acceptodds
Under review as a conference paper at ICLR 2027

Neurons as Worlds: Learning Adaptive Mesoscopic Dynamics from Particles

Abstract

Coarse particle group representations can reduce the communication cost of particle simulation, but changing particle groups disconnects their current contents from the accumulated predictive context. The challenge is to reorganize the representation without restarting its history. We introduce Neuronal World Dynamics (NWD), which treats regrouping as a state transition of recurrent particle group representations called neuronal worlds. As the particle groups splits and merges, physical moments are recomputed from the unchanged particles, while compact memories are copied or combined locally. Shared recurrent dynamics then refine the inherited context using each successor's current observations, without replaying trajectories. The theoretical analysis identifies when a larger decoder cannot resolve coarse-state ambiguity and quantifies the trade-off between suppressing reorganization and retaining learned dynamics. Across four particle benchmarks, NWD achieves the lowest mean 100-step position NRMSE in all five settings, with 51.8–52.6% reductions over the recurrent-token baseline on N-body. On N-body, NWD reduces this error by 10.7–14.5% relative to fixed-cardinality variants and 51.8–52.6% relative to a recurrent-token baseline; memory ablations provide additional evidence for retained context and its association with particle support.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
neuronal representation, particle simulation, particle dynamics, particle representation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_jjMEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Neuronal World Dynamics (NWD), an adaptive framework for particle simulation that maintains recurrent memory across particle-group split–merge operations. The method combines adaptive grouping, memory inheritance, and physics-aware constraints to improve long-term prediction. Theoretical analyses examine coarse-state ambiguity and reorganization-induced errors. Experiments on four benchmarks demonstrate improved long-horizon accuracy over existing baselines, with ablation studies supporting the importance of persistent memory.

Critical strengths
  1. The paper introduces an interesting approach that preserves predictive context across particle-group split–merge operations by separating physical state updates from recurrent memory inheritance.

  2. NWD achieves consistent improvements over existing baselines, particularly on N-body simulations. The ablation studies further demonstrate the importance of persistent memory, while physical diagnostics and computational analyses strengthen the evaluation.

Critical weaknesses
  1. The experiments do not clearly isolate the benefit of memory inheritance across split–merge events. Removing the handoff mechanism causes only a 2.8% performance degradation, raising questions about whether the improvements primarily come from persistent memory and adaptive grouping.

2.The split–merge scheduler relies on predefined rules rather than explicitly learned event decisions. Moreover, the public benchmark experiments use only one training seed and a fixed training budget, limiting evidence of robustness and generalization.

The paper presents an interesting and well-motivated framework for adaptive particle simulation, with convincing long-horizon improvements on the N-body benchmarks. However, the experimental evidence does not yet sufficiently isolate the effectiveness of the proposed memory inheritance mechanism, which constitutes the central novelty of the work. The limited multi-seed validation on public benchmarks also weakens the generality of the empirical conclusions.

Questions for authors
  1. Can the authors provide stronger evidence that memory inheritance specifically improves predictions at split–merge events? For example, an event-centered analysis comparing prediction discontinuities and subsequent rollout errors with and without memory inheritance would be informative. This could help distinguish the contribution of the proposed handoff mechanism from the more general advantage of maintaining recurrent memory.

  2. How sensitive is NWD to the lifecycle scheduler and event-attenuation settings? In particular, how does performance change with different split–merge frequencies, cardinality targets, and event attenuation coefficients? Given that the standalone N-body configuration sets the event-step learned-update coefficient to zero, it would be useful to clarify how much of the improvement is attributable to memory inheritance versus event suppression and adaptive allocation.

  3. Do the improvements on Dam2D and Sand2D remain consistent across multiple independent training seeds and longer training schedules? The existing comparisons use a common number of optimization updates but do not match computational runtime or establish convergence. Additional evidence would help determine whether the hybrid's advantages reflect a robust architectural improvement.

6 comments
AuthorsOpenReview ↗

Rebuttal response (1/2)

Thank you for your careful assessment and for identifying the controls needed to distinguish inheritance, event attenuation, and public-benchmark robustness. Your comments helped us make these comparisons more direct.

