acceptodds
Under review as a conference paper at ICLR 2027

Precise About the Wrong Law: Uncertainty Misses Pooled Sources

Abstract

Foundation inference models estimate a dynamical law and its uncertainty from a set of trajectories in a single forward pass. This interface assumes that every trajectory comes from the same system, yet the input carries no record of where each trajectory came from. We show that when trajectories from different systems are pooled, the models’ native uncertainty does not reveal the violation. In the released model for Markov jump processes, a two-source context and a single-law context matched to its one-step transition statistics receive nearly identical uncertainty. The complete trajectories still tell the two apart: a likelihood-ratio witness that knows the candidate laws, but not the source labels, separates them almost perfectly. Decoders trained on the uncertainty output, by contrast, stay near chance on held-out tasks. We call this failure context aliasing: the trajectories expose source heterogeneity that native uncertainty fails to flag. As in classical misspecification, the model fits a pooled pseudo-law while losing fidelity to its sources, with little change in uncertainty. Routing the same trajectories by source cuts generator and first-passage-time error by more than half. A simple transition-count test provides an effective check for source heterogeneity. In controlled simulations, the model for ordinary differential equations shows the same decoupling of error and uncertainty. On recorded motion, the response varies across record groups. Uncertainty from these models is conditional on an experiment; it alone does not establish whether its trajectories should be pooled.

Primary area:
learning on time series and dynamical systems
Keywords:
foundation inference models, uncertainty quantification, model misspecification, source heterogeneity, Markov jump processes, dynamical systems

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 3 reviews, above 27% of the others.
Reviewer_LsKSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines a failure case for foundation inference models that aim to infer one dynamical law, along with its uncertainty, from a collection of trajectories. The usual interface presumes every trajectory comes from the same underlying system, yet the input tensor provides no information about trajectory provenance or which source each trajectory belongs to.The authors refer to this failure mode as context aliasing: when trajectories from different systems are combined, the model produces a confident estimate of a pooled pseudo-law, and its built-in uncertainty does not indicate that the single-law assumption has been breached.

The paper is technically rigorous, well written, and notably open in distinguishing between uncertainty conditional on a one-law experiment and evidence that the experiment itself is sound. My primary concerns relate to how novel this is compared with standard work on model misspecification and posterior predictive checks, the limited coverage of backbones, the somewhat advantageous setup of the controlled stationary construction, and the disconnect between a strong oracle-style witness and the much weaker detector that could actually be deployed.

Critical strengths
  1. The paper addresses an important and often overlooked failure mode in scientific foundation models, and it provides suitable reproducibility details along with a discussion of limitations.

  2. The Markovization approach is technically straightforward and cleanly separates the impact of missing trajectory provenance, while the comparison between the oracle path-level detector and the model’s native uncertainty offers strong evidence of a representation gap.

  3. The paper tests meaningful control strategies and shows their practical implications for generator estimation and first-passage-time prediction.

Critical weaknesses
  1. The core idea is not clearly distinct from traditional notions of model misspecification. As a result, the contribution’s novelty mainly lies in extending this perspective to foundation-model interfaces.

  2. Just two base inference models are assessed. They both originate from a closely connected research line and use the same set-to-one-law interface.

  3. The paper contrasts the model’s native uncertainty with source error, even though that uncertainty was not intended to reflect that quantity. The authors might consider highlighting the semantic mismatch in what the uncertainty represents, rather than treating it as standard miscalibration.

Questions for authors

How do the results change with fewer trajectories or shorter trajectories?

How would the proposed test handle observation noise or partially observed states?

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

The paper studies foundation inference models (FIM-MJP, FIM-ODE), which infer a dynamical law and its uncertainty from a set of trajectories. The authors show that when trajectories from two different systems are mixed, the models fit a compromise law that matches neither system. Their uncertainty stays about the same as for single-system inputs. The authors call this "context aliasing" and suggest passing source labels or using a transition-count test to detect mixing.

