acceptodds
Under review as a conference paper at ICLR 2027

Updating Beliefs over Eventual Outcomes as Evidence Arrives

10 views

Abstract

Real-world decisions often require updating beliefs about the same eventual outcome as evidence arrives. Learning these belief paths is challenging because the true outcome probabilities at intermediate stages are not directly observable, even after outcomes are known. We introduce **BELIEF**, a generative framework for learning belief paths from event streams. Rather than directly predicting outcomes, BELIEF learns how incoming evidence is distributed under each candidate outcome, given the full preceding history. Observed evidence then updates the relative support for these candidate outcomes through accumulated log likelihoods and Bayesian inference, yielding entire belief paths without direct supervision of intermediate beliefs. It supports binary, multiclass, and continuous outcomes. We provide quantitative guarantees for belief-path recovery and prove exact recovery under correct prior and evidence models. Experiments show accurate belief-path recovery on synthetic evidence streams and the closest alignment with reference probability paths on real data, while preserving downstream task performance. Each belief update decomposes into per-event evidence contributions, providing a basis for risk monitoring and timely decisions.

Primary area:
learning on time series and dynamical systems
Keywords:
Belief-Path Learning, Event Streams, Bayesian Inference, Probabilistic Prediction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Tu4NOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes BELIEF, which uses the observed distribution under the final result condition and Bayesian update to learn the changing path of the probability of the final result in the event stream. Experiments show that it can more accurately recover the probability path than the baseline on synthetic data and some real tasks, and maintains competitive final prediction performance. However, it does not fully outperform other methods in final prediction, and the reference path in real data is not equivalent to the true posterior.

Critical strengths

The method performed well and achieved competitive results in multiple datasets. Although the method proposed in the paper is quite simple, its performance on multiple datasets is excellent, which proves the effectiveness of the method.

Critical weaknesses

I don't see much practical value in this topic. For events that change fast in the duration, the predictions fall to overfit, since the predictions can not be made by incomplete datas. For example events sharing the same prefix that may have different results in the end.

The method relies on a known set of outcomes. The paper does not discuss how to handle new outcomes or candidate sets that change at test time.

The efficiency evaluation is incomplete. The reported timing results cover only binary synthetic data. It remains unclear how the cost grows with the number of candidates or compares with direct predictors.

The method can not be used in open world.

Questions for authors

For cases that sharing same prefix but leading to different results, did they share the same label during training? How to handle this conflict? In other words, the gt labels are made because of the full context, why should an incomplete prefix predict the same labels? This is my main concern of this paper.

How do inference time and memory usage change as the number of candidate outcomes increases, compared with PREFIX, TERMINAL, and other direct predictors?

10 comments
AuthorsOpenReview ↗

Response to Reviewer Tu4N (1/2)

Thank you very much for your careful review and for your many thoughtful and insightful comments. Your main question on shared prefixes concerns the core design of BELIEF, so we first clarify this point and then respond to each comment.

Q1 — Shared prefixes and the training label

For cases that sharing same prefix but leading to different results, did they share the
same label during training? How to handle this conflict? In other words, the gt labels are
made because of the full context, why should an incomplete prefix predict the same labels?
This is my main concern of this paper.

We would like to clarify that BELIEF does not train an incomplete prefix to predict the final label. Instead, it learns the distribution of observations conditioned on each candidate outcome:

MethodModeled distributionRole of the outcome yy
PREFIXp(y∣Hj)p(y\mid H_j)prediction target at every prefix
BELIEFp(xj∣Hj−1,τj,mj,y)p(x_j\mid H_{j-1},\tau_j,m_j,y)conditioning input

Training. At every step, BELIEF takes the observation history, the time and variable of the next observation, and the realized outcome as inputs, and predicts the distribution of the next observed value. It is trained by maximizing the likelihood of the value actually observed (Eq. 3).

Inference. The outcome is unknown. For each new observation, BELIEF evaluates the learned distribution at the observed value under every candidate outcome and updates the posterior over outcomes by Bayes' rule (Eq. 4), rather than predicting the final label directly.

