acceptodds
Under review as a conference paper at ICLR 2027

Convergence of Few-Step Importance Weights: A Cheap, Reference-Free Certificate

Abstract

Diffusion and flow models increasingly use few solver steps as proposals for importance sampling, where the model likelihood enters the final estimator and numerical convergence becomes part of the sampling problem. Few-step likelihoods are only approximate, and confirming their accuracy would require exactly the expensive many-step reference that few-step sampling was meant to avoid. The standard cheap diagnostic, effective sample size, cannot fill this gap: it cancels a constant likelihood shift exactly, and its first-order response to a fluctuating error is governed by a squared-weight alignment that differs from the observable covariance governing estimation bias, so an observable can be biased while ESS is unchanged. We introduce a cheap, reference-free, a-posteriori certificate for the coupled log-weight discretization error. For a fixed source draw the log weight admits a structured expansion in the step size, so extrapolation from a few additional short solves reconstructs the error without invoking the reference; the certificate is provably sound once the step size falls below an explicit threshold, tightens as the budget grows, and uses no hand-tuned constants. The same expansion yields a same-sample truncation diagnostic, and an elementary bridge lemma converts per-sample error bars into a bound on the self-normalized estimate. Across controlled synthetic flows, one-step average-velocity maps, and energy-based molecular samplers, the certificate attains high empirical coverage, substantially reduces importance-sampling error on a real average-velocity map, and reaches near-reference-quality estimates at a fraction of the reference cost on a molecular system, whereas ESS-based budget rules accept severely biased budgets and cannot rank candidates at the degeneracy floor.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
importance sampling, diffusion models, flow matching, few-step sampling, likelihood estimation, Richardson extrapolation, effective sample size, Boltzmann generators, uncertainty quantification

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_EAmgOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the reliability of importance weights when diffusion/flow (ODE-based) models are used as few-step proposals for self-normalized importance sampling (SNIS), such as in Boltzmann generators. The paper separates three sources of error:

  • O1: discretization error in the log weight
  • O2: error between the likelihood returned by the implementation and the exact likelihood of the deployed K-step map.
  • O3: the resulting bias in SNIS estimates.

Experiments include synthetic GMMs, iDEM samplers, MeanFlow, FALCON.

Critical strengths
  1. The problem is important. Few-step flow/diffusion models are attractive for sampling, but their likelihoods usually no reliable for importance sampling.
  2. The authors split the question into O1/O2/O3 parts. It makes clear that discretization error, likelihood error, and SNIS bias are analyze seperately.
  3. The result for O1 is clean. The Richardson-based certificate is reference-free, and Theorem 2 gives explicit conditions for soundness and asymptotic tightness. Proposition 3 gives a useful explanation of the limitation of ESS.
Critical weaknesses
  1. The strongest theory is not for the error that actually causes SNIS bias. The theorem which is theoretically proved, is mainly for O1. As the paper notes, if the exact density of the discrete K-step map is used, SNIS remains consistent. O2 is the error that can actually bias SNIS. However, O2 is handled mainly through the heuristic diagnostic, without a guarantee. Lemma 4 gives a bias bound if per-sample O2 error bounds are already known, but the paper does not provide a practical way to obtain such bounds. This makes the paper feel like two separate contributions. The experiments also blur this distinction. I also have no idea to estimate/improve the bounds of existing models, e.g., Falcon, Meanflow.

  2. For FALCON, the goal is not simply to approximate the continuous ODE with fewer steps. The model learns large, invertible jumps along the flow so that the discrete map itself has a valid likelihood. From this view, closeness to the continuous-flow limit (O1) is not the main question. More relevant questions are: a). How much invertibility error does the learned map have? b). How does this error affect the likelihood and SNIS bias? c). How close is the few-step proposal to the target distribution? These seem closer to the actual failure modes of few-step learned maps. In the FALCON experiment, the paper instead introduces an artificial log-determinant truncation. This tests the proposed O2 diagnostic, but it does not test the main approximation error of FALCON itself, such as the reported cycle residual.

  3. The related-work discussion should include several papers: (a) F2D2 (Ai et al., 2025), which jointly distills trajectories and cumulative divergence for few-step sampling and likelihood evaluation. (b) Flow Map Matching (Boffi, Albergo & Vanden-Eijnden, 2024), which is closely related to the flow-map models considered here. (c) Free-form Flows (Draxler et al., 2024), which studies likelihood estimation for approximately invertible learned maps. Authors should introduce the connection and difference with these work.

  4. The error analysis leads to little practical algorithmic improvement. The main correction is to subtract the signed cKc_K​ from the log weight. However, this only applies to the first-order log-determinant approximation and works only when the series converges. (This setting is also somewhat artificial.) The analysis also does not show how to train few-step maps with smaller O2O_2 error. The practical contribution would be stronger with a theory-guided improvement, such as: (a) an extrapolated or bias-corrected weight; (2) improvement of Falcon training objective.

[1] Ai X, He Y K, Gu A, et al. Joint distillation for fast likelihood evaluation and sampling in flow-based models[C]//International Conference on Learning Representations. 2026, 2026: 44940-44977. [2] Boffi N M, Albergo M S, Vanden-Eijnden E. Flow map matching with stochastic interpolants: A mathematical framework for consistency models[J]. arXiv preprint arXiv:2406.07507, 2024. [3] Draxler F, Sorrenson P, Zimmermann L, et al. Free-form flows: Make any architecture a normalizing flow[C]//International Conference on Artificial Intelligence and Statistics. PMLR, 2024: 2197-2205.