Critical strengths
  • The problem is practically relevant.
  • The controls are done carefully.
Critical weaknesses
  • The main finding is expected. Under misspecification, the Bayesian posterior concentrates around the pseudo-true parameter and becomes more confident as data accumulate (Kleijn & van der Vaart, 2006; 2012). So it does not seem surprising that the uncertainty doesn't flag pooled sources. The authors acknowledge this however.
  • The theory seems trivial. Proposition 1 is the triangle inequality and Proposition 2 is the chain rule for KL divergence.
  • The writing is unnecessarily technical and heavy on jargon.
Questions for authors
  • The uncertainty head is trained only on single-law contexts. Why should it react to pooled ones, and what happens if mixtures are included in training?
  • Would existing misspecification detectors for amortized inference help in this scenario, e.g., [1]?

[1] https://arxiv.org/abs/2112.08866

Reviewer_ryjwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates context aliasing in scientific foundation inference models, where trajectories from different dynamical systems are pooled under a single-law assumption, potentially producing inaccurate source-specific estimates without a corresponding increase in native uncertainty. The authors formalize the limitations imposed by missing source provenance and construct a stationary Markovization control that matches the pooled adjacent-pair transition distribution while preserving differences in complete trajectory distributions. The main experiments with FIM-MJP demonstrate that source-relative estimation error increases with source separation while native uncertainty remains relatively stable. An oracle path-likelihood test achieves 0.999 AUROC in distinguishing heterogeneous from homogeneous contexts, compared with 0.555 AUROC for the best native raw-variance decoder. Oracle source-aware routing reduces generator and first-passage-time errors by 60.5% and 55.5%, respectively. Additional controlled experiments with FIM-ODE reproduce the observed error–uncertainty decoupling. Overall, the paper demonstrates that native uncertainty may fail to indicate violations of the common-source assumption, motivating explicit source provenance and heterogeneity diagnostics in scientific inference.

Critical strengths

1. Well-controlled experimental design isolating context aliasing. The stationary Markovization construction provides a particularly useful control by matching the adjacent-pair transition distribution while preserving differences in complete trajectory distributions. The comparison between near-perfect oracle path-level discrimination (0.999 AUROC) and substantially weaker native-variance decoding (0.555 AUROC) provides convincing evidence that source heterogeneity can remain undetected by the model's uncertainty output. This controlled isolation of the phenomenon is a key strength of the paper.

2. Empirical validation and scientific relevance. The experiments demonstrate that increasing source separation leads to substantially higher estimation error without a corresponding increase in native uncertainty. The reported reductions of 60.5% in generator error and 55.5% in first-passage-time error under oracle source-aware routing establish that the issue affects scientifically meaningful quantities beyond uncertainty estimates alone. The additional FIM-ODE experiments provide supporting evidence that the phenomenon is not restricted to a single inference architecture.

Critical weaknesses

W1. Limited novelty relative to existing work on model misspecification and heterogeneous dynamical systems. The central phenomenon of confident inference around a pooled pseudo-law is well established in the misspecification literature, while prior work has investigated detecting misspecification in amortized inference and identifying mixtures of Markov chains. The theoretical results largely formalize these existing principles, and the main novelty appears to lie in their empirical demonstration using pretrained foundation inference models and the matched Markovization control. A clearer differentiation from prior work would help establish the significance of the contribution beyond applying known statistical phenomena to a new class of models.

W2. Limited end-to-end validation of the proposed non-oracle remedies. The paper demonstrates substantial improvements in generator and first-passage-time estimation through oracle source-aware routing and separately evaluates training-free heterogeneity detectors. However, it does not establish whether routing trajectories using inferred source partitions achieves comparable improvements. Furthermore, detection performance is primarily reported using AUROC, without evaluating false-positive rates and detection power at operational thresholds. These gaps limit the evidence supporting the proposed deployment recommendations when source provenance is unavailable.

Questions for authors

Q1. Novelty relative to prior work. Could the authors clarify how the proposed context-aliasing characterization and stationary Markovization analysis differ from existing work on Markov-chain mixtures and misspecification detection in amortized inference?

Q2. Practical effectiveness of non-oracle correction. Have the authors evaluated source-specific generator and first-passage-time errors when routing trajectories using the inferred partitions rather than oracle source labels? Such a comparison would help establish whether the proposed heterogeneity-detection procedure translates into meaningful improvements in scientific inference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.