Hence, for two trajectories that share a prefix but end in A and B, one contributes to the observation distribution under A and the other to that under B. These are distributions under different conditions, so the two trajectories never provide conflicting targets. If the prefix observations are equally likely under A and B, their likelihood ratio is close to one and the posterior stays near the prior. Identical prefixes receive identical beliefs until later evidence separates them (Fig. 3C).

W1 — Practical value

I don't see much practical value in this topic.

In many applications, the probability of the eventual outcome must be updated in real time while the process is only partially observed, and this evolving probability is the belief path that BELIEF learns. In intensive care, for example, a patient's risk of in-hospital mortality is updated as vital signs and laboratory results arrive, supporting early warning (P12, MIMIC-III). In futures markets, the probability that a price reaches a given level at expiry is revised as supply and demand data are released, guiding hedging (Steel-RB).

Such evolving probabilities are in fact common in practice, although they are usually produced by mechanistic models or empirical formulas, such as precipitation probability forecasts, option-implied probabilities and outcome probabilities in sports. Our real-data reference paths (Weather, Steel-RB, Soccer) are of exactly this kind. These applications need an accurate probability at every moment, not only at the end.

W2 — Overfitting on incomplete prefixes

For events that change fast in the duration, the predictions fall to overfit, since the
predictions can not be made by incomplete datas. For example events sharing the same prefix
that may have different results in the end.

Training incomplete prefixes directly against the realized final label may lead to unreliable intermediate predictions when the available evidence is insufficient to distinguish outcomes. As explained in Q1, BELIEF never uses the final label as a target for an incomplete prefix: an uninformative observation has a likelihood ratio near one and leaves the belief almost unchanged, so the belief moves only as distinguishing evidence arrives.

We also evaluated BELIEF on fast-changing events, using all 1,306 qualifying binary synthetic test trajectories from the original generator whose exact posterior reverses rapidly (two opposite moves, each within six hours, Appendix D.6). BELIEF tracks these reversals much more closely than all baselines:

Reversal-window KL↓BELIEFPREFIXTERMINAL
Linear0.01010.18490.2561
Nonlinear0.01850.04800.1989
AuthorsOpenReview ↗

Response to Reviewer Tu4N (2/2)

W3 / W5 — New outcomes and open-world use

The method relies on a known set of outcomes. The paper does not discuss how to handle new
outcomes or candidate sets that change at test time.
The method can not be used in open world.

In general classification tasks, the candidate outcomes are defined by the task. In BELIEF, once labeled trajectories of a new outcome are available, the outcome can be added to the model. Since each candidate enters the evidence model through an outcome embedding, BELIEF only needs to train the new embedding and the shared weight branch while keeping the encoder fixed (about 6% of the parameters), whereas PREFIX and TERMINAL update all parameters. We withheld one class in the synthetic four-class settings and added it later (Appendix D.7, Table 10):

LinearNonlinear
Path KL↓Old-class F1 change↑Path KL↓Old-class F1 change↑
BELIEF0.0515−0.03610.0394−0.0256
PREFIX0.4402−0.11590.1447−0.0792
TERMINAL0.5870−0.11540.5037−0.0804

After the new outcome is added, BELIEF has much lower path KL and less forgetting of old classes. For outcomes not yet labeled, BELIEF admits a natural extension: its posterior-predictive likelihood (Appendices B.4 and D.7) can flag trajectories whose evidence is poorly explained by all known outcomes, and the corresponding outcomes can then be added as new candidates once labeled, so that the candidate set grows as new outcomes appear.

W4 / Q2 — Cost versus number of candidates

The efficiency evaluation is incomplete. The reported timing results cover only binary
synthetic data. It remains unclear how the cost grows with the number of candidates or
compares with direct predictors.
How do inference time and memory usage change as the number of candidate outcomes
increases, compared with PREFIX, TERMINAL, and other direct predictors?

We have added a comprehensive efficiency evaluation (Section 4.7, Fig. 4 and Appendix D.5), measuring complete-path inference across different numbers of candidate outcomes and events against PREFIX and TERMINAL on the same backbone.

