acceptodds
Under review as a conference paper at ICLR 2027

Efficient Rare-Event Estimation Can Distort Conditional Trajectories

Abstract

Rare-event proposals can supply both probability estimates and trajectories used to explain how an event occurs. On a controlled diffusion, learning raises estimation efficiency, reduces estimated global conditional fidelity, and improves route-conditioned hitting-time accuracy on the same evaluation paths. This combination occurs in all five original models and all five new fits trained through the deployed control cap, with evaluation on fresh paths. RareFlow characterizes this separation through the exact factorization , where is event discovery and is the reciprocal of one plus the forward conditional discrepancy. A mechanism decomposition separates route allocation from within-route distortion. An independent guide-family comparison shows that nearly matched route shares can conceal timing bias, while an exact conditional-law ceiling limits what route reallocation alone can repair. A finite-state control exposes severe overestimation of this ceiling when an important submode receives sparse proposal coverage. A second planar landscape reproduces the qualitative discovery–fidelity separation. Together, these results show that improved probability estimation does not by itself certify faithful unweighted event trajectories, and that full conditional fidelity and observable-specific accuracy should be evaluated separately.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Rare-event generation, Conditional trajectory fidelity, Mechanism-frequency decomposition, Stopped-path likelihood, Guided diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7oosOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether improving the probability estimation efficiency of a learned rare-event proposal also improves the fidelity of its unweighted event-conditioned trajectories. The authors introduce RareFlow, an auditing framework that decomposes overall estimation efficiency S into event discovery H and conditional path fidelity F through the exact identity S = HF. A mechanism decomposition further separates conditional mismatch into route-allocation and within-route distortion components. The framework also derives a route-reallocation ceiling and an observable bias bound. Controlled experiments on a modified Muller-Brown potential and a three-hole landscape examine three interventions: route feedback, residual learning, and guide refinement. The central finding is that residual learning consistently improves probability estimation efficiency and route-conditioned hitting-time accuracy while reducing global conditional fidelity, and this combination persists across five original models and five new fits trained through the deployed control cap.

Critical strengths
  • The problem formulation is original and well-motivated. The observation that estimation efficiency and conditional path fidelity are distinct evaluation targets, and that improving one can degrade the other, addresses a genuine gap in the rare-event sampling literature. The S = HF factorization provides a clean and interpretable lens for diagnosing this separation, and the fact that all three quantities can be estimated from a single weighted proposal sample makes the framework practically deployable.

  • The mechanism decomposition is technically elegant and diagnostically valuable. Splitting conditional mismatch into between-route and within-route components, then deriving the fidelity-optimal route allocation and its exact ceiling, gives practitioners actionable guidance on where to focus improvement efforts. The finding that route reallocation offers almost no headroom while within-route distortion dominates is a strong result that directly shapes future proposal design.

  • The experimental design is thorough and commendably honest. The three controlled interventions isolate distinct sources of efficiency change, development and evaluation data are strictly separated, and the consistent-cap follow-up rules out a plausible alternative explanation for the trade-off. The finite-state hidden-submode stress test, which exposes severe overestimation of the route ceiling under sparse coverage, demonstrates intellectual honesty about the limitations of the diagnostic itself.

Critical weaknesses
  • The experiments are confined to two planar synthetic landscapes with supplied mechanism labels and exact likelihood ratios. While the generality of the qualitative finding is supported by replication across two geometries and multiple noise levels, the absence of any higher-dimensional or application-driven case study leaves the practical scope of the framework unclear. Extending to molecular dynamics or autonomous-system failure sampling would substantially broaden the impact.

  • The relationship between global fidelity loss and observable-specific improvement lacks deeper mechanistic explanation. The paper demonstrates that hitting-time accuracy improves while forward chi-squared fidelity declines, and connects this via the Cauchy-Schwarz bound, but does not explain why this particular observable benefits. Understanding which classes of observables tend to improve versus degrade under learned proposals would make the finding more actionable.

Questions for authors
  • Could the framework be applied to proposals with estimated rather than exact likelihood ratios, and how sensitive are the fidelity estimates to ratio-estimation error in practice?

  • Is the efficiency-fidelity trade-off specific to the cross-entropy training objective used for residual learning, or would alternative objectives such as variance-based or backward-KL formulations exhibit a different trade-off profile?

Reviewer_F4kaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies importance-sampling proposals that make rare events easier to observe. Correct importance weights can recover the target event probability, but unweighted successful trajectories still follow the proposal's conditional distribution. Hence accurate probability estimates do not guarantee that these trajectories represent how the event occurs. Interpreting unweighted trajectories as representative examples of failure mechanisms can therefore be misleading.