Questions for authors
  1. Proposition 3(i): a constant likelihood shift does not always “bias nothing.” This is true for self-normalized estimates, but not for unnormalized IS estimates such as partition functions or free-energy differences.

The same as Weaknesses.

Reviewer_kG9WOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Few-step flow and diffusion samplers are increasingly used as importance-sampling proposals, so their approximate likelihoods are used in the weights weights. The paper proposes a per-sample Richardson extrapolation from solves at K, 2K and 4K steps. It estimates the gap between the K-step and continuous-flow log-weights without needing an expensive reference solve, and the paper proves it is sound once the step size is below a threshold h*. It also shows that the effective sample size (ESS) is not always a good diagnostic, so an observable can be biased while ESS is unchanged. the paper proposes c_K, the next term of the log-det series, as an empirical diagnostic and correction for a first-order truncated log-det likelihood. It also gives a lemma (Lemma 4) that would bound the SNIS bias if valid per-sample likelihood-error bounds were available, though the paper does not seem to supply one. Experiments cover synthetic GMM flows, iDEM samplers (DW4, LJ13), MeanFlow, and a FALCON reimplementation.

Critical strengths

The paper targets an important question of how accurate are the few-step likelihoods used in importance sampling with flow and diffusion models. It decomposes the error into distinct components (discretization drift, deployed-likelihood error, and downstream estimation bias). It provides a cheap, reference-free method to estimate how far the K-step log-weight is from its continuous-flow limit. The paper also reports several limitations and failure modes clearly.

Critical weaknesses

The paper motivates its contribution by the need for a “trustworthy and numerically converged model likelihood” in few-step importance sampling. However, convergence is not what importance sampling requires. Any proposal, converged or not, yields consistent SNIS estimates as long as the weights use the exact density of the deployed sampler. The authors acknowledge this, noting that SNIS is consistent at every K and that the deployed-likelihood error is what corrupts SNIS. Yet the abstract, introduction, and the paper’s only theorem-backed contribution focus on convergence of the K-step log-weight to its continuous-flow limit. The critical gap is between the computed likelihood and the exact density of the sampler actually used.

Perhaps the biggest weakness is that clarity problems make the results hard to interpret.

Questions for authors

My main concern is clarity. The paper contains potentially useful ideas, but the presentation makes it very difficult to understand and assess the technical contributions. Key terms (“certificate,” “deployed likelihood,” h*) are not clearly defined. Several estimators and constants are introduced only in appendices or not at all (e.g., C_rich). The framing emphasizes convergence to the continuous-flow limit rather than the error that actually biases importance sampling. I believe major revisions would be needed before the contributions can be properly evaluated.

Questions:

  • SNIS is consistent at every K when the exact discrete log-det is used. What decision should the K-step drift E_K then drive?

  • Please formally define the deployed likelihood q^dep for each case you claim to cover (truncated series, Hutchinson trace, approximate inverse). In each case, is δ deterministic or random, and which of your results apply?

  • Please define “certificate.” Given that the condition h ≤ h* cannot be checked, in what sense is C_sc a certificate rather than an asymptotically exact a-posteriori error estimator?

  • Please define C_rich and its “2σ” allowance. Please also provide a table showing which estimator (C_sc, C₃, C_gap, C_full, c_K), with which constants, produced each reported result. It is currently difficult to tell which results are supported by Theorem 2.

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

The paper provides a posterior error estimate for few step samplers used for importance weights. Few step samplers typically use approximation for the log determinant needed for the log-likelihood computation and this approximation makes the likelihood inexact. This error can have a negative effect on the importance weighted estimates and the ESS estimate.

It separates three sources of error: (1) the discretization error of the K-step log-weight relative to its continuous-flow limit, (2) the error from computing the log-likelihood with a cheap approximation of the Jacobian log-determinant (e.g. a truncated series, Hutchinson trace estimates, or an approximately invertible map), which makes the likelihood inexact for the generated samples, (3) the resulting bias in the self-normalized importance sampling estimates.

To solve (1), the paper bounds the error using three cheap evaluations with an increasing number of solver steps (K, 2K, 4K). The authors propose to estimate the first two terms of an asymptotic expansion of the log-weight by solving a 2x2 linear system on the differences of the evaluation.

For (2), they propose a diagnostics c_K, the next term of the log-determinant series, which costs one extra trace per step. It detects the truncation error and can be subtracted to correct it where the series converges. A bridge lemma converts per-sample bounds on (2) into a bound on (3).

On synthetic flows, MeanFlow, FALCON and molecular samplers, the estimates reliably cover the true error, and the c_K correction substantially reduces importance-sampling bias where ESS fails to detect it.

Critical strengths
  • First theoretical error estimate on importance sampling estimates for few step samplers.
  • Could potentially be very useful for the generative modelling community
  • Results look convincing
Critical weaknesses
  • Writing is very hard to follow and needs to be improved (Often symbols are not introduced or intuition is missing on what they are. Same symbol is used for different quantitites which is confusing (e.g. C)
  • Missing comparisons to baselines such as Pareto-smoothed importance sampling or standard numerical-analysis baselines (plain Richardson, embedded error estimates).
Questions for authors
  • Please try to introduce and motivate symbols and try to avoid overlap in the notation.
  • We need 7K evaluations to estimate the error for K steps. Would it there not be more efficient to directly solve the likelihood for 7K steps as this would significantly reduce the error (without estimate)? In which applications does it really pay of to invest these amount of steps?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.