The cost does not grow linearly with the number of candidates, because all candidates are evaluated in one batched operation. The history encoding and the mixture components are shared by all candidates and computed once per event. Only the mixture weights depend on the candidate, so stacking all candidate embeddings into a matrix yields every candidate's likelihood in a single matrix operation, including on the grid of values used for continuous outcomes. With 200 events on one RTX 3090:

CandidatesLatency (ms)↓Peak memory (MiB)↓
BELIEFPREFIXBELIEFPREFIX
20.8100.56916.4215.27
160.7830.57216.4415.28
640.7840.58217.6515.30

Compared with direct predictors, the extra cost is small. TERMINAL has essentially the same cost as PREFIX.

Additional response

In addition, we respond to the following remark in your summary.

However, it does not fully outperform other methods in final prediction, and the reference
path in real data is not equivalent to the true posterior.

Final prediction. Our main task is to learn belief paths, that is, how the probability of the eventual outcome should evolve as evidence arrives. The downstream experiments are therefore designed to verify that this objective does not come at the cost of final prediction. BELIEF achieves the best results or stays close to the best method on these tasks (Table 3), confirming that learning belief paths does not noticeably affect downstream performance.

Reference paths. We acknowledge that the true posterior is rarely observable in real settings, and this is what motivates the task. BELIEF provides a theoretically grounded way to recover it from evidence and outcomes alone (Section 3.5). On real data, we therefore compare with reference paths, which represent the beliefs of professional forecasters and markets about the same outcome. BELIEF is not trained on them, so its close agreement with them shows that it learns comparable beliefs directly from raw evidence.

In response to W4/Q2, could the authors please clarify the following observations:

  1. As the number of candidates increases, the latency of BELIEF decreases, whereas that of PREFIX increases. This appears counterintuitive, since conventional prediction methods should be largely insensitive to the number of candidates, while BELIEF is expected to involve substantially more computation. Could the authors explain the underlying reasons for this behavior?

  2. Why does the peak memory consumption of PREFIX increase as the number of candidates grows?

  3. Why was TERMINAL not included in the comparison? If possible, could the authors provide the corresponding results?

Reviewer_im3UOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BELIEF, a generative framework for learning belief paths from event streams. It learns observation distributions conditioned on history and candidate outcomes, then constructs posterior paths through Bayesian updating. The authors provide theoretical guarantees for belief-path recovery, and evaluate BELIEF on synthetic data, real-world reference probability paths, and downstream prediction tasks.

Critical strengths

1.BELIEF connects outcome-conditioned evidence modeling with sequential Bayesian updating, providing explicit per-event likelihood-ratio contributions. This is consequential because it offers a unified mechanism for both constructing evolving predictions and interpreting how incoming observations revise them.

2.The synthetic experiments use analytically known posterior paths and matched PREFIX and TERMINAL controls across six encoders. The consistent improvements provide evidence that the benefits extend across architectures. Evaluating entire paths also reveals errors that endpoint prediction metrics alone would miss, directly supporting the paper’s central objective.

Critical weaknesses

1.The paper cites temporal-consistency training for incremental sequence classification (Maystre et al., 2025) but does not include it as a baseline. Although the matched PREFIX and TERMINAL comparisons are useful, they do not establish whether BELIEF improves upon training objectives specifically designed for evolving predictions. This omission is consequential because it leaves the empirical advantage of the proposed evidence-learning formulation over a closely related alternative unresolved.

Questions for authors

1.When the observation mechanism violates the conditional independence assumption in Equation (7), how are BELIEF’s belief-path recovery error and calibration affected? Does it become overconfident or fail to fully exploit information conveyed by the observation process?

2.If distributional error is computed directly from the predicted posterior on the outcome grid, rather than from a moment-matched Gaussian approximation, does BELIEF retain its advantage on continuous-outcome tasks?

3 comments
AuthorsOpenReview ↗

Response to Reviewer im3U

Thank you very much for your recognition of our work and for your constructive comments, in particular the suggestion to consider informative observation mechanisms. We respond to each comment below.