Weakness #1 / Question #1. Memory inheritance beyond recurrence

Thank you for identifying this distinction. Your feedback helped us formulate the control that directly tests inheritance. The original “no handoff” contrast changes event-step output attenuation; both arms retain memory inheritance. It therefore does not isolate successor-memory transfer, as clarified in the revised Table 3 caption. The new control resets only successor memories at split/merge events while retaining recurrence elsewhere, the scheduler rule, attenuation, training budget, and checkpoint-selection rule. Matched retraining at γ = 0.25 uses three fitted seeds and all 300 test trajectories. Inheritance improves both metrics in every seed:

Training condition100-step NRMSE, mean (seed SD)Position MSE
Reset successor memories0.277 (0.010)0.0215
Inherit successor memories0.266 (0.012)0.0197

The earlier 21.0% contrast tests the value of persistent memory; this successor-only control tests whether that memory should be retained when the representation reorganizes. Its benefit therefore cannot be attributed merely to adding recurrence.

We separately report two fixed-weight diagnostics. In the output-discontinuity diagnostic, a matched parent-topology sham defines the reference learned acceleration. The relative output jump falls from 0.025 to 0.020 for splits and from 0.040 to 0.030 for merges, i.e., reductions of 20% and 25%. The jump is the RMS difference in pre-attenuation learned acceleration relative to the sham, normalized by sham RMS with a training-derived floor.

The subsequent-error diagnostic below forks the same checkpoint and pre-event state, applies the same topology transformation, and resets only successor memory before the common message/memory update. Its nonzero event coefficient, γ = 0.25, avoids immediate equality caused by suppressing the event-step learned output:

EventPre-event paired gapEvent-step NRMSE, inherit/resetOffset-10 NRMSE, inherit/reset
Split00.0724 / 0.07450.0926 / 0.0986
Merge00.1136 / 0.12380.1392 / 0.1468

The pre-event gap is zero because the branches start from the same state. At offset 10, inheritance lowers error by 6.1% for splits and 5.2% for merges. Repeated successor resets increase 100-step NRMSE from 0.263 to 0.285; equivalently, inheritance lowers error by 7.7%. These diagnostics complement the retrained successor-only control by measuring event discontinuity and later rollout consequences directly.

Weakness #2 / Question #2. Scheduler sensitivity and event attenuation

We appreciate your careful examination of the scheduler. The full model retains its mean advantage when event attenuation is completely disabled. On the same 32 trajectories and three frozen fitted seeds, the one-factor sensitivity results are:

FactorValues tested100-step NRMSE, same order
Event-check interval2 / 4 / 80.2373 / 0.2315 / 0.2372
Target cardinality16 / 20 / 320.2307 / 0.2315 / 0.2365
Event coefficient γ0 / 0.25 / 10.2315 / 0.2323 / 0.2431
Split geometryPCA / mass-median / longest-axis0.2315 / 0.2312 / 0.2316

Fixed-K gives 0.2711. With γ = 1, NWD still lowers mean error by 10.3%. Thus, the full adaptive model's advantage persists without event-step suppression. Separately, the successor-reset experiment isolates inheritance at fixed γ = 0.25. These controls distinguish adaptive allocation, inherited context, and output attenuation. At γ = 0, memory refresh, known forcing, and physical integration continue; only the event-step learned output is suppressed. A learned gate over legal events gives 0.2290 under the same event/world budget, showing that the inheritance interface also supports learned decisions.

AuthorsOpenReview ↗

Rebuttal response (2/2)

Weakness #2 / Question #3. Seed replication and longer matched training

The mean gains persist beyond the original single-seed, 2,000-update comparison under both update-matched and runtime-matched training. Both architectures receive symmetric validation-based schedule selection and three fitted seeds, evaluated on all 25 Dam2D and 30 Sand2D test trajectories. Parentheses report seed SD; reductions are computed from the displayed mean MSEs.

Dataset / budgetEnlarged GNS MSEGNS+NWD MSEReduction
Dam2D / equal time, 8,366 s0.3409 (0.0384)0.2513 (0.0508)26.3%
Sand2D / equal time, 8,167 s0.0386 (0.0023)0.0344 (0.0018)10.9%
Dam2D / extended time, 12,548 s0.3341 (0.0258)0.1955 (0.0236)41.5%
Sand2D / extended time, 12,251 s0.0378 (0.0045)0.0340 (0.0027)10.1%

The fixed-update gain is favorable in all three Sand2D seeds and two of three Dam2D seeds; the Dam2D equal-time gain is also favorable in two of three. The extended comparison increases the equal-time budget by 50% and uses three fresh seeds per architecture. The mean advantage remains 41.5% on Dam2D and 10.1% on Sand2D. Together, the update-matched, runtime-matched, and extended-runtime comparisons show that the mean advantage persists across seed replication and substantially longer training.

