acceptodds
Under review as a conference paper at ICLR 2027

Foundation Models for Variance Reduction in Stochastic Processes

Abstract

Estimating expectations of functionals of stochastic processes is a fundamental problem in many disciplines, from finance to life sciences and engineering. However, obtaining accurate estimates with Monte Carlo requires prohibitively many samples. We introduce foundation models for variance reduction that use simulator paths and black-box functional evaluations, without access to the explicit dynamics or task parameters. A pretrained model reads a small set of observed paths and outputs a task-specific variance-reduction controller, an importance sampling proposal or control variates integrand. This architecture leads to estimators which are unbiased by construction and adds little computational overhead. Because the models separate identification from estimation, the allocation of paths between the two provides a natural test-time scaling mechanism. We pretrain on a broad family of synthetic stochastic processes, to ensure strong downstream performance even in out-of-distribution applications, especially for importance sampling. We demonstrate the performance of the models on a variety of applications in finance and biology, showing that the proposed model provides significant variance reduction with matched wall-clock speed.

Primary area:
learning on time series and dynamical systems
Keywords:
foundation model, variance reduction, stochastic processes

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_ZwFiOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the problem of estimating expectations of functionals of stochastic processes. Obtaining accurate estimates with Monte Carlo is expensive but the cost can be reduced through use of a well-chosen importance sampling proposal or well-chosen control variates integrand. The paper introduces the use of foundation models for this purpose thereby reducing computational costs of Monte Carlo estimation whilst retaining unbiased estimates. The foundation models are tested on a range of applications in finance and biology and demonstrate significant variance reduction whilst matching wall-clock speed. Improvements in this area have important applications in many disciplines.

Critical strengths

Whilst task-specific trained models have previously been used to achieve Monte Carlo variance reduction in related ways, the use of reusable foundation models that adapt to new tasks without retraining or explicit task parameters appears to provide a significant and novel development. The construction of the models is clever and technically imaginative, and the experimental results show considerable promise for this new approach. This is a strong and well-written paper that lays the foundations for a new methodology with potential applications across a wide range of disciplines.

Critical weaknesses

It is unclear whether the models fully capture the breadth implied by the term ‘foundation model’.  Broader evaluations and stronger ablations would help determine the full generality of the approach. But in my view, for a paper that is introducing this new approach to the ML community, such limitations do not detract too significantly from the novelty or promise of the paper

I think it would be helpful to expand and summarise limitations of the method in one place (at present several caveats are dispersed across the paper). This would be helpful to the reader and, I believe, strengthen the paper rather than weaken it.

Questions for authors

(1) Can you expand on the way in which the engineering and stressed financial benchmarks are out of distribution relative to foundation model pre-training distribution and to what extent the resulting experiments support the claimed generality of the approach.

(2) Do you have suggestions for how users might select the pilot and estimation budgets?

(3) Are you able to describe what formal protection the importance-sampling mixture provides and/or how sensitive the results are to its mixing weight?

3 comments

We thank the reviewer for the efforts and time spent reviewing the manuscript and for the encouraging feedback and comments that help improve the manuscript.

Broader evaluations

This is a good point, we included a new set of experiments on extrapolating far outside the training domain in terms of dimension to see how the model performs on the engineering and finance tasks. We extrapolate in terms of both dimension and time and include the results in Figure 14 in the appendix. The results indicate the models still provide variance reduction on inference for higher, unseen dimensions with a gradual degredation as dimension increases.

Limitations section

Thank you for this suggestion, we included a new limitations section at the end of the main text. We specifically included some of the assumptions we required and commentary on how we can relax such assumptions in future work.

Out of distribution meaning

By out of distribution, we mean that the benchmarks use stochastic-process and function constructions not sampled by the pretraining generator. Specifically, we change the governing dynamics and the path dependence of the payoff from pretraining to test. Note that some of the components may overlap with pretraining: for example, pretraining includes cubic reaction terms, while the Allen–Cahn benchmark uses this with a nearest-neighbor spatial Laplacian which was never used in pretraining.