W1 — Temporal-consistency baseline

The paper cites temporal-consistency training for incremental sequence classification
(Maystre et al., 2025) but does not include it as a baseline. Although the matched PREFIX
and TERMINAL comparisons are useful, they do not establish whether BELIEF improves upon
training objectives specifically designed for evolving predictions. This omission is
consequential because it leaves the empirical advantage of the proposed evidence-learning
formulation over a closely related alternative unresolved.

We have added TC-λ (Maystre et al., 2025) as a baseline in Tables 1–3. TC-λ (Appendix C.2.1) trains the prediction at each prefix toward a recursive target built from the model's own later predictions and the realized outcome, so that successive predictions stay temporally consistent. In most experiments TC-λ improves on PREFIX, confirming that this objective helps evolving predictions. BELIEF, however, outperforms TC-λ on 16 of the 17 metrics, on synthetic paths, real reference paths and downstream tasks, showing the advantage of evidence learning over temporal-consistency training.

Synthetic (Table 1, mean of 6 encoders)Path KL↓Path MAE↓mv→1AUROC↑AUPRC↑corr↑CRPS↓
BELIEF0.01500.09850.96650.97140.94590.78650.3410
TC-λ0.12870.32481.68670.90900.85310.61420.4150
Real data (Table 2)WeatherSteel-RBSoccer
PACE-MAE↓AUROC↑PACE-MAE↓AUROC↑PACE-MAE↓AUROC↑
BELIEF0.26370.77620.25240.85730.08960.9667
TC-λ0.27130.84350.37970.56160.09540.9606
Downstream (Table 3)P12 AUROC↑MIMIC-III AUROC↑Geolife AUROC↑P12-LoS MAE↓
BELIEF0.86080.84530.78340.6062
TC-λ0.82670.81430.78320.6367

Q1 — Informative observation mechanisms

When the observation mechanism violates the conditional independence assumption in
Equation (7), how are BELIEF’s belief-path recovery error and calibration affected?
Does it become overconfident or fail to fully exploit information conveyed by the
observation process?

Thank you for raising this point. Our original formulation considered observation processes that carry no outcome information, as in many financial or sports records. Your question made us consider settings such as healthcare, where the decision to observe can itself carry information and violate Eq. (7). Following your suggestion, we have extended our theory to include this information (Appendix E.1). Writing zj=(τj,mj)z_j=(\tau_j,m_j) for the time and variable of an observation,

\underbrace{p_\theta(x_j\mid H_{j-1},z_j,y)}_{\text{value likelihood}} \underbrace{r_\psi(z_j\mid H_{j-1},y)}_{\text{observation-mechanism likelihood (added)}}.$$ Under Eq. (7), the added term is the same for all candidate outcomes and cancels, which recovers the original BELIEF. **Synthetic data.** We constructed tasks in which observation times, variables, or both depend on the outcome (Appendix E.2, Table 11). Without the added term, BELIEF does not become overconfident, but it leaves this information unused and its path KL increases. Adding the term lowers both path KL and log loss under every informative mechanism: |Mechanism|Path KL↓ (6 tasks)||Log loss↓ (4 classification tasks)|| |-|-|-|-|-| ||Value only|+Mechanism|Value only|+Mechanism| |Original (Eq. 7 holds)|**0.0162**|0.0165|**0.5369**|0.5384| |Informative time|0.0356|**0.0311**|0.5394|**0.5328**| |Informative variable|0.0291|**0.0271**|0.5406|**0.5382**| |Informative time+variable|0.0434|**0.0327**|0.5422|**0.5304**| **Real data.** The added term brings no consistent improvement and can lower performance (Table 12). A likely reason is that the outcome information in real observation processes is weak relative to the error of learning their likelihood, and this error accumulates over a trajectory. Weather reports and soccer matches follow largely fixed schedules, and even irregular clinical measurements add limited information beyond the measured values. Without the term, BELIEF avoids this error and stays well calibrated on most real tasks. ## Q2 — Direct posterior-grid evaluation > ```text > If distributional error is computed directly from the predicted posterior on the outcome > grid, rather than from a moment-matched Gaussian approximation, does BELIEF retain its > advantage on continuous-outcome tasks? > ``` Yes. We evaluated all predicted posteriors directly on a common grid (Table 7, Appendix D.1). The grid KL is very close to the original KL, so the conclusion is unchanged: BELIEF has the lowest grid KL in all 12 setting/encoder combinations.

