acceptodds
Under review as a conference paper at ICLR 2027

When Are Joint Predictive Models Reusable?

Abstract

Joint predictive models are reused after training under new compositions, observations, and questions. Familiar-task accuracy alone does not establish which later answers remain available. We ask what training records must reveal and what a retained state must preserve for such reuse. For known finite operators, we characterize the information required by future queries and evidence updates. Across 32 newly sampled worlds, records targeting a theoretically identified missing three-way interaction reduce conditional-query excess Brier score by 0.0206 relative to equal-budget irrelevant records. At the primary budget, direct fits of a sufficient target marginal with seven free parameters and the full 64-state law achieve nearly identical mean risks. In a separate dynamic intervention, removing past–current coupling preserves tested forward predictions but stops revision of the past; a five-coordinate state preserves revision for a fixed past event. Within the supplied mechanism class, local mechanisms learned from chain endpoints also predict unseen trees and use their outputs to revise later predictions with fixed parameters. Thus familiar-task accuracy can conceal different reuse abilities, while the information sufficient for a future use can be smaller than a full world law.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, sufficient statistics, predictive state representations, belief state compression, compositional generalization, probabilistic inference

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ZCLSOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Authors study when a learned probabilistic model remains useful after training when the graph composition, available evidence, or query changes. Its central theoretical message is sensible: good performance on the original prediction task does not guarantee that the learned representation contains the joint information required for later conditional queries. The paper formalizes this through identifiability and retention criteria and supports the theory with several controlled experiments.

Critical strengths

I find the theoretical perspective interesting and several controlled experiments convincing as demonstrations of the proposed information criteria. In particular, the distinction between information needed for current prediction and information needed for subsequent conditioning is useful. The authors report uncertainty across multiple generated worlds and disclose several negative or failed checks rather than presenting only successful experiments.

Critical weaknesses

However, the paper's practical contribution remains unclear. The motivating applications are broad, while the formal results assume known finite operators, declared future queries/evidence, and supplied structural interfaces, and most experiments are constructed probabilistic systems designed around those assumptions.

The paper motivates the work using world models, LLMs, compositional inference, and prediction under changing evidence, but the connection between these applications and the formal experimental setting remains abstract.

I also found the terminology around "prediction" insufficiently precise. Several experiments are more naturally understood as probabilistic classification or conditional-probability estimation followed by Bayesian updating. This matters because the paper's central statement that "prediction quality does not imply reusability" could be interpreted much more broadly than what is actually demonstrated. The paper would benefit from explicitly distinguishing these notions and narrowing its claims accordingly.

Questions for authors

I would ask the authors to explicitly define what they mean by prediction, how it differs from probabilistic classification in their experiments, and whether the results apply to arbitrary predictive models or specifically to models intended to represent coherent probabilistic laws.

I would like to see at least one experiment on an established world-model environment showing two models with comparable familiar-task prediction/classification performance but meaningfully different downstream reuse ability.

The conclusions should therefore distinguish more sharply between: what is proved for known finite probabilistic systems? what is demonstrated in synthetic learned mechanisms? and what is merely suggested for modern neural world models?

1 comment

Thank you for asking how prediction, evidence updating and world models connect. We clarified the scope of the finite criteria and added the requested test on simulated walking (DeepMind Control Suite Walker).

Matched familiar prediction. The new simulated-walking comparison shows that models with similar familiar accuracy can differ after the same feedback. For two neural models trained independently on the same trajectories, all four paired 95% intervals for familiar-task Brier differences lay within ±0.005 on 8,192 fresh episodes. Their reversed-plan scores were close before feedback. Given the same step-four height, step-eight Brier was 0.1193 vs 0.1430, with paired gap 0.0237 [0.0216, 0.0258]. A directly supervised classifier for that query scored 0.1273. Both action plans occur in training. A separate added Gaussian control uses its own 2,048 episodes and holds the single-time distributions fixed while changing cross-time covariance (revised §8 and Appendix A). The supplement's walker_md_aggregate/verify.py checks the neural Brier point estimates from score sums.

Prediction and the finite criteria. The revised Introduction defines a prediction as the expectation E[f], which for an indicator is the class or event probability used in probabilistic classification. Evidence with likelihood L changes it to E[fL]/E[L]. For coherent laws, known finite operators, strictly positive initial laws and linear expectation summaries, our exact criteria characterize when records and states determine both conditional answers and evidence probabilities. Revised §4 clarifies that a recursively updated state must also tell which evidence updates are available and determine the resulting summary (see also original §4).

Which records to collect. In original §6, the test identifies a triple interaction missing from pair records. In 32 fresh worlds, collecting relevant triples reduced mean excess Brier from 0.022783 to 0.002182 compared with equal-cost unrelated triples, with improvement in every world.

Reuse on unseen trees. In the original experiment on eight binary worlds (§5 and Figure 2), frozen local mechanisms with the supplied generating factorization transfer to unseen trees under the same interface used by the graph-conditioned neural comparison. Familiar-chain risks meet the prespecified matching margin. On unseen trees, excess Brier is 0.000320 for the local mechanisms and 0.021835 for the graph-conditioned model.

Additional neural evidence and next test. In a separate added 512-episode check, both released DreamerV3-port checkpoints improve their own Walker forecasts after noisy height evidence (revised Appendix A). A next question for general-purpose pretrained world models is whether they can identify which intermediate measurement will most improve a later forecast (Discussion).

