowness, this work provides a meaningful problem formulation and proposes a framework to improve the reliability and faithfulness of multimodal time-series prediction.
Critical weaknesses
  1. Limited Evidence for Genuine Cross-Modal Reasoning. Although the proposed future-consistent grounding encourages the model to distinguish matched from mismatched event–history pairs, it remains unclear whether this objective necessarily promotes meaningful cross-modal reasoning rather than simpler event-dependent associations. Additional analyses would help establish whether the learned representations truly capture interactions between temporal dynamics and textual events.

  2. Limited Experimental Comparisons. Although DualGround-TS consistently outperforms its Task FT baseline, the main comparisons do not fully isolate the contributions of the proposed grounding objectives from those of the additional architectural components. Moreover, the method does not consistently outperform stronger baselines on the CiK benchmark. More comprehensive comparisons under matched architectural settings would strengthen the empirical validation.

Questions for authors
  1. Cross-modal Reasoning How does the proposed grounding objective ensure genuine cross-modal reasoning rather than reliance on simpler event-dependent associations?

  2. Explanation Faithfulness How does the proposed behavioral grounding ensure that generated explanations faithfully reflect the model's reasoning, beyond identifying evidence that influences predictions?

6 comments
AuthorsOpenReview ↗

Response to Reviewer dDRE [Part 1]

Thank you for your constructive and detailed review. Below, we respond to each of your comments and summarize the additional experiments.

W1: Although the proposed future-consistent grounding encourages the model to distinguish matched from mismatched event–history pairs, it remains unclear whether this objective necessarily promotes meaningful cross-modal reasoning rather than simpler event-dependent associations.

We sincerely appreciate your insightful comments, especially your recognition of the value of testing whether a prediction truly depends on the relationship between an event and the preceding time series, rather than prediction correctness alone. Our central design uses the observed future to supervise conditional event comparisons: under the same history, the model learns which event better fits what actually happened afterward. We agree that easy matched-versus-mismatched pairs alone cannot establish this ability. We therefore introduce harder event–history contrasts and an additive-fusion control to test whether the gains depend on learning the interaction.

(1) Learning from what actually happened. Future Grounding uses the observed future as a training signal. We keep the historical series and its actual future fixed, replace the event, and compare which event–history combination better matches that future. The model learns to favor the actual combination. This gives the model a reason to learn how an event's effect depends on the preceding history, instead of relying only on whether the event sounds positive or negative.

(2) Hard counterfactual experiments. We constructed three additional Finance contrast families: priced-in versus surprise events, changes in historical market regime, and events with the same polarity but different strengths. They test whether the model handles situations in which broad event sentiment or an obvious mismatch is less informative. We also added an Additive Fusion baseline, which combines temporal and textual information without the same learned interaction.

We have summarized the new findings in Sec. 5.3 of the revised manuscript‾\underline{\text{Sec. 5.3 of the revised manuscript}}, with the full construction and results in Appendix J.3‾\underline{\text{Appendix J.3}}. For convenience, the main results are shown below.

ModelTarget Accuracy (%) ↑Hard-Pair Correctness (%) ↑
Task-only57.2218.89
Future-only58.2420.09
Behavioral-only58.6124.08
Additive Fusion57.9620.56
Full59.7225.18

Findings.

(1) Full performs better on challenging event–history pairs. Hard-Pair Correctness increases from 18.89% to 25.18%, an improvement of 6.30 percentage points over Task-only (paired 95% CI: [+4.63, +8.06]).

(2) Learning the interaction is more effective than simple fusion. Full exceeds Additive Fusion by 4.62 points (95% CI: [+2.50, +6.70]).

(3) The advantage is consistent across the three types of contrasts. Compared with Task-only, the gains are +7.22, +5.83, and +5.83 points, respectively.

Overall, these experiments provide direct support for our central goal: learning how an event's predictive meaning changes with the historical conditions, rather than relying only on broad event cues. The improvements hold even on harder counterfactual cases and against an additive combination of the two modalities.