Thank you for your response. It has addressed most of my concerns, and I will maintain my positive rating.

AuthorsOpenReview ↗

Thank You for Your Response

Dear Reviewer im3U,

Thank you very much for taking the time to read our response and for maintaining your positive assessment of our work. We are glad that our response has addressed most of your concerns. We deeply appreciate your constructive feedback, which has helped improve the quality and clarity of our paper.

We would appreciate knowing whether there are any remaining concerns that prevent you from giving a higher rating. We would welcome the opportunity to address any such concerns promptly and provide further clarification.

Thank you again for your time and consideration.

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

This paper considers the question of forecast revisions as new information comes in before resolution. This is an interesting and timely problem, especially with the rise of things like markets where one can bet on whether an event is going to occur or not. The framework BELIEF proposed by this paper aims at learning the entirety of the belief distribution conditional on the preceding history. This is done through an encoder-decoder framework which is trained by using an autoregressive manner. The model learned is then used for Bayesian belief updating. Finally, the proposed methods are tested on some real-world data sets and demonstrate some improvement in performance.

Critical strengths

This is an interesting problem to be sure, and the authors make a meaningful contribution to it. The architecture proposed seems sufficiently flexible yet sufficiently easy to be used in practice. I appreciate the clear description of the architecture.

Critical weaknesses

The experiments could be written up a bit more clearly. It is a bit hard to understand how it performed for some specific examples. I cannot really appreciate the experiments to their full extent as they are presented currently. More typical belief paths should be shown.

Questions for authors

Can you sample over paths in the posterior?

2 comments
AuthorsOpenReview ↗

Response to Reviewer RCrE

Thank you for your careful review and constructive comments. We are glad that you find the problem timely and the architecture flexible and easy to use in practice. We address your comments below.

W1 — Clearer experiments and typical belief paths

The experiments could be written up a bit more clearly. It is a bit hard to understand
how it performed for some specific examples. I cannot really appreciate the experiments
to their full extent as they are presented currently. More typical belief paths should be shown.

In the experiments section (Sections 4.1–4.4), we have explained the synthetic results and the evaluation metrics in more detail, and we have revised the table captions so that the results are easier to read.

We have also added typical belief paths (Fig. 3, Section 4.6) for common situations in which the baselines can fail. When all methods reach similar final predictions, BELIEF tracks the posterior more closely at intermediate events. When the evidence reverses, the baselines are misled and end with the wrong prediction, whereas BELIEF follows the fall and recovery of the posterior. When observations carry no outcome information, BELIEF stays neutral more stably than the baselines.

Q1 — Sampling over paths

Can you sample over paths in the posterior?

Yes. BELIEF can sample future belief paths by generating future evidence and applying its Bayesian update. As shown in the following equation, the current belief qθ,jq_{\theta,j} weights the outcome-conditioned observation distributions to form a posterior-predictive distribution for the next value:

pθ(x∣Hj,z)=∫qθ,j(y) pθ(x∣Hj,z,y) dμ(y),p_\theta(x\mid H_j,z)=\int q_{\theta,j}(y)\,p_\theta(x\mid H_j,z,y)\,\mathrm{d}\mu(y),

where z=(τ,m)z=(\tau,m) gives the time and variable of the next observation. We sample the next value from this distribution, append it to the history, and update the belief. Repeating this generates possible future belief paths. Fixing a specified outcome instead yields outcome-conditional paths. We have clarified both procedures in Appendix B.4.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer RCrE,

Thank you very much for your careful review and positive assessment of our work.

We have addressed your comments in our previous response and the revised paper. As the discussion period is coming to an end, we would greatly appreciate knowing whether our response has addressed your concerns or whether any questions remain. Given the limited granularity of the four-point rating scale, if you feel that our work and response justify a higher rating, we would be grateful if you would consider raising your score. We would be happy to provide further clarification if needed.

