acceptodds
Under review as a conference paper at ICLR 2027

AdaRecurr: State-Adaptive Recurrent Computation for Neural PDE Simulations

Abstract

Neural simulators can substantially reduce the cost of PDE simulation, but autoregressive rollouts typically apply the same amount of inference computation to every physical transition even though different states may benefit differently from additional computation. Some states require only limited recurrent refinement, whereas others continue to improve with deeper refinement. This raises a natural question: can computation be allocated dynamically according to the current physical state as the trajectory evolves? We introduce AdaRecurr, a state-adaptive recurrent computation framework that turns recurrent depth from a global test-time parameter shared across an entire rollout into an online computation budget selected separately for each physical transition. AdaRecurr builds on a pretrained recurrent neural PDE simulator that can perform a variable number of recurrent refinement iterations for each prediction. Keeping this simulator frozen, AdaRecurr observes the current physical field together with the latent refinement already executed and sequentially decides whether to STOP or CONTINUE. Continuing reuses the current latent state and performs further recurrent refinement, while stopping accepts the current prediction and advances the physical rollout. The controller is trained on self-generated autoregressive trajectories to balance prediction error against additional recurrent computation. Across three neural PDE benchmarks, AdaRecurr reduces trajectory-level relative error by 3.6–15.3% over fixed recurrent-depth inference at comparable mean recurrent depth. On viscous Burgers, states receiving more computation exhibit a 36.2% higher mean aggregate estimated shock strength than states receiving less computation. In a controlled compute-demand heterogeneity experiment, we further observe larger gains from adaptive allocation when recurrent-computation requirements vary more strongly across physical states. These results show that test-time recurrent computation can be allocated dynamically along evolving physical trajectories rather than using a single fixed recurrent depth for the entire rollout.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural PDE simulation, adaptive computation, recurrent-depth inference, test-time computation, scientific machine learning, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_6S8aOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

AdaRecurr adds a learned stopping controller to a pretrained, frozen recurrent neural PDE simulator. For each physical transition, the controller observes the current physical field and the latent refinement checkpoints already computed, then decides whether to stop or continue recurrent refinement. The actor and critic are trained with PPO on model-generated autoregressive rollouts, using a reward that balances next-state prediction error against additional recurrent iterations. On Burgers, NS2D, and SWE2D, the paper reports 3.6–15.3% lower trajectory-level relative L2 error than nearby fixed-depth operating points at comparable mean recurrent depth. It also examines the association between allocated computation and estimated shock strength, and evaluates controlled populations with different levels of compute-demand heterogeneity.

Critical strengths

The paper provides evidence that state-dependent recurrent-depth allocation can improve neural PDE rollout accuracy at comparable mean recurrent depth. The evaluation covers three benchmarks and includes fixed-depth, random budget-matched, and heuristic adaptive baselines. This is a useful empirical demonstration that the choice of where to allocate recurrent computation can matter, beyond simply varying depth without conditioning on the state. Keeping the pretrained backbone frozen also makes the controller a potentially useful addition to an existing recurrent simulator.

Critical weaknesses
  1. The incremental contribution and reproducibility of the controller are insufficiently established. Learned stopping is an established idea, and the paper cites relevant approaches such as ACT and PonderNet. Conditioning on physical states and executed latent traces, reusing latent computation, and training on model-generated rollouts are sensible design choices. However, I am not convinced that the paper establishes a sufficiently substantive adaptation to the PDE setting beyond applying learned stopping to an existing recurrent simulator. The return and advantage are computed within each physical transition, so training on autoregressive rollouts should not be interpreted as directly optimizing a trajectory-wide reinforcement-learning return. Comparisons with existing learned stopping approaches, or targeted ablations, would help establish which design choices are essential. Furthermore, Section 3.2 and Appendix A.4 describe the controller modules without specifying their concrete architectures, dimensions, or parameter counts. This omission affects both reproducibility and the assessment of controller overhead.

  2. The practical accuracy–cost benefit is not established once the added controller is included. Mean recurrent depth measures backbone refinement and omits controller inference, repeated stopping decisions, and any additional execution overhead. At some lower-budget operating points, increasing the depth of a simpler baseline appears capable of reaching similar accuracy; the relevant question is whether AdaRecurr remains advantageous at matched end-to-end runtime. Controller training also requires repeated forward inference through the frozen backbone, together with backpropagation and parameter updates for the controller. The paper does not quantify these additional training costs or establish how readily the controller transfers when simulation conditions change. This leaves the amortized benefit uncertain, particularly if a transferable backbone nevertheless requires a newly trained controller. Finally, the NS2D result in Figure 2 appears to improve on the fixed K = 32 baseline despite using much less mean recurrent depth. This is possible, but the mechanism and robustness of the improvement need clarification to determine whether the result reflects computation allocation, avoidance of harmful additional refinement, or changes in error propagation during rollout.

