acceptodds
Under review as a conference paper at ICLR 2027

When Does a Neural Sampler Actually Need the Target Score? Isolating the Role of Langevin Parameterisation

Abstract

Nearly every effective neural sampler for unnormalised targets injects the target score into its drift, a design known as Langevin parameterisation. he2025notrick showed that removing it causes severe mode collapse and concluded that an explicit score term is required; their comparison, however, changes the score term and the velocity parameterisation at once. We re-run it inside a single architecture family, varying only the score's role. The term is indeed causal—coverage drops from modes to —but it works just as well as an ordinary input feature, and the capacity does not substitute for it, so what matters is the availability of the gradient rather than the residual structure around it. Most unexpectedly, within this setting the score is needed only during a specific interval of the sampling trajectory: restricting it to is numerically identical to having no score at all, recovers full coverage, and the last alone is worthless again. We give a partial account. The optimal control provably converges to the target score as , which explains the design but deepens the puzzle; a transport bound then shows the cost of redistributing mass inside a window of width diverges as , providing a control-cost mechanism consistent with the failure of late windows, and varying the mode separation moves the threshold consistently with the implied . The failure of early windows remains unexplained, and we report it as open rather than attach one of the two intuitive stories our own experiments falsify.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
target score, diffusion samplers, stochastic optimal control, mode collapse

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_YrqDOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper extend the analyze by He et al 2025 on the role of Langevin preconditioning. It extends the conclusion in 2 ways: first, they confirm the conclusion of He et al 2025, and also show that inputing the score to the network can also help avoid mode collpasing. Second, they explore what chunk of the trajectory requires the score.

Critical strengths

This is an interesting analysis of the role of Langevin preconditioning. I appreciate the potential effort the author took to analyze these kind of techniques. However, I am sorry that I might have to say that I do not believe this paper has substantive strength that support acceptance.

Critical weaknesses
  1. The conclusion is superfacial. The paper only anaysis the PIS style loss, and only analyze different way of inputing the score (and different time to input the score). There are many other more interesting things that the paper failed to find and discuss (or just leave it open). It is hard to see what this conclusion guide us for a betterr neural sampler.

  2. The theory is not informative and not relevant. Prop 1 is so obvious and not means anything. Also, the score agrees with the optimal control does not mean the score is important at the point they agree. Think about diffusion model! Most of them has bad score agreement at data end, but it does not matter.

  3. The experiments are not representative and not convincing. Only GMM and WM is analyzed. The window test also did not fully conduct on WM. We know different target can have very different behiour (also in different number of dims). However, the paper did not discuss or test any of the conclusion can transfer to more complicated / higher dimension target.

  4. The scope is limited to old methods PIS. (See my question)

Questions for authors

There are many recent methods that do not need preconditions. But they implicitly learning to regress the score at some point (flow sampler, adjoint sampler, BMS, etc). How you explain them? When these score are effective or not?

Reviewer_4NFVOpenReview ↗
Rating
1(−2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(−1)
Read review
Paper summary

This paper studies how and when neural samplers benefit from the target score. Within the controlled experiments, supplying the score as an input feature works as well as an additive term, while its usefulness depends strongly on when it is available along the trajectory. The theory offers a partial explanation for the failure of short terminal windows.

Critical strengths
  • Although I was unfamiliar with neural samplers, I could follow the motivation and main results. The paper provides an accessible introduction to the problem.

  • I found the controlled experiments valuable, particularly those separating score availability, parameterisation, and timing.

  • Propositions 1 and 2 are clear, and their proofs are straightforward to check. I did not notice an obvious flaw. While I cannot confidently assess their novelty, I believe they seem to support the proposed interpretation within the stated limitations.

Critical weaknesses
  • It would have been helpful to discuss the practical implications of these findings. It is not clear to me what guidance they offer for sampler design.

  • Both target families are evaluated at d=16d=16. Since reference evaluation remains tractable in higher dimensions, experiments at higher dimenions (say d=64,128d=64, 128), or with more modes, would help assess the generality of the window finding.

Questions for authors
  • What practical guidance do these findings offer for sampler design, particularly for choosing a score window on a new target?

  • Does the window phenomenon persist in higher dimensions or with more modes?

Reviewer_Cn5aOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper looks at what it is about the score that improves the performance of neural sampling techniques. They attempt to answer what aspect of the score actually matters, and when during sampling does introducing the score actually help.

Critical strengths
  • honest scoping of claims with open questions stated rather than papered over.
  • good experimental design
  • one architecture family, only the score's role varies, exact ground truth.
  • Time-gating result is novel and surprising.
Critical weaknesses
  • The early-window failure is explicitly unexplained (the authors are upfront about this).
  • Prop. 2 is a necessary condition only; it doesn't predict where the threshold falls.
  • it assumes the control must do the moving. Remark 2 admits the Brownian noise moves samples comparable distances (RMS ≈ 5.7 at w = 0.5 vs inter-mode R ≈ 8.5). The "noise is undirected" argument is plausible but not proven.
  • The scaling evidence is thin: four radii, one excluded as degenerate, one only a lower bound. The authors concede it can't distinguish R² from e.g. R^1.8.
  • "Covered" = ≥ 1% of samples in a mode. This hides proportions: a mode that should get 30% counts as covered at 1.1%.
  • On the GMM, the successful 8/8 arms still have mode-weight L1 ≈ 0.8 (scale 0–2) vs 1.56 for collapsed arms. So "full coverage" still means substantially wrong weights; the headline 8/8 vs 2/8 looks cleaner than it is. While coverage is meant as a collapse detector it might be worth making this sort of thing more explicit.
  • On Many-Well, exact target samples score only 16/32 under the 1% rule, so 32/32 means overspreading into tiny modes. The authors acknowledge this and re-run with a target-aware criterion, but the headline tables use the raw count.

Also I noticed that there are Identical failure numbers / possible optimisation artifacts. Many failing arms (no score, 16× capacity, score-early, score-late) give L1 = 1.562 to 3–4 decimals, i.e. the same collapsed solution. With NN_2 zero-initialised, did NN_2 ever move away from zero in the score-early arm? If not, the early-window "mystery" could be an optimisation issue.

Questions for authors

what are the learned NN_2(t) values or magnitude of the score contribution per window?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.