To give a few examples, we included the following table which describes the held out structure being tested for the benchmarks:

BenchmarkHeld-out structure
HestonCoupled price–variance dynamics with square-root volatility
AutocallsPayoffs determined by the first threshold crossing at specified observation times
MertonCompound-Poisson multiplicative jump dynamics
Allen–CahnCubic reaction coupled through a nearest-neighbor Laplacian
Lorenz–96Cyclic quadratic interactions
QueuesInteracting queue lengths constrained to remain nonnegative
and added more details in appendix A.5.

Finally, the real data experiments involve data that are not used during pretraining.

Choosing pilot and estimation budget

Thank you for this suggestion. As an example, we performed a scaling law analysis for the control variate foundation to predict the matched time variance ratio compared to MC as a function of the number of parameters of the model, the number of estimation paths and the number of pilot paths. At a fixed time budget, one can then predict the optimal number of estimation and pilot paths: while both numbers of paths need to generally increase as the time budget increases, the increase for the number of estimation paths is smoother than for the number of pilot paths, see appendix F.3.

IS mixing weight

This is a great suggestion, we included a figure (Figure 16 in appendix E.2) in the revision on how the α\alpha parameter plays a role in the variance reduction of the method. The results indicate that the model is fairly robust across different values of α\alpha. Note that the model construction is guaranteed to be unbiased for all α\alpha.

The defensive mixture provides an explicit bound on importance weights. Specifically, $ q_\alpha=\alpha p+(1-\alpha)q_\theta $ with $ 0<\alpha\leq1 $, where pp is the original sampling distribution, we have qα≥αpq_\alpha \geq\alpha p, and w=pqα≤1αw=\frac{p}{q_\alpha}\leq\frac1\alpha which gives us $ \mathbb E_{q_\alpha}[(Fw)^2] \leq\frac1\alpha \mathbb E_p[F^2]. $ This means that under our default α=0.2\alpha=0.2 level, the importance weight is bounded by 5 and preserves a finite second moment whenever the original payoff has one. This protects against extreme likelihood ratios caused by proposal mismatch.

We thank you again for your review, and we hope that our response provides additional clarification on the points in the review.

Thank you for addressing my comments in detail and for making corresponding changes to the pdf. I have no further questions nor need for further clarification and maintain my acceptance-recommendation score of 4.

We thank you again for your time and efforts in providing improvements to the manuscript!

Reviewer_es4aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes pretrained “foundation models” for Monte Carlo variance reduction in discretized diffusions and jump-diffusions. At test time, a small set of pilot paths and their functional values is encoded. In parallel, a function-probe encoder evaluates the functional on deterministic perturbations of the pilot paths. The two encoding are used to condition the decoder, which outputs a variance reduction method:

  1. An importance-sampling (IS) proposal over the driving noise.
  2. A control-variate (CV) integrand hth_t for a stochastic-integral control variate.

The estimator is unbiased regardless of model quality, since the model only recommends methods that affect the variance. This also allows the procedure to preserve other important properties of monte carlo estimators.

Critical strengths
  1. A well-posed, practically relevant problem formulation

Framing variance reduction as in-context identification from black-box pilot paths is a natural and, to my knowledge, new use of the pretrained/amortized paradigm. The design also keeps the property that matters most for Monte Carlo practitioners: unbiasedness that does not depend on model quality.

Critical weaknesses
  1. Inconsistent and underspecified method description

The CV description is inconsistent in the main body and the appendix - the estimator in Appendix B.4 (Eq. 6) differs significantly from the main text (section 3).

Moreover, the control variate process and the important sampling rely on facts that arrive from Brownian motion. For example, the important sampling procedure uses Girsanov’s theorem and the control variate method uses the Martingale representation theorem. Both of these assume a Brownian motion filtration and the density corresponding to Brownian motion. So, for the jump-diffusions the paper targets, the control variate can’t remove the variance coming from jumps, and the importance sampling never changes the jump law.

  1. Limited scale experiments

All pretraining and evaluation tasks have d≤16d \le 16 and T≤128T \le 128. The 60-dimensional alanine representation is truncated to its first 8 or 16 coordinates (A.4), and the engineering-task dimensions are not reported. Monte Carlo is most relevant, and variance reduction most valuable, in high dimension and over long horizons. These are also the regimes where IS weights degenerate and a fixed probe set covers fewer directions of the functional.