Thank you again for the constructive questions that helped strengthen our work and manuscript.

Thank you for your detailed responses and additional experiments. I appreciate the authors' efforts, which have addressed many of my initial concerns. I have two remaining questions: Does the benefit of memory inheritance remain consistent across different split–merge frequencies? Could the authors further clarify the relatively large variability across training seeds, particularly on Dam2D?

Reviewer_iu8aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Neuronal World Dynamics (NWD), an adaptive mesoscopic representation for learning particle dynamics. Specifically, particles are organized into neuronal worlds whose memberships and number of active units can change during simulation through routing and split/merge operations. To preserve historical information during these changes, the method transfers recurrent memories between neuronal worlds while recomputing their physical properties from the associated particles. The proposed method also incorporates force–torque constraints when mapping neuronal-world predictions to particle-level dynamics. Experiments demonstrate improved long-horizon prediction accuracy over the evaluated baselines across several particle simulation benchmarks.

Critical strengths
  • The proposed method explicitly considers the preservation of historical information during adaptive particle regrouping. The memory inheritance mechanism allows recurrent states to be transferred across split/merge operations, maintaining temporal context as the representation changes.

  • The method incorporates physical constraints through carefully designed particle-level output mechanisms, aiming to preserve aggregate force and torque consistency.

Critical weaknesses
  1. The presentation of the proposed method is difficult to follow, with several important concepts and implementation details insufficiently explained in the main text. For example, Equation (1) introduces neuronal worlds, particle memberships, and the router without first clearly explaining their meanings or relationships to the original particle states. Although these concepts are formally defined in Section 3.2, a more intuitive introduction would help readers understand the overall framework. Similarly, Section 3.3 introduces terms such as support residuals, age, cooldown, and cardinality without adequately explaining how they affect split/merge decisions. While the memory inheritance rules are provided, the actual split/merge procedure and its execution within a simulation step are only briefly described. Important implementation details are deferred to Appendix B, making it difficult to understand the complete computational pipeline from the main text alone.

  2. The experimental results do not sufficiently isolate the contribution of the proposed split/merge mechanism and memory inheritance strategy. According to Table 3, persistent memory provides a substantial benefit, with full NWD achieving a 21.0% reduction in 100-step position NRMSE compared with its memoryless variant. In contrast, the improvement over the variant without memory handoff is only 2.8%. The authors also acknowledge that the existing ablations do not fully isolate the benefit of memory transfer specifically at split/merge events. Although comparisons with fixed-cardinality variants demonstrate some benefits from adaptive grouping, additional experiments would help establish the contribution of the proposed memory inheritance mechanism. In particular, how frequently do split/merge events actually occur during a rollout, and how do they affect prediction accuracy immediately before and after an event? More targeted ablations and event-level analyses would strengthen the evidence supporting this central contribution.

  3. The comparisons with closely related methods could be more comprehensive. In Section 5.1 (Lines 336–337), the authors state that STT-R is a compact recurrent-token baseline rather than a reproduction of WorldParticle. However, the differences between STT-R and the original WorldParticle are not sufficiently clarified. Since WorldParticle is closely related to the proposed method in its use of particle aggregation and mesoscopic representations, a direct comparison would help establish the advantages of NWD. Could the authors clarify the architectural differences between STT-R and WorldParticle and explain why the original WorldParticle is not included in the experiments?

  4. The evaluation of generalization to unseen particle counts and physical scenarios is limited. The N-body training data contain particle counts ranging from 256 to 2,048, while the main robustness experiments in Section 5.5 primarily evaluate models within this range. Although the appendix includes two out-of-distribution examples with 4,096 and 8,192 particles, these are individual trajectories rather than a systematic evaluation across multiple unseen scenarios. A more comprehensive study with multiple trajectories, different initial conditions, and potentially unseen physical configurations would provide stronger evidence of the model's generalization capabilities. Given that scalability and adaptability are important motivations for the proposed mesoscopic representation, these results would also be valuable to include more prominently in the main text.