Thank you again for your time and careful consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We have provided detailed responses to the reviewers’ comments, completed additional experiments and analyses, and already incorporated the corresponding revisions into the paper.

Following Reviewer RCrE’s suggestions, we improved the presentation of the paper. We also addressed Reviewer im3U’s concerns through additional experiments and clarifications. In addition, we provided point-by-point responses to Reviewer Tu4N’s questions and concerns. We believe that our detailed explanations and additional results have adequately addressed their main concerns.

Core Contributions of Our Work

We introduce belief-path learning, which aims to recover how beliefs about the same eventual outcome evolve as evidence arrives. In many applications, this probability must be updated while the process is only partially observed. In intensive care, a patient’s mortality risk is updated as vital signs and laboratory results arrive, supporting early warning. In futures markets, the probability that a price reaches a given level is revised as new market data are released, guiding hedging. In these and other fields, decisions require tracking how outcome probabilities evolve as new evidence arrives, rather than predicting only the final outcome. We believe this task could open a promising research direction for bringing time-series learning to real-world applications.

We propose BELIEF, a flexible framework that models incoming observations conditioned on the preceding history and candidate outcomes, and updates beliefs through Bayesian inference. It does not require true posteriors or belief paths as supervision, so it can be trained directly on historical event streams and their realized outcomes. We provide quantitative error bounds and exact recovery under the stated assumptions. Experiments show accurate recovery of analytic posterior paths on synthetic data and the closest alignment with reference paths on real data, while maintaining competitive downstream predictive performance.

Clarifications and Presentation

  • Clarifying training with shared and incomplete prefixes (Tu4N, Q1/W2). BELIEF uses the realized outcome as a conditioning input and the next observed value as the training target. Thus, trajectories sharing a prefix but ending with different outcomes do not create conflicting training targets. This design also avoids directly supervising incomplete prefixes with their final labels, which can lead to unreliable intermediate predictions (Response Tu4N (1/2), Q1/W2).
  • Final prediction and reference paths (Tu4N, Additional response). The downstream experiments verify that learning belief paths does not come at the cost of final prediction performance. On real data, true posteriors are generally unavailable, so we evaluate alignment with external probability paths from forecasters and markets. BELIEF does not use these paths for training, so close alignment shows that it learns comparable beliefs directly from raw evidence.
  • Clearer experiments and typical belief paths (RCrE, W1). We improved the experimental presentation and added representative paths (Section 4.6, Fig. 3).
  • Sampling over paths (RCrE, Q1). We clarified posterior-predictive sampling of future observations and resulting belief paths (Appendix B.4).

Additional Experiments and Analyses

  • Adding new labeled outcomes (Tu4N, W3/W5). Once labeled trajectories of a previously withheld class become available, BELIEF can add it by updating only a small subset of parameters. It achieves lower path KL and less forgetting of existing classes than the baselines in both settings (Appendix D.7, Table 10).
  • Cost versus number of candidates (Tu4N, W4/Q2). We added inference-latency and peak-memory comparisons (Section 4.7, Fig. 4; Appendix D.5, Fig. 9). With parallel evaluation, BELIEF keeps its latency close to direct predictors. Our follow-up response further explained the timing and memory results.
  • Tracking rapid evidence reversals (Tu4N, W2). BELIEF tracks rapid changes in evidence with lower reversal-window KL than the baselines (Appendix D.6, Table 9).
  • Temporal-consistency baseline (im3U, W1). We added TC-λ, and BELIEF retains its path-recovery advantage (Tables 1–3).
  • Direct posterior-grid evaluation (im3U, Q2). BELIEF has the lowest grid KL in all 12 setting–encoder combinations (Appendix D.1, Table 7).
  • Informative observation mechanisms (im3U, Q1). We extended the update and recovery guarantee to informative observation times and variable identities. This improves synthetic path KL and log loss, but not consistently on real data (Appendices E.1–E.2, Tables 11–12).

We believe these additional results and clarifications strengthen the evidence for BELIEF’s contribution. Thank you again for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.