Questions for authors
  1. All tasks have d≤16d \le 16 and T≤128T \le 128. Can the model handle larger dd and TT without retraining, and how well does it perform there? This matters for calling it a foundation model - It should generalize to much larger dimensions and time scales.
5 comments

We thank the reviewer for the review and for recognizing that we propose "A well-posed, practically relevant problem formulation". We are grateful for the valuable feedback which helps improve the manuscript.

Implementation differences between appendix and main text

The description of the control variate method in the main text indeed abstracts away some of the details for the sake of clarity, while the appendix version provides all the details. In particular, the appendix version describes how the control variate head learned is made invariant to the scale of the applied function while the main-text estimator absorbs this normalization and gating mechanisms for stability into the effective control coefficients. We added a note in the main text to refer the reader to the appendix for more details (l281-282).

Accommodating jumps

We are happy to clarify this point.

The control variate model can handle jumps natively: it only requires zero-mean driving noise. This is done in some experiments such as the financial benchmark. We added a note in the main text (l189) and details in appendix (B.5).

For importance sampling, we did not experiment with controlling the jumps, and a Brownian-only martingale construction would indeed only pertain to the continuous noise component. However, our estimator does not require that the simulated process have a Brownian filtration. Jump increments can be constructed via inverse transform sampling, i.e. J=FJ−1(Φ(z)),z∼N(0,1)J=F_{J}^{-1}(\Phi(z)), z \sim \mathcal{N}(0,1). Therefore, we can propose a new distribution for the underlying Gaussian distribution generating the noise with importance weight evaluated in the Gaussian input space: Ep[f(J(z))]=Eq[f(J(z))dpdq]E_p[f(J(z))] = E_q[f(J(z)) \frac{\mathrm{d} p}{\mathrm{d} q}]. This means that the models are not intrinsically restricted to only Brownian increments. We have updated the manuscript with this discussion (l285).

Scalability to arbitrary time lengths and dimensions

The reviewer is correct that the model was trained for a maximum d=16d = 16 and T=128T = 128 (d=32d = 32 maximum including auxilary variables). However, we note that this kind of limitation is common for foundation models on numerical data: they are designed and trained to perform well up to a fixed dimensionality.

We note that, although the state dimension is at most 16, each discretized trajectory contains an effective dimension of up to d×T=2048d\times T=2048 scalar state values. The benchmarks include path-dependent payoffs and rare events, where Monte Carlo estimates can require substantial sampling.

The related question of the model's ability to generalize to higher dimensions is worth studying. Apart from pretraining directly at higher dd and TT, which nothing in our model prevents, one can use our current models at higher dd and TT by extrapolating: we split the target dimensions into 32 dimensional encoding and the full dimensional function value. For example, if we want a 128 dimensional process, we can estimate 4 separate groups of 32 with the function value of the 128 dimensional system. For the probe paths, we vary the particular 32 dimensional group and hold the other dimensions as their average. The model then infers the correction to the 32 dimensional subgroup. For time-grid extrapolation, we interpolate the learned temporal queries and evaluate on the finer grid, adjusting IS proposal shifts and CV noise increments consistently with the new time-step size. We then do this for each subgroup to sample the full system. We included a new figure (Figure 14) in the Appendix D that illustrates this zero-shot performance by scaling up to 1024 dimensions and 2048 time steps: this is 64×64 \times and 16×16 \times out of distribution extrapolation but the degradation in performance is slow which is notable since the model was never trained on these dimensions. For example, at d=1024d=1024 (which is a 64×64\times extrapolation factor), X-Large IS achieves an aggregate estimator vs Monte-Carlo variance ratio of 0.873, with a 95% confidence interval of [0.806,0.947], across the benchmarks.

We thank you again for your constructive comments and hope that these resolved the issues described in the review.

Thank you for the detailed response.

I still have some reservations about scalability. While I understand that foundation models often have limitations, T≤128T \leq 128 and d≤16d \leq 16 seem very restrictive. The proposed grouping approach is also not entirely convincing, as dependencies between different groups are lost when they are treated separately.