  5. The qualitative evaluation of the predicted dynamics is relatively limited. Figure 4 presents representative snapshots for two benchmark scenarios, while Appendix Figure 9 provides additional N-body comparisons across several scenarios and baselines. However, these static visualizations do not fully characterize the temporal evolution of the predicted dynamics. Since the proposed method focuses on long-horizon particle simulation, rollout videos or sequences of representative frames would help readers assess the realism and temporal consistency of the predictions, particularly during particle separation, merging, and other complex interactions. More comprehensive visual comparisons across different physical systems and baselines, including potential failure cases, would strengthen the qualitative evaluation.

Questions for authors

Please refer to the weaknesses.

5 comments
AuthorsOpenReview ↗

Rebuttal response (1/2)

Thank you for your detailed and constructive feedback on both the presentation and the empirical evidence. We have clarified the method in place and added controls for memory inheritance, a direct WorldParticle comparison, systematic generalization tests, and a temporal Sand2D comparison.

Weakness #1. Definitions and the computational pipeline

Thank you for identifying where the presentation was difficult to follow. Your feedback helped us make the pipeline more explicit. We have clarified the concepts before Eq. (1) and the scheduler terms in Sec. 3.3. A neuronal world is a weighted, possibly overlapping particle group with physical summaries and recurrent memory; the router produces its particle-membership weights. Each step first encodes and routes the particles and measures the world moments. Any legal split or merge then transfers memory and remeasures successor moments, after which the model applies shared interactions, refreshes memory, decodes the update, and integrates particle motion. Thus, the event reorganizes the representation before physical prediction, leaving particle positions and velocities unchanged. The revised Sec. 3.3 defines support residual as within-world feature dispersion, age as time since world creation, cooldown as temporary event ineligibility, and cardinality as the active-world count. Residual and geometry guide proposals; age and cooldown determine eligibility; cardinality controls the allocation budget. Splits copy the parent's memory, and merges use mass-weighted averaging. Each successor's current observations then enter the shared recurrent update. Appendix B.2 provides the complete one-step operator.

Weakness #2. Isolating memory inheritance and adaptive grouping

We appreciate this insightful point. It helped us design a more direct test of inheritance beyond recurrence. The original 21.0% memoryless contrast measures persistent recurrence. The original “no handoff” contrast changes event-step output attenuation while retaining inheritance in both arms, as now clarified in the Table 3 caption. The new control directly isolates inheritance: it resets only successor memories at split/merge events while preserving recurrence elsewhere. With matched retraining, event coefficient γ = 0.25, validation-based selection, three fitted seeds, and all 300 test trajectories:

Training condition100-step NRMSE, mean (seed SD)Position MSE
Reset successor memories0.277 (0.010)0.02150
Inherit successor memories0.266 (0.012)0.01973

Inheritance lowers NRMSE by 3.7% and MSE by 8.2%, with both gains present in all three fitted seeds. This tests the additional value of carrying context through reorganization when both models already have recurrence. Default-controller events occur 18.70–21.53 times per 100 steps.

The event-centered diagnostic forks the same checkpoint and pre-event state, applies the same split/merge, and changes only successor memory before the common message/memory update. At γ = 0.25, we measure the learned-output jump relative to a matched parent-topology sham and then continue the rollout:

EventPre-event paired gapinherit/resetEvent-step NRMSEOffset-10 NRMSE
Split00.020 / 0.0250.07247 / 0.072550.09263 / 0.09357
Merge00.030 / 0.0400.11367 / 0.113790.13927 / 0.14068

The NRMSE columns also list inherit/reset. Inheritance reduces the output discontinuity by 20% for splits and 25% for merges, and the paired prediction error remains lower ten steps later. The zero pre-event gap denotes equality between branches, not zero prediction error. Output jump is the RMS change in pre-attenuation learned acceleration relative to the sham, normalized by sham RMS with a training-derived floor.

Separately, Fixed-K retains routing and recurrence while disabling cardinality-changing events; the submitted five-seed comparison gives a 10.9% reduction for NWD. Together, these controls distinguish adaptive grouping, persistent recurrence, and the event-specific value of inherited state.

AuthorsOpenReview ↗

Rebuttal response (2/2)

Weakness #3. Direct comparisons with STT-R and WorldParticle

The original STT-R comparison tests a compact recurrent-token alternative: learned-query pooling, fixed-slot GRU updates, and an attention/MLP readout. WorldParticle has a different architecture: continuous-convolution particle features, hierarchical super-token aggregation, and a RoPE-based particle decoder. Its super-tokens are constructed within each prediction step, whereas NWD carries particle-supported recurrent state through cross-step split/merge events.