AuthorsOpenReview ↗

Response to Reviewer dDRE [Part 2]

W2: The main comparisons do not fully isolate the contributions of the proposed grounding objectives from those of the additional architectural components. Moreover, the method does not consistently outperform stronger baselines on the CiK benchmark.

We sincerely appreciate your thoughtful comments, especially your recognition that DualGround-TS improves on task fine-tuning while addressing a meaningful problem in reliable event-conditioned prediction. These gains should be attributed to the grounding objectives themselves, not simply to the larger temporal–textual pathway. We agree that the original comparison does not fully isolate the two factors. We therefore compare four models with identical architectures and trainable parameter counts, changing only the learning objectives. For CiK, our aim is reliable evidence use alongside competitive forecasting, rather than the highest raw score on every forecasting benchmark.

(1) Architecture-matched comparison. The original Task FT baseline trains 29.10M parameters, whereas DualGround-TS trains 61.74M and includes an additional temporal–textual pathway. To isolate the learning objectives, we trained four variants with the same architecture, 61.74M trainable parameters, and matched training conditions. Only the Future and Behavioral Grounding objectives differ.

The complete experiment is included in Appendix J.1‾\underline{\text{Appendix J.1}}, with its main findings in Sec. 5.3‾\underline{\text{Sec. 5.3}}.

VariantNatural Acc. (%) ↑Series CF-PC (%) ↑Event CF-PC (%) ↑HPC (%) ↑
Task-only46.3812.8229.2219.72
Future-only47.9113.9246.6020.93
Behavioral-only46.8411.4734.4021.23
Full48.5012.9553.7521.15

(2) CiK and our main research goal. Our goal is not simply to obtain the lowest forecasting error on every benchmark. We aim to make predictions that remain useful and competitive while improving whether they depend on the relevant historical and textual evidence and whether their explanations are supported. TimeLitmus and the intervention experiments directly test that reliability goal; CiK tests whether the learned model also forecasts well on an independent benchmark.

We evaluated the four matched variants on CiK without CiK-specific training or model selection, and clarified the comparison with stronger external systems in Sec. 5.2‾\underline{\text{Sec. 5.2}}. The category-wise CiK results are reported in Table 3 and Appendix D.2‾\underline{\text{Table 3 and Appendix D.2}}, while the architecture-matched comparison is provided in Appendix J.7‾\underline{\text{Appendix J.7}}.

ModelCiK Average RCRPS ↓
Llama-3.1-405B (DP)0.159
Llama-3-70B (LLMP)0.236
Qwen3.5-9B Task FT0.285
Architecture-matched Task-only0.283
Future-only0.267
Behavioral-only0.274
Full0.244

Findings.

(1) Grounding still helps when capacity is fixed. Relative to matched Task-only, Full improves Natural Accuracy by 2.12 points, Series CF-PC by 0.13, Event CF-PC by 24.53, and HPC by 1.43.

(2) The benefits also appear outside our diagnostic benchmark. On CiK, Full reduces average RCRPS from 0.283 to 0.244 against the architecture-matched baseline, and from 0.285 to 0.244 against Task FT.

(3) The CiK results remain competitive without making forecasting accuracy our only target. In comparison with stronger external systems, our controlled results show better forecasting from the same Qwen3.5-9B architecture, alongside the grounding improvements directly measured on TimeLitmus.

Overall, the new results establish that the grounding objectives provide benefits beyond architectural capacity. They also support our intended balance: stronger evidence-based prediction and explanation, with useful forecasting performance on an independent benchmark.

AuthorsOpenReview ↗

Response to Reviewer dDRE [Part 3]

Q1: How does the proposed grounding objective ensure genuine cross-modal reasoning rather than reliance on simpler event-dependent associations?

Thank you for raising this question. Future Grounding is designed to learn how an event's predictive meaning changes with the preceding history, not merely whether event wording is positive or negative. We compare alternative events under the same history against the same observed future. We agree that this design needs behavioral evidence beyond the training loss, so we also test alternative future-learning objectives and harder event–history contrasts.