Moreover, the fixed function-probe evaluations seem problematic in high dimensions - points become very sparse and do not represent the function well anyore.

Overall, I quite like the idea and the method proposed. However, for this to be called a "foundation model", there needs to be clear evidence that it scales to larger dimensions. Therefore, I have decided to not change my score.

Best Regards, Reviewer es4a

We thank the reviewer for the response and the continued engagement. However, we would like emphasize that the revision provides direct evidence of extrapolation beyond the pretraining limits to higher, unseen dimensions. Our models trained with d≤16d \leq 16 and T≤128T \leq128 are evaluated on full systems with up to d=1,024d=1,024 and T=2,048T=2,048, see Figure 14 in Appendix D.

Although grouping limits the controller’s expressivity, the full coupled dynamics and full-dimensional payoffs are retained, including dependencies across groups. The probes provide complementary information to the pilot trajectories and their payoffs, and they are not intended to reconstruct the payoff throughout the high-dimensional domain. They provide targeted evaluations alongside full-dimensional pilot-path payoffs, so sparse probe coverage does not by itself preclude variance reduction.

This brings us to the numerical evidence: for models trained up to d≤16d \leq 16, at OOD d=128d=128, IS and CV achieve aggregate variance ratios of 0.669 [0.609, 0.732] and 0.798 [0.725, 0.879]; at OOD d=1024d=1024, these remain 0.873 [0.806, 0.947] and 0.820 [0.761, 0.875] with the numbers in bracket representing 95% confidence intervals. These results demonstrate statistically supported variance reduction far beyond the training dimensions (x64!), which we believe this evidence supports the central contribution of reusable pretrained models that transfer across stochastic estimation tasks and dimensions without task-specific weight updates.

Nevertheless, we are also open to adopting the more descriptive term “pretrained models for variance reduction” if this would communicate the contributions more clearly.

Reviewer_M7JtOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes pretrained “foundation models” for variance reduction in Monte Carlo estimation of functionals of stochastic processes. Given a small set of pilot paths and black-box functional evaluations, a pretrained model outputs either an importance-sampling (IS) proposal (a Gaussian mixture over the driving noise) or a control-variate (CV) integrand (a neural function of the state and running statistics). The estimators are unbiased by construction because pilot and estimation paths are independent, and the split between pilot and estimation paths provides test-time scaling. The models are pretrained on synthetic diffusion and jump-diffusion families, and evaluated on financial benchmarks (in- and out-of-distribution), engineering rare-event tasks (Lorenz-96, power grid, Duffing, queueing, Schlögl, Allen–Cahn, Müller–Brown), market-calibrated volatility surfaces, a limit-order-book simulator, and stochastic optimization (Newton's method, deep hedging, market making).

Critical strengths
  1. Amortizing variance reduction across tasks, without task parameters and without retraining, is a natural and useful direction. The “identification vs estimation” separation is clean, and the pilot/estimation split is a genuine test-time scaling mechanism.

  2. Breadth of evaluation. Finance, engineering, real volatility data, limit-order-book data, and stochastic optimization is a broad set of applications for a single method; the matched-time metric is the right one for a method whose selling point is wall-clock efficiency.

  3. Real-world experiments.The market-calibrated volatility and limit-order-book experiments go beyond the usual synthetic benchmarks, even if the gains are modest.

Critical weaknesses
  1. The “black-box” claim is too strong. Section 3.1 states the assumptions: for IS, “we assume that we can inject driving noise into the process”; for CV, “that we can access the noise increments.” Many simulators of interest (Gillespie-type chemical kinetics, molecular dynamics with constraints, commercial pricing engines) do not expose the driving noise or its density. The abstract’s “without access to the explicit dynamics or task parameters” is technically true but misleading: the method does need access to the noise, which is a core part of the dynamics. The paper should state this assumption prominently, characterize which simulators satisfy it, and discuss the failure mode when they do not.
  2. No uncertainty quantification. The figures and tables appear to be point estimates; there are no error bars, no seed counts, and no repeated-run variability. For claims of 5–10× variance reduction, this is a serious omission. Several entries in Table 3 are identical to three decimals across model sizes (e.g., Real torsion 5D: 0.820 for Small, Medium, and Large; Synthetic 2D: 0.855 for all three), which suggests either that the model size does not affect the result or that the table is not based on independent runs. The authors should report independent runs with confidence intervals and explain the identical entries.
Questions for authors

see weaknesses

2 comments

We would like to thank the reviewer for the time and efforts in reviewing and for the helpful feedback provided to improve the manuscript. We address individual concerns below.