We now include an adaptation built from WorldParticle's official implementation, retaining continuous convolution, iterative k-partite aggregation, RoPE decoding, and its predictor–corrector structure. The reduced configuration uses a 72-dimensional continuous-convolution embedding, three aggregation iterations, and three decoder layers. Both methods use three fitted seeds, 2,000 training updates, validation-based checkpoint selection, and all 300 N-body test trajectories. At 100 steps, WorldParticle gives 0.4299 ± 0.0071 NRMSE and NWD gives 0.2814 ± 0.0230: NWD is 34.5% lower, with the gain present in every seed. This supplies the requested official-architecture comparison under the stated budget alongside the original controlled recurrence comparison.

Weakness #4. Systematic particle-count and initial-condition generalization

Thank you for encouraging a systematic evaluation across unseen particle counts and initial conditions. The new evaluation extends the two illustrative count-OOD examples to multiple unseen initial conditions. Frozen models trained on 256–2,048 particles are evaluated at 4,096 and 8,192 particles, covering collapse/capture, cluster merger, flyby, and tidal split:

Test countDistinct trajectoriesNWD reduction vs Fixed-Kvs STT-R
4,0963224.0%50.3%
8,1921624.5%45.9%

At each unseen count, the mean 100-step error favors NWD over both comparators in each of the three fitted seeds. These cohorts cover all four scenario families and establish particle-count and initial-condition transfer under the same governing physics, without count-specific finetuning.

Weakness #5. Temporal qualitative comparison

Thank you for suggesting a sequence of frames; this helped us make the temporal behavior easier to assess. Appendix G now provides five successive Sand2D frames comparing reference particles, enlarged GNS, and the NWD prediction as contact develops into deformation and spreading. The ordered frames and per-frame coordinate RMSE make both the temporal progression and the remaining differences in the advancing front and particle distribution directly inspectable. This sequence complements the existing N-body gallery with a temporal comparison on a granular-material benchmark. The quantitative conclusions are based on the complete test cohorts, while the sequence illustrates how the predictions evolve over time.

Overall, our revision makes the pipeline clearer and adds targeted quantitative controls, systematic generalization tests, and the Sand2D temporal comparison.

Thank you again for helping us improve the presentation and empirical support.

Area_Chair_T28yOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer iu8a,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

Reviewer_6WJbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

Neuronal World Dynamics (NWD) reframes coarse‑grained particle simulation as learning a recurrent state‑transition system over mesoscopic units called "neuronal worlds." Each world carries physical moments (mass, momentum, angular momentum) plus a fixed‑width learned memory. The key novelty is that worlds can split and merge during rollout -- successor moments are recomputed from unchanged particles, compact memories are copied or averaged, and shared recurrent dynamics refine the inherited context using new observations, without replaying trajectories or re‑encoding from scratch. The paper theoretically establishes that (1) a larger decoder cannot fully resolve coarse‑state ambiguity, (2) preserving aggregate physics leaves room to correct trajectories via a projected residual, and (3) reallocation introduces representation drift that must be damped. Empirically, NWD reduces 100‑step position NRMSE by 51.8‑52.6% over STT‑R on N‑body, and by 10.7‑14.5% over fixed‑cardinality variants on N‑body. For bonded dynamics the improvement is smaller (~8.6%). The hybrid GNS+NWD model achieves 37.0% MSE reduction on Dam2D and 11.3% on Sand2D at matched parameter budget.

