acceptodds
Under review as a conference paper at ICLR 2027

Predicting and Unpredicting: Anchor Choice in Joint Inference

Abstract

How can dynamics learned by prediction help explain hidden states from later evidence? We reuse a frozen library of locally invertible mechanisms to propose trajectories on unseen compositions, then weight them under a joint posterior when innovations and observations are uncertain. A conditional bound separates inverse execution, model error and uncertainty in the missing inputs. In a predeclared 24-world held-out study with terminal and midpoint observations, the tested midpoint-anchored proposal lowers effective sample size and the fraction of runs meeting full-trajectory quality criteria in expanding chains relative to terminal anchoring, but improves both in contracting chains. Both proposals target the same learned posterior. At matched full-trajectory posterior quality, a direct-inverse policy selected on separate development worlds lowers complete query cost against continuous-grid smoothing in expanding chains with both observations at both tested horizons. The result identifies where direct recovery helps joint inference and where unresolved uncertainty still governs the answer.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian inference, state-space smoothing, importance sampling, invertible neural networks, learned dynamics, compositional generalization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_kFVCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how dynamics models learned from forward prediction can be reused for inverse state inference. The authors propose a framework that uses frozen, locally invertible mechanisms to reconstruct hidden trajectories from later observations. The method combines direct inverse execution with importance weighting to account for uncertain innovations and noisy observations. The paper further derives a conditional error bound and compares terminal-anchored and midpoint-anchored proposals on unseen mechanism compositions. Experiments across 24 held-out synthetic worlds show that the preferred anchor depends on whether the dynamics are expanding or contracting. The results also demonstrate computational advantages over continuous-grid smoothing in certain expanding-dynamics settings.

Critical strengths

The main strength of this paper is its systematic investigation of inverse inference using previously learned dynamics without retraining separate inverse models. The framework explicitly distinguishes inverse execution error, model error, and uncertainty in unobserved inputs. Furthermore, the held-out experiments compare different proposal anchors under a common posterior target, providing useful insights into how dynamical properties affect inference efficiency.

Critical weaknesses

The main limitation is that the experimental validation remains restricted to synthetic systems with scalar hidden states and relatively simple transition mechanisms, making it difficult to assess the generalizability of the findings. Although direct inversion demonstrates computational advantages in certain settings, Table 1 shows that only two of the eight experimental configurations satisfy the predefined matched-quality cost criterion. The learned-versus-true posterior discrepancies also remain substantial, raising concerns about the practical effectiveness of inverse inference when the learned dynamics are imperfect. In addition, the conditional error bound does not directly explain the observed differences in effective sample size between proposal anchors.

Questions for authors
  1. The experiments focus primarily on scalar-state synthetic dynamical systems. Could the authors clarify whether the proposed approach can be extended to higher-dimensional or more complex nonlinear systems? How might the computational advantages change as the state dimension and uncertainty increase?
  2. In Table 1, only two experimental configurations establish a reliable computational advantage over grid-based smoothing. Could the authors clarify the conditions under which direct inversion is expected to outperform structured smoothing methods? How sensitive are these results to the proposal distributions and observation noise?
  3. The learned-versus-true posterior discrepancies remain relatively large despite accurate inference under the learned model. Could the authors further discuss how prediction errors affect downstream state recovery and whether improved forward-model accuracy would change the observed anchor preferences?
1 comment

Your distinction between an accurate sampler and an accurate model gets at why we froze the dynamics. Once that fit is reused for a reverse query, changing the proposal tests the sampler while keeping the model error fixed. Thank you for asking where this works as dimension, observation precision and cost change.

Higher dimensions and cost. We added nonlinear vector-state tests and a matched-accuracy budget comparison in revised Appendix W. Against the same fitted posterior, both guided anchors and HMC passed every marginal check in 8D and 16D. Gauss–Newton missed one in 16D. The original proposals passed at most 3/24 checks in 8D despite using four times as many draws. In a separate cost comparison, the 2,048-path midpoint guide cost less in two of three qualifying settings, one in 8D and one in 16D. The baseline was the cheapest tested Gauss–Newton budget that met the same 80% pass criterion. The fourth setting did not meet that criterion (revised Appendix W).