Black-box claim

We have updated the phrasing of the abstract and of the manuscript to better reflect the requirements of our models and included a more detailed discussion in the limitations section. Specifically, IS requires controllable driving innovations and likelihood ratios in that innovation space; CV requires access to conditionally centered innovations; and, the probe encoder requires evaluating the function on constructed paths. These requirements are satisfied by simulators implemented with explicit random inputs. For IS, the likelihood ratio is evaluated in the driving-noise space, so a tractable transition or path density for the simulated process is unnecessary.

Concerning the remark on simulators, we would like to note the following: while exposing the driving noise or densities is not standard for most simulators, these quantities are actually computed inside the simulator and can be made available to a user. In that sense, there are no fundamental obstacles to exposing this for the purposes of the proposed model.

Uncertainty quantification

Thank you for pointing this out, we have updated the manuscript to include figures and tables with confidence intervals. Specifically, we address two sources of randomness in our experiments. First, the randomness in evaluation: we re-ran all most of our evaluations 10 times and report the mean and 95% confidence intervals. The confidence intervals are small enough to not change the conclusions from the original reported figures. More precisely, we updated:

  • Figure 2a: importance-sampling and control-variate validation variance ratios across training checkpoints.
  • Figure 2b: control-variate performance versus path count on financial benchmarks.
  • Figure 3: control-variate performance versus runtime under severe stress.
  • Table 1: importance-sampling matched-time variance ratios on real local-volatility pricing tasks.
  • Figure 4a: importance-sampling performance on engineering tasks.
  • Figure 4b: autocallable hedging results.
  • Figure 6: Müller–Brown estimator variance.

as well as figures 8 to 14 and tables 2 to 7 in the appendix, where tables 6 and 7 now correspond to the Alaline benchmark.

Note that in the original Alaline table some cells had the same numbers due to rounding, eg for Real Torsion 5D: small achieved 0.82056, medium 0.82085 and large 0.82079, which were (incorrectly) rounded to 0.820.

Second, we address randomness in training dynamics: we retrained the "small" CV/IS models for 5 seeds and show that performance changes very little (Figure 15 in Appendix E.1, reporting standard deviation across five training seeds).

Thank you again for the constructive feedback on the manuscript, and we hope that we addressed the concerns with the submission.

Thank you for the feedback, I'll keep my score unchanged.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Changes

We would like to again thank all the reviewers for their time and efforts in contributing valuable feedback to improve the manuscript. We thank the reviewers for noting the work is "well-posed" and "practically relevant"; evaluated on a wide range of experiments and real-world conditions; and a "significant and novel development".

We want to outline the changes we made in response to the feedback provided:

  1. Uncertainty quantification on the experiments

We included uncertainty bands across the figures and tables in the manuscript and reproduced the training across several seeds, as recommended by Reviewer M7Jt.

  1. New evaluations on extrapolation out of time and dimension distributions

As suggested by Reviewer es4a, we test how well the models can extrapolate well out side the pretraining time lengths and dimensions and provide new figures showing this performance, which further reinforces the foundation model perspective.

  1. Prediction of optimal Nest,NpilotN_{est}, N_{pilot} for a given wall-clock evaluation budget

We included new experiments to show how the performance of the model and its inference time scale with NestN_{est} and NpilotN_{pilot}. This informs the user when faced with a practical problem, as suggested by Reviewer ZwFi.

  1. Dedicated limitations section

We included a new limitations section where we describe the major limitations of the method in a unified fashion, as suggested by Reviewer ZwFi.

All the textual changes are in a blue font color in the revised manuscript.

We hope these new experiments and modifications answer remaining questions, and we are happy to discuss further should there be additional topics for discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.