Reviewer_CqW6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In my opinion, the submission is unreadable. I cannot identify or summarize the contributions.

Critical strengths

None

Critical weaknesses

In my opinion, the submission is unreadable. No clear objectives, no clear contributions, no mathematical rigor.

Questions for authors

None

1 comment

Thank you for raising the question of clarity. The revised abstract distinguishes what to observe and what to retain. The Introduction poses both questions through predicting a third observation from two.

What to observe together. When predicting a third observation from two, pair records can miss a three-way interaction. Relative to equal-cost unrelated triples, the relevant measurements reduced mean excess Brier from 0.022783 to 0.002182 and improved prediction in all 32 fresh worlds (original §6). Revised §7 now reports that the five-coordinate state matched both the selected past-event forecast and per-step evidence probability in the existing replay (original §7 and Appendix E.2).

Why the information test is exact. For known finite operators, strictly positive initial laws and linear expectation summaries, population records determine the declared conditional answers and evidence probabilities exactly when the span of the observation functions contains the required functions. The same test applies to stored state coordinates. A missing direction permits two strictly positive laws with identical records but different requested answers or evidence probabilities. When all directions are present, linear combinations recover the query numerator and evidence probability. Their ratio gives the conditional answer. Recursive state use also requires well-defined updates (original §4, Eq. (9), Appendix C). Revised §4 states which evidence histories enter the finite-horizon test and clarifies when the next update is determined by the summary.

Reviewer_75uxOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is very clearly AI-generated as authors also clearly state in the AI use disclosure:

""" AI tools assisted hypothesis and protocol development, mathematical derivations and proof checks, implementation of synthetic-data generators and training/evaluation code, result interpretation, literature search, manuscript preparation, and artifact-consistency checks. """

As it stands the paper is completely unreadable. Consider the first two lines of the introduction section as an example:

"In world-model research, learned models are increasingly used with new conditioning roles, compositions, and feedback after training (Hafner et al., 2025; Alzayer et al., 2026; Zhu et al., 2025; Bakermans et al., 2025). Large language models face a related question: accessible component facts need not yield a correct multi-hop answer (Press et al., 2023; Yu et al., 2026)."

It is not at all clear to me what the authors mean by "accessible component facts" or how are "new conditioning roles" and "compositions" being defined. The remaining text is also riddled with the same issues. I have tried to make multiple passes over the abstract and introduction section and I am still struggling to understand the problem statement of this work. Also, given the AI use statement, I am not really sure what was the contribution of the authors, since everything from mathematical derivations, developing the hypothesis, writing the manuscript, and even interpreting the results was done with AI tools.

Critical strengths

NA

Critical weaknesses
  • Completely unreadable paper
  • Authors' own contributions not at all clear, since everything from mathematical derivations, developing the hypothesis, writing the manuscript, and even interpreting the results was done with AI tools.
Questions for authors

NA

1 comment

Thank you for the feedback on the opening. The revised Introduction opens with a concrete example about three observations and uses Figure 1 to explain composition and how an observation changes roles.

From chains to trees. In the original example, we learn two mechanisms from short chains, freeze them and connect them into a new tree. This reconnection through the supplied ports is what we mean by composition. The output starts as a prediction target. Once observed, it becomes evidence about the shared source and changes the next tree's forecast (original §5 and revised Introduction, Figure 1).

What to measure together. Even exact pairwise relationships can leave the probability of a third observation given the first two undetermined. Our test selects three to record together, improving prediction in all 32 fresh worlds versus equal-cost unrelated triples (original §§4 and 6). Here separately known facts are the relations available one pair at a time. Multi-hop language questions motivate asking whether separately known facts determine the combined answer.

Author contributions. The author team led problem formulation and collection of the theoretical and empirical evidence. The paper's AI-use statement describes the tools' assistance with hypothesis and protocol development, derivations, implementation, interpretation and writing.

Comments on the paper

1 comment

We thank the reviewers for asking what makes a predictive model reusable after new evidence. We revised the abstract around that use. The introduction now asks how to predict a third observation after seeing two others. For a given query, we identify what is needed to predict after seeing the evidence and to assign a probability to the evidence itself. One test asks whether the population training records determine that information. The other asks whether the model’s retained state preserves it. For known finite operators, strictly positive initial laws and linear expectation summaries, each test is an exact linear-span criterion (original §4).

The original experiments illustrate both design decisions. In the six-bit acquisition example, exact pairwise frequencies leave a relevant triple interaction undetermined. The test selects three variables to observe together before collecting new data. At equal cost, those records reduced mean excess Brier from 0.022783 to 0.002182 and improved prediction in all 32 fresh worlds. In twelve filtering worlds, five coordinates selected for a particular past event reproduced the full fitted filter’s predictions for that event after each observation. Six coordinates selected for a different event can give a different answer (original §§6–7). Revised §7 adds the existing check on per-step evidence probability to the main text (original Appendix E.2 and revised Appendix F.2).

In response to the request to test whether two models with similar familiar-task accuracy can use new evidence differently, we added a simulated walking comparison (DeepMind Control Suite Walker). Two independently trained neural trajectory models had comparable Brier loss on familiar height-event predictions. After both received the same intermediate height measurement, their Brier losses for a later height event when the two action blocks were applied in reverse order were 0.1193 and 0.1430. A separate Gaussian control fixes single-time forecasts while changing cross-time covariance (revised §8 and Appendix A). Familiar prediction accuracy can conceal differences in how models use feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.