Cost against grid smoothing. At matched fitted-posterior accuracy, the original policy was cheaper in two of eight settings. Both used expanding chains and two readings, at 16 and 32 steps, where grid cost was 1.53× and 1.57× the direct method’s cost (original Table 1). Both methods had to pass at least 80% of 36 held-out cells. The 95% world-bootstrap interval for log(grid/direct cost), computed on cells where both passed, had to lie above zero.

Proposal sensitivity. In a post-review 16-step comparison, replacing clipped-gradient with quadrature guidance raised midpoint full-path passes from 8/36 to 36/36 with the anchor proposal, two readings, fitted posterior and 8,192-path budget fixed. Grid smoothing also passed all 36 checks and cost less than the guided midpoint (revised Figure 6 and Appendix C).

Observation noise. In the added Appendix U sweep on expanding 16-step chains, both quadrature-guided anchors passed every full-path check at the tested sensor-noise standard deviations from .05 to .20. At .01 under the refined reference, the original proposals passed none, while guided terminal and midpoint passed 25/36 and 18/36 (revised Appendix U).

Forward-model error and anchor choice. In the added true-dynamics comparison, the forward fit accounted for the larger discrepancy at the initial state. Fitted-to-true posterior KS was 0.409 in expansion and 0.272 in contraction, while guided sampling had KS of 0.009–0.013 against its fitted posterior. Using the true dynamics changed the effective sample size (ESS) gap. Under both models, the terminal anchor had higher ESS in expansion and the midpoint anchor in contraction (revised Appendix U). A next test is to vary forward-fit accuracy at fixed observation settings and sample budget, and check whether the ESS ranking changes within each proposal family.

Recovery bound and ESS. Under the inverse identities and domain conditions in original §3, the forward-fit error contributes at most ρ(1−ρH)δ/(1−ρ)\rho(1-\rho^H)\delta/(1-\rho) to the starting-state recovery error when the inverse-state gain is ρ<1\rho<1, forward error is uniformly bounded by δ\delta, the terminal state and disturbances are correct, and inverse execution is exact. Revised §3 adds the ESS limit for independent draws, with proposal coverage and finite weight variance. It makes explicit how ESS depends on the full target-to-proposal density ratio.

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

According to my limited understanding, this paper studies how to make inverse estimation of the one-step transition function (likely 1-D?) that are fitted with input state xtx_t and noisy next state xt+1x_{t+1}. Then at inference time, the learned transition function is applied to a new trajectory without intermediate states observed and the task is to perform posterior inference over the hidden trajectory.

To be frank, I think the relationship between what being studied here and the "world model" as outlined at the start of the intro is very weak. I don't think this paper is well motivated and many places I feel confusing about. For example, I'm confused by what the setup is, how it is motivated, and what are the challenges, how this work differs from the previous. It could be that I'm not familiar with this field and not qualified for reviewing this line of work.

Critical strengths

NA

Critical weaknesses
  1. The problem is poorly motivated. It's not clear why choosing the anchor of this kind of sampler is an important question. The paper gives no application where recovering earlier hidden states with a frozen invertible model is the bottleneck, and no argument that the answer would change how people build or use world models. The intro leans on world models like Dreamer and LeWM, which learn high-dimensional, latent and usually non-invertible dynamics for planning. The paper instead studies a scalar state with a few monotone splines learned from fully observed pairs, so the link to world models is weak. Even the water-tank example falls outside the contraction regime the theory needs.
  2. The setup and its challenges are unclear. As far as I can tell, the problem is posterior smoothing in a known scalar state-space model, and the question is which time step should seed an importance-sampling proposal. The paper doesn't explain why this is hard or how it differs from existing smoothers. Sec. 6 even notes that two-filter smoothers already sample a midpoint state and complete both sides. The heavy private vocabulary ("unpredicting", "gate attainment", "strata") also makes the paper hard to follow.
  3. The main finding is expected and the proposals compared are naive. That backward proposals work better in expanding chains and worse in contracting ones follows from standard SMC intuition and from the paper's own Sec. 3. Both proposals are simple hand-designed ones, and the midpoint proposal ignores the tight terminal reading when drawing the forward segment. Stronger options such as bridge proposals, particle smoothers or HMC on the joint are not compared.
  4. The practical gain is small and hard to verify. The cost advantage over the grid smoother is about 1.5×1.5\times in only 2 of 8 settings, and elsewhere the grid is cheaper or the direct method fails. "Matched quality" is measured against the learned posterior, which differs noticeably from the true one. The supplement can't reproduce the full experiments, and Appendices E–K study a largely unrelated problem.
Questions for authors

See my weakness.