Questions for authors
  1. Contribution and implementation: What substantive adaptation to autoregressive PDE simulation distinguishes this controller from existing learned stopping approaches? Please specify the architectures and dimensions of the spatial, latent, and sequence encoders, the actor/critic, and their parameter counts. Can comparisons or targeted ablations establish the importance of the proposed conditioning and training choices?
  2. Actual computational cost: When controller inference and all stopping decisions are included, does AdaRecurr retain an accuracy–cost advantage over fixed-depth and heuristic adaptive baselines at matched end-to-end runtime? Please provide runtime measurements and quantify the additional controller-training cost, including frozen-backbone forward inference and controller optimization.
  3. Transfer and amortization: Under a change in simulation conditions that the frozen backbone can accommodate, can the trained controller be reused directly, or does it need adaptation or retraining? Please clarify the demonstrated scope and, if adaptation is required, its cost and implications for amortizing the training overhead.
  4. NS2D behavior: What explains the NS2D operating point in Figure 2 that improves on fixed K = 32? With the same input state and latent initialization, can further refinement increase next-state error, or does the advantage mainly emerge from altered autoregressive error propagation? Please provide the numerical fixed-depth results and variability across paired repeated evaluations, alongside the controller-seed results already reported in Table 7.
Reviewer_2p7qOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose AdaRecurr, a state-adaptive recurrent computation framework that decides whether to stop or refine the iteration in a recurrent-depth module. They validate their proposed method on 3 different PDEs: Burgers, Navier-Stokes 2d, and Shallow-water 2d.

Critical strengths

They develop a strategy that addresses the limitations of the RecurrSim framework. The paper is well written and easy to follow, and the authors position their contribution clearly within the existing literature.

Critical weaknesses

The results are modest. The authors introduce an adaptive controller that must be trained after RecurrSim pre-training to achieve performance comparable to a heuristic adaptive controller, which requires no training and is widely used in other recurrent-depth architectures, such as DEQs.

Questions for authors

In a multi-physics scenario where the number of iterations matters, do you think the choice of model would have a greater impact on performance? For example, if you're working with equations of varying complexity, such as the heat equation, Burgers' equation, and the Navier–Stokes equations, you might be able to solve the heat equation with relatively few iterations, while Burgers' equation requires a moderate number, and Navier–Stokes requires significantly more. In this case, would your model make a bigger difference in computational efficiency and convergence?

Reviewer_AuQsOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

AdaRecurr uses a PPO controller to choose how much recurrent computation a frozen neural PDE simulator performs at each physical step. The controller uses the predicted state and intermediate latent representations to decide when to stop. Experiments on three PDEs show lower rollout error at comparable average recurrent depths.

Critical strengths

The motivation is clear. The method is straightforward to integrate with a pretrained recurrent simulator, and training on self-generated rollouts is a sensible choice. The experiments are reasonably broad and show encouraging results.

Critical weaknesses
  1. It is not yet clear how much of the improvement comes from adapting to the physical state. The examples in Figure 3 show recurrent depth decreasing over time, which could also be captured by a time-dependent schedule. A comparison with such a schedule would help establish the value of state-dependent decisions.

  2. Although training uses self-generated rollouts, returns terminate at each physical transition. Future errors are therefore not directly credited to earlier stopping decisions; the implications of this local objective for longer-horizon accuracy remain unclear.

  3. The evidence linking heterogeneity to larger gains is not conclusive. In Table 1, the improvement increases from LOW to HIGH relative to the lower fixed-depth reference, but decreases relative to the upper reference. The interpretation therefore depends on the comparison being made.

Questions for authors
  1. How does AdaRecurr compare with a time-only schedule? Does the latent refinement trace provide a clear advantage over using the current physical state alone?

  2. Why does credit assignment terminate at each physical transition? Could this miss benefits of additional refinement that emerge only in later predictions?

  3. How should the different trends against the two fixed-depth references in Table 1 be interpreted? Does HIGH still show a larger benefit when the methods are compared at similar mean recurrent depths?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.