(1) How Future Grounding works. Imagine keeping the historical time series fixed but replacing the accompanying event. We compare both versions against the same observed future trajectory. The model is trained so that the version with the actual event better matches the outcome that followed. This gives it a training signal about whether that event makes sense under the particular historical conditions, rather than only whether its wording looks positive or negative.

(2) Comparing different ways of learning from the future. We tested alternatives that reconstruct the future or learn a future-related representation from the series alone, using the same architecture. We have clarified the procedure in Sec. 4.3‾\underline{\text{Sec. 4.3}} and report the relative objective-level improvements in Appendix F.1 (Table 23)‾\underline{\text{Appendix F.1 (Table 23)}}, discussed in Appendix J.2‾\underline{\text{Appendix J.2}}, alongside the original analyses in Figure 4‾\underline{\text{Figure 4}}.

ObjectiveΔ Natural Acc. (pp) ↑Δ Series CF-PC (pp) ↑Δ Event CF-PC (pp) ↑Δ HPC (pp) ↑
Future Reconstruction+0.78+3.48+8.71+0.16
Series Latent+3.37+5.28+12.29+0.55
Conditional Latent (Ours)+3.99+6.55+14.08+2.14

Findings.

(1) Using future information improves performance, and how it is used makes a difference.

(2) Conditional Latent achieves the largest improvements across all four metrics, including +14.08 points in Event CF-PC over Task-only, compared with +8.71 for Future Reconstruction and +12.29 for Series Latent.

(3) The separate hard-counterfactual results in W1 further show stronger performance on cases where simple event cues are less useful.

Overall, Future Grounding gives the model a concrete way to learn how an event and the preceding history jointly relate to the outcome. Both the objective comparison and the hard counterfactual tests support its value for learning this relationship.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to all reviewers & Summary of revisions

We sincerely thank all the reviewers for their insightful comments and valuable suggestions, which have greatly helped us improve our paper.

This paper introduces DualGround-TS, a framework for addressing Hollowness in event-conditioned time-series prediction. Hollowness reveals a shared failure pattern behind shortcut-dependent predictions and behaviorally unsupported explanations: apparent evidence use does not guarantee behaviorally valid predictive dependence. Addressing either failure alone is insufficient for reliable prediction and explanation. Based on this insight, DualGround-TS combines future-consistent cross-modal grounding with behavioral explanation grounding to address both. We evaluate DualGround-TS on TimeLitmus (controlled grounding), CiK (zero-shot forecasting), and Time-MMD (supervised forecasting), presenting four key findings that we hope will provide a systematic foundation for developing more reliable evidence-grounded time-series models.

The reviewers recognized the importance of our work, describing Hollowness as an "important yet underexplored problem" and a "meaningful problem formulation". They highlighted the "practical value" of evaluating genuine event–history dependence and explanation faithfulness beyond prediction accuracy, as well as the detailed intervention protocols. Our experiments were also recognized for demonstrating "strong empirical gains across diverse regimes", spanning controlled diagnostics, zero-shot forecasting, and supervised multimodal forecasting.

AuthorsOpenReview ↗

Response to all reviewers & Summary of revisions – Continued

The reviewers also provided many insightful comments. We have made every effort to address all the concerns. Below is a summary of the major revisions:

  • Clarifying the scientific motivation, novelty, and method presentation (Reviewer RjuU & Reviewer qiPC): We explain why both failures must be addressed together: solving either alone cannot ensure reliable prediction and explanation. We expand comparisons with TESS, TimeX, and TimePNS, redesign Figure 1, and clarify task-conditioned evidence readouts and training-only mechanisms in Figure 2.

    Revised → Secs. 1–4, Figures 1–2, and Appendix A‾\underline{\text{Secs. 1–4, Figures 1–2, and Appendix A}}.

  • Architecture-matched grounding ablations (Reviewer dDRE, Reviewer wjx6, Reviewer RjuU & Reviewer qiPC): We compare Task-only, Future-only, Behavioral-only, and Full using the same Qwen3.5-9B architecture (61.74M trainable parameters) and matched training conditions. Relative to Task-only, Full improves Event CF-PC by 24.53 points, Sufficiency by 8.16 points, and Comprehensiveness by 14.64 points.