1 comment

Thank you for asking how this relates to predictive world models. We clarified what a forward fit can infer on new sequences and added stronger inference comparisons.

Connection to latent world models. DreamerV3 trains a policy using imagined rollouts, and LeWorldModel (LeWM) plans with predicted latent trajectories. Both illustrate how forward-trained dynamics can be reused. We revised the abstract and §1 to ask what the same fit can tell us about the past from later readings. Section 1 now distinguishes the observed state pairs used to fit each locally invertible transition from the earlier states and stochastic disturbances hidden at query time. The mechanism sequence and observation model are supplied. Extending the test to latent world models would require specifying an earlier quantity of interest and checking whether the learned representation and later observations retain enough information to infer it. Hafner et al. (2025) · Maes et al. (2026)

Relation to smoothing. Original §6 already describes midpoint sampling and completing both sides. Revised §6 explains how the backward filter in two-filter smoothing can use observations to guide proposals. On unseen four- and eight-step sequences in ten scalar worlds, inverse sampling closely matched numerical inference under the same fit (CDF risk 0.002209 vs. 0.002191 against the true posterior, original Appendix B.1). The tank illustrates inverse execution with inverse gain 1/0.9>11/0.9>1. The §3 bound applies under inverse contraction, with the terminal state and disturbances given.

Measured temperatures. We added an offline study in 20 held-out Catania homes in response to questions about practical use. A single forward fit per fold supported forecasting and reconstructing an earlier indoor reading masked for evaluation by analytically conditioning on outdoor context and two later readings with added noise. Mean within-home reconstruction RMSE was 0.411°C for the shared model and 0.424°C for a separately trained predictor given the same inputs (revised Appendix X).

Joint inference. We added a two-reading guidance comparison with the midpoint fixed. Replacing clipped-gradient with quadrature guidance raised full-path passes from 8/36 to 36/36 in 12 expanding 16-step worlds, with the anchor proposal, both readings, fitted posterior and 8,192-path budget fixed. Guided terminal and grid smoothing also passed all 36 checks, and grid smoothing took less mean query time than the guided midpoint (revised Figure 6 and Appendix C).

Other inference methods. In a separate added scalar cohort, the guided particle smoother and the regime-selected two-reading anchor passed 25/36 and 35/36 in expansion and both passed 36/36 in contraction (Appendix C). In the 8D study, the guided anchors, Gauss–Newton importance sampling and Hamiltonian Monte Carlo all passed every fitted-posterior marginal check (Appendix W).

Cost and model error. The original policy cost less than grid smoothing at matched fitted-posterior accuracy in two of eight settings, both expanding chains with two readings at 16 or 32 steps. Grid cost was 1.53× and 1.57× the direct method’s cost (original Table 1). The added comparison against the true posterior shows that discrepancy from the forward fit is larger than sampling error under that fit (revised Appendix U).

Reproduction. The original supplement replayed saved results. The revised package adds code and configurations for the original joint study, unseen compositions and guided smoothers, plus a saved-record figure replay (revised Appendix V and ZIP README.md). The same-host rerun matched 17,485 of 17,496 original trajectory/weight pairs within 10⁻¹⁰ (revised Appendix V). A separate run reproduced all 576 guided-SMC non-timing records (supplement code-config/README.md). An isolated Python environment on that host regenerated 48 cases. An independent metric implementation agreed on comparable quantities for all 48 cases using the same fitted-posterior reference (revised Appendix V). Full external reproduction by an independent team and measuring gains in real workflows are future work.

Context experiments. Original Appendices E–K document a separate context-inference study.

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

It is very hard to discern what the purpose of this paper is.

Critical strengths

It is very hard to discern what the purpose of this paper is.

Critical weaknesses

I am curious to the answers the authors provide to my questions.

Questions for authors

Where in the paper is a clear description of the problem that is to be solved? Where can I find a clearly expressed definition of what the benefit to machine learning would be if this problem was solved? Where can I find a quantifiable definition of success? Where can I find a justification of the metrics that are used? What is the most realistic use-case that is demonstrated in the paper? If there are no realistic use cases, what is the argument that the use cases demonstrated in the paper contribute to solving realistic problems?