The authors propose RareFlow, a rare-event auditing framework built around S=HFS=HF. This identity explains how increased event discovery can compensate for worse conditional fidelity. The framework also decomposes conditional mismatch into incorrect allocation across mechanisms and distortion within mechanisms. For example, a proposal could generate the correct proportions of upper-route and lower-route transitions while misrepresenting the trajectories within each route.

The main experiments use a 2-D modified Müller–Brown landscape with two transition routes. Across both sets of five learned-model comparisons, learning improves probability-estimation efficiency and route-conditioned hitting-time accuracy while reducing estimated global conditional fidelity.

Critical strengths
  • The paper identifies an important limitation of efficiency-based evaluation. It distinguishes probability estimation, fidelity of the full conditional trajectory distribution, and accuracy of a chosen observable.

  • The framework provides a useful breakdown of proposal quality. Beyond reporting overall efficiency, RareFlow separates event discovery from conditional fidelity and distinguishes between-mechanism mismatch from within-mechanism distortion. This helps explain why matching mechanism frequencies alone may not produce representative trajectories and what route reallocation can repair.

Critical weaknesses
  • The writing and presentation are difficult to follow and messy. In particular, the abstract and introduction introduce technical terminology and experimental details before providing enough intuition. Specific examples include:

    • Unexplained experimental labels in the abstract. Such as "five original models" and "five new fits" appear before readers understand what models are being trained.
    • The abstract describes FF as the "reciprocal of one plus the forward conditional discrepancy". A plain-language explanation would help readers understand its mathematical intuition.
    • The introduction uses "rare-event proposals" without first giving an intuitive account of how a proposal changes the sampling process.
    • The key quantities SS, HH, and FF are poorly explained in the introduction. The authors introduce the identity S=HFS=HF without clearly defining what these quantities represent or how they are calculated. For example, it is unclear whether HH refers to the probability of reaching the rare event, Q(E)Q(E), or simply the number of successful trajectories. Similarly, FF and SS are not adequately explained. The authors should also clarify that SS measures statistical efficiency per attempted path, not computational speed.
    • Route-allocation terminology is poorly introduced. The authors use technical terms such as "fidelity-optimal route allocation," "route-only ceiling," and "allocation-regret identity" without adequately explaining their meaning or intuition.
  • The identity S^=H^F^\hat{S}=\hat{H}\hat{F} holds by construction, so it cannot be used to verify adequate sample coverage. The consistency result does not tell practitioners whether a particular sample is large enough to trust the estimates.

  • The experimental scope is limited. The experiments involve two planar synthetic landscapes and exact path-density ratios for the analysis. A higher-dimensional or application-driven experiment would help the audience to understand the framework's practical utility.

Questions for authors
  • Can RareFlow be applied to sampling from generative diffusion models? For example, could it be used to audit guided sampling conditioned on a rare output event? What assumptions are required to compute the necessary path-density ratios.
Reviewer_ZJE5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a factorization for the quality of a rare-event proposal distribution. Two sources can weaken a propoal: a weak event-discovery probability Q(E) and a shape mismatch for the event-conditional distribution of samples. Quantifying the latter two with H and F respectively, the overall distribution mismatch between proposal and target is derived as S = HF.

The paper goes on to estimate these individual metrics using samples and then uses the methodology to evaluate multiple proposal distributions.

Overall the contribution seems too narrow for acceptance.

Critical strengths

1/ The derivations of factorization and estimation seem correct but are mostly straight-forward.

Critical weaknesses

The paper's presentation is very dense and not easily accessible for readers.

1/ The paper does not introduce the problem space clearly. While initially it seems like a paper broadly around importance sampling it later 1/ turns out to be fairly narrow in scope focusing on transition path sampling, which should be more clearly introduced.

2/ The introduction of the factorization S=HF is very ad-hoc and then only explained in (5) and it's unclear whether this is just a shorthand for a conceptual split of a mathematical relationship that's proposed. H and F are not defined on first mention beyond the names these variables stand for

3/ Some notational aspects are not defined

  • chi2 in (2) - the actual chi2 distribution or some distributional distance measure?
  • P ≪ Q - is this a statement about supports of P and Q?
Questions for authors

1/ are the proposal distributions standard proposals used within relevant fields? Why were these three chosen ? 2/ can you clarify the notation ? 3/ can the setup be generalized to broader conditional sampling problems other than trajectory sampling?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.