    Newly added → Appendix J.1: Architecture-Matched Grounding Ablation (Tables 31–32)‾\underline{\text{Appendix J.1: Architecture-Matched Grounding Ablation (Tables 31–32)}}. Revised → Sec. 5.3‾\underline{\text{Sec. 5.3}}.

  • Harder event–history comparisons and future-objective controls (Reviewer dDRE, Reviewer wjx6 & Reviewer RjuU): We compare conditional future grounding with Future Reconstruction and Series Latent, and test three harder Finance contrasts (priced-in/surprise, historical-regime, and same-polarity event-strength changes) alongside Additive Fusion. Full achieves 25.18% hard-pair correctness versus 18.89% (Task-only) and 20.56% (Additive Fusion), with paired CIs supporting both gains.

    Newly added → Appendices J.2–J.3: Future-Grounding Objectives and Harder Event–History Comparisons‾\underline{\text{Appendices J.2–J.3: Future-Grounding Objectives and Harder Event–History Comparisons}}. Revised → Sec. 5.3‾\underline{\text{Sec. 5.3}}.

  • Independent validation of evidence and explanations (Reviewer dDRE, Reviewer wjx6 & Reviewer RjuU): We add equal-budget selection controls, blinded human judgments (100 examples), 200 claim-level explanation checks, and 200 unseen-intervention cases. Human Explanation Support rises from 61.0% to 76.0% (paired 95% CI: [+5, +25]), while held-out Explanation Consistency rises from 73.0% to 88.0%.

    Newly added → Appendices J.4–J.6: Evidence Selection, Independent Explanation Validation, and Unseen Interventions‾\underline{\text{Appendices J.4–J.6: Evidence Selection, Independent Explanation Validation, and Unseen Interventions}}. Revised → Sec. 5.4‾\underline{\text{Sec. 5.4}}.

  • Transfer to independent forecasting benchmarks (Reviewer dDRE, Reviewer RjuU & Reviewer qiPC): We add architecture-matched comparisons on zero-shot CiK and supervised Time-MMD. Compared with matched Task-only, Full reduces CiK RCRPS from 0.283 to 0.244 and Time-MMD normalized MSE from 3.4385 to 2.1842; it also outperforms the original Task FT in all nine Time-MMD domains.

    Newly added → Appendices J.7–J.8: Architecture-Matched CiK and Time-MMD Transfer‾\underline{\text{Appendices J.7–J.8: Architecture-Matched CiK and Time-MMD Transfer}}. Revised → Sec. 5.2‾\underline{\text{Sec. 5.2}}.

  • Grounding across backbone sizes and families (Reviewer RjuU & Reviewer qiPC): We extend evaluation to Qwen3.5-4B and Llama-3.1-8B, with three-seed comparisons and paired 95% confidence intervals. Full improves Sufficiency and Comprehensiveness in Finance and Traffic on both backbones, with all eight corresponding intervals excluding zero. Complete five-metric results are reported.

    Newly added → Appendix J.10: Grounding Across Backbone Sizes and Families (Tables 42–43)‾\underline{\text{Appendix J.10: Grounding Across Backbone Sizes and Families (Tables 42–43)}}. Revised → Sec. 5.3‾\underline{\text{Sec. 5.3}}.

  • Training and inference efficiency (Reviewer RjuU): We profile training time, peak GPU memory, and inference latency, separating training-only grounding from inference. Full adds 41.42% training time per update over matched Task-only, while inference latency differs by less than 1% in the tested settings.

    Newly added → Appendix J.9: Training and Inference Cost (Tables 40–41)‾\underline{\text{Appendix J.9: Training and Inference Cost (Tables 40–41)}}. Revised → Sec. 4 and Figure 2‾\underline{\text{Sec. 4 and Figure 2}}.

After a full 7 days of experiments and revisions, we have added over 150 new experimental results across 13 new tables to address the reviewers’ concerns. The revisions are highlighted in orange‾\underline{\text{orange}}.

The reviewers’ valuable suggestions have been extremely helpful in improving our paper. We are happy to answer any further questions.

We look forward to your feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.