2 comments

Thank you for asking us to clarify the task and how we evaluate it. We infer what could have happened before a noisy reading using the same model that predicts what happens next.

Setup and benefit. We revised the abstract and §1 to put the reverse query first. Section 1 now distinguishes observed state pairs at training from noisy readings on a new sequence at query time. The same fit can answer questions absent from training. In ten scalar worlds, sampling through the learned inverse on unseen four- and eight-step sequences closely matched numerical inference under that fit (original §4 and Appendix B.1).

How we measure success. Revised §5 explains the purpose of the three original checks against the fitted posterior. A distance between trajectory distributions (path MMD², ≤2×10⁻⁵) checks joint accuracy. Marginal KS (≤0.03) checks the distribution of the state at each of the initial, midpoint and final times. Root-sign error (≤0.05) checks how much probability goes to each sign of the initial state when a squared reading makes either sign possible. A cost advantage requires each method to pass at least 80% of runs and a positive paired 95% world-bootstrap interval for log(grid/direct query cost) on runs where both methods pass. The original policy met this rule against grid smoothing for expanding chains with two readings at 16 and 32 steps (original §5, Table 1 and Appendix A).

Demonstrated use. We added a study in 20 held-out Catania homes. One forward-fitted model per fold forecast four hours ahead and reconstructed an earlier indoor reading masked for evaluation by analytically conditioning on outdoor context and two later readings with added noise. Mean within-home reconstruction RMSE was 0.411°C for the shared model, 0.424°C for a separately trained predictor with the same inputs and 0.432°C for backward carry (revised Appendix X).

From controlled paths to measured readings. With the fitted model, midpoint anchor, both readings and sample budget fixed, switching the disturbance guide from clipped gradients to quadrature improved agreement with the fitted posterior over entire paths (revised Figure 6). The held-out-home study demonstrates the same forward fit serving forecasting and reverse reconstruction (revised §7 and Appendix X). Measuring gains in real workflows is future work.

After reading the rebuttal, I still cannot understand this paper. This paper requires serious rewriting before any resubmission.

Comments on the paper

1 comment

World models use learned dynamics to predict what happens next. A tank’s current water level and a known addition predict its later level. Given a noisy later reading, we ask which earlier levels and additions could have produced it. We test whether the same fitted dynamics can answer this reverse question on new sequences.

We thank the reviewers for asking how this reverse query relates to world models with learned latent states. We revised the abstract and §1 to distinguish training on observed pairs of successive states from querying with a known transition sequence and noisy readings. Earlier states and disturbances are hidden at query time. Our tests use locally invertible transitions and a specified observation model. This focus led us to change the title from Predicting and Unpredicting: Anchor Choice in Joint Inference to Predicting and Unpredicting: Reusing Learned Dynamics for Reverse Inference.

On unseen four- and eight-step sequences, the inverse sampler closely matched numerical inference under the same fit (original Appendix B.1). Original §3 bounds error accumulation when backward steps shrink state errors and have bounded sensitivity to disturbances on a common domain. In an added simulation with known dynamics, replacing the true dynamics with the fitted model changed the inferred starting-state distribution more than imperfect sampling under that model (revised §7 and Appendix U).

To check whether the midpoint result depends on how we sample disturbances, we kept candidate paths starting at the midpoint and fixed the two noisy readings, fitted model and 8,192 paths per query. Replacing clipped-gradient guidance with numerical quadrature raised the number of runs that passed checks against the fitted model’s path distribution from 8/36 to 36/36 across 12 systems where forward dynamics separate nearby states, with 16 steps and three repeats per system. Guided sampling starting at the final reading and grid smoothing also passed 36/36. Grid smoothing took less mean time per query than the guided midpoint method (revised §5, Figure 6 and Appendix C).

On 20 held-out homes in Catania, the same model fitted on other homes forecasts four hours ahead and reconstructs a masked earlier indoor reading from outdoor temperature and two later readings with added noise. It reconstructed about as accurately as a separate predictor trained for that query with the same inputs (revised §7 and Appendix X). A next step for latent world models is to test whether their learned dynamics and later readings can recover a specified earlier quantity on new sequences, measuring fit and inference errors separately.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.