Critical strengths
  1. Novel conceptual framing: mesoscopic units as learnable state‑transition worlds. The idea of treating coarse‑grained particle groups as recurrent state units with explicit split/merge lifecycle operations, rather than fixed partitions, is a fresh contribution to the learned simulation literature. It connects graph neural network simulation to hierarchical representation learning in a principled way.
  2. Theoretical analysis accompanies the method. The paper does not just propose an architecture -- it proves capacity bounds on coarse‑state ambiguity, shows how projected residuals preserve aggregate physics, and characterizes the representation drift from reallocation. This theory‑to‑experiment loop is rare and strengthens the paper's contributions.
  3. Broad benchmark coverage across four physics domains. Evaluation spans N‑body, bonded dynamics, free‑surface fluids, and granular materials. N‑body and bonded dynamics use in‑house DyWorlds benchmarks, while Dam2D and Sand2D are established public benchmarks. The results show consistent improvement over Fixed‑K, STT‑R, UPT, and Transolver baselines, which supports the generality claim.
  4. Careful ablation design. The component ablation (Table 3) isolates the contribution of persistent memory, shared context, local lift, handoff, and graph‑lift reallocation, and uses paired wins with Bonferroni intervals. This level of statistical rigor is above typical ICLR papers.
Critical weaknesses
  1. The advantage over fixed‑cardinality baselines is modest on fluids and granular materials. While standalone NWD achieves meaningful gains on N‑body, the hybrid GNS+NWD achieve 37.0% and 11.3% MSE reduction on Dam2D and Sand2D respectively over enlarged GNS‑A.The core value proposition -- adaptive mesoscopic dynamics -- needs to show larger standalone gains on problems where fixed partition is clearly inadequate.
  2. Computational cost analysis is incomplete. Wall‑clock latency comparisons between NWD, GNS‑PF and STT‑R on N‑body are reported. However, controlled matched‑particle‑count comparison against UPT and Transolver is missing, and the isolated overhead introduced purely by split‑merge bookkeeping logic is not quantified. Without these measurements, the practical benefit over simpler baselines is less clear.
  3. The split/merge heuristics are hand‑designed rather than learned. The paper argues that discrete reallocation can coexist with continuous dynamics, but the actual decision rules (covariance‑based split, nearest lead‑particle merge) are simple heuristics. There is no analysis of whether learned split/merge gates would improve results, nor whether the current heuristics fail on specific physical regimes (e.g., highly turbulent flows, dense particle collisions).
Questions for authors
  1. The split/merge rules are deterministic: splits along leading covariance, merges centers by nearest lead particle. How sensitive is the result to these heuristics? Have you tried learned split/merge decisions, or do simpler rules (e.g., equal‑mass split, nearest‑neighbor merge) work as well?
  2. The hybrid GNS+NWD model adds only 137,354 parameters (8.6% over GNS‑A). The branch‑removal ablation shows 5.98‑12.07% NRMSE loss on Sand2D across 20‑100 steps. Is this magnitude of improvement consistent across all benchmarks? Note that the 14.0% NRMSE reduction in Table 3 originates from shuffled‑memory ablation rather than correction projection ablation.
  3. The theoretical analysis shows that reallocation introduces drift (Eq. 14). In practice, how often do splits/merges occur per 100‑step rollout, and does the drift accumulate over longer trajectories (e.g., 500 or 1000 steps)?
  4. The particle‑to‑world mapping uses a sparsity constraint (q=8 candidate units, r=2 nonzeros per row). What happens when a world's effective particle support becomes too small? Is there a garbage collection mechanism, or do underpopulated worlds persist throughout rollout?
3 comments

Thank you for your careful reading and for recognizing the representation-level contribution and theoretical analysis.

Weakness #1. Standalone gains and adaptive cardinality

The standalone comparison directly tests the value of adaptive cardinality. Fixed-K retains NWD's encoder, learned routing, physical moments, and recurrent memory, while disabling cardinality-changing events (submitted Appendix C.4). NWD reduces 100-step NRMSE by 10.9% across five reported seeds. At 4,096/8,192 particles, the reductions increase to 24.0%/24.5% over 32/16 distinct trajectories, with favorable mean differences in all three fitted seeds.

We additionally compare standalone NWD with its matched Fixed-K counterpart on Sand2D, using three seeds and 8,000 updates. The eight contact-heavy trajectories are selected from reference trajectories before evaluating predictions.

100-step position MSEFixed-KStandalone NWDReduction
All 30 test trajectories0.2120.16920.3%
Eight highest-contact trajectories0.2840.21225.4%

This comparison demonstrates a benefit from adaptive allocation without a GNS backbone, with the larger reduction occurring on contact-heavy trajectories.

Weakness #2. Matched computational cost

We agree that the practical cost deserves a clear answer. We now report matched-particle-count latency against the official-code UPT and Transolver adaptations, together with their accuracy. Accuracy uses all 300 test trajectories; latency uses 1,024 particles on an RTX 4090 in FP32 with the corresponding checkpoints and optimized implementations.

Method100-step NRMSE, mean (seed SD)Latency (ms/step)
UPT adaptation0.5120 (0.0329)4.01
Transolver adaptation0.3172 (0.0282)7.65
NWD0.2814 (0.0230)12.52

NWD reduces mean NRMSE by 45.0%/11.3% relative to UPT/Transolver. These measurements quantify the accuracy–latency trade-off. In the same optimized timing comparison, Fixed-K takes 18.37 ms/step. Separate component instrumentation attributes 3.4% of the instrumented NWD time to split/merge bookkeeping, with instrumented and reference outputs agreeing exactly. This fraction measures bookkeeping within the instrumented execution; the end-to-end timings above report the cost of the complete architectures.

Weakness #3 / Question #1. Scheduler choices and demanding regimes

We appreciate this thoughtful suggestion, which helped us examine how much the gains depend on the scheduler. The benefit persists under alternative geometric rules and a learned legal-event gate. A merge proposes the nearest pair of world centers and then applies age/cooldown checks; it does not select nearest lead particles. We clarify this order in Sec. 3.3 and Appendix B.3. On 32 trajectories and three frozen fitted seeds:

Split/event rule100-step NRMSE
Covariance split0.2315
Mass-median split0.2312
Longest-axis split0.2316
Learned legal-event gate0.2290
Fixed-K comparator0.2711

The geometric alternatives give closely matched mean errors, while the learned gate gives a further improvement under the same event/world budget. Thus, the benefit does not depend on the particular covariance-based split rule, and the inheritance interface also supports learned decisions. On the seven highest-vorticity Dam2D test trajectories, the hybrid retains 15.0% lower MSE than enlarged GNS, complementing the contact-heavy Sand2D comparison above. These subsets directly probe demanding interactions within the evaluated physical systems.

Question #2. Branch contribution and the 14.0% contrast

Your question helps us clarify what each comparison measures. The active branch contribution is task-dependent. The reported 5.98–12.07% values are position-MSE reductions, measured on four reused trajectories at fixed hybrid weights (submitted Appendix E.3):

Full hybrid versus branch off20 steps50 steps100 steps
Dam2D0.41%0.61%0.02%
Sand2D5.98%10.58%12.07%

Branch removal retains the backbone learned jointly with the correction and measures that correction's contribution at fixed weights. Independently training the parameter-matched enlarged GNS tests the complete architecture and gives 22.4%/12.0% lower MSE on Dam2D/Sand2D after 8,000 updates. N-body and Bonded use standalone NWD and have no separate GNS branch to remove. The 14.0% contrast concerns the association between carried memory and particle support; it is not a correction-projection ablation.

(to be continued)

AuthorsOpenReview ↗

rebuttal response (2/2)

Question #3. Event frequency and long-rollout drift

Thank you for raising the question of accumulated error; it motivated us to examine both individual events and longer rollouts. The default controller executes 18.70–21.53 events per 100 steps. Regrouping itself leaves particle positions and velocities unchanged. Equation (14) separately accounts for the change in the learned update and the update omitted by attenuation. We evaluate the resulting prediction error on 16 long trajectories per fitted seed:

HorizonNWD position NRMSESTT-R position NRMSE
5000.65410.7732
1,0000.68580.8055

NWD retains 15.4%/14.9% lower position NRMSE at 500/1,000 steps despite repeated reorganization. All evaluated trajectories remain finite through 1,000 steps. The momentum-error norm is also 99.1% lower than STT-R at 1,000 steps, with favorable comparisons in every fitted seed. Matched event-centered diagnostics reduce the relative output jump by 20% for splits and 25% for merges. Together with the extended rollouts, these results address both the immediate effect of reorganization and its consequences over longer trajectories.

Question #4. Underpopulated worlds

We appreciate this careful question, which helps us clarify the behavior of worlds with small support. Small worlds can persist: NWD does not use a general garbage collector. The constraints q = 8 and r = 2 bound candidate and nonzero memberships per particle, rather than the number of particles per world. Coverage repair keeps active-world masses and centroids defined; no empty supports occurred in 190,884 world-step observations. Small worlds continue their recurrent updates and may gain support through routing or participate in legal merges. Split-legality checks prevent invalid children. The safeguard ensures nonempty support rather than a permanent minimum effective-support size.

Thank you again for helping us sharpen these comparisons, improving our work and manuscript.

Area_Chair_T28yOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer 6WJb,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.