Graph Mamba Operator: A Latent Simulator for Interacting Particle Systems
Karn Tiwari, Niladri Dutta, N M Anoop Krishnan, Prathosh A P
Abstract
Modeling interacting dynamical systems requires capturing spatial interactions alongside long-range temporal dependencies. Graph neural networks (GNNs) provide a natural representation but typically rely on autoregressive rollouts and treat spatial and temporal dynamics separately, leading to error accumulation over long horizons. Existing approaches also focus on local interactions and short temporal contexts, limiting their ability to capture multi-hop dependencies and global structure. We introduce the Graph Mamba Operator (GraMO), a latent-space simulator that integrates state-space models with graph-based interaction learning. In contrast to prior work that sequences nodes or applies spatial and temporal updates in separate stages, GraMO couples graph-based interactions and temporal state updates within a single recurrence. The update is linear in the latent state, with input-dependent coefficients that adapt across regimes. We evaluate GraMO on N-body systems, motion capture, and robotics datasets, achieving the lowest error across benchmarks and the largest gains in long-horizon prediction.
AI Impact Assessments
(1 models)Scientific Impact Assessment: Graph Mamba Operator (GraMO)
1. Core Contribution
GraMO introduces a latent-space simulator that unifies graph-based spatial message passing with selective state-space model (SSM) temporal evolution within a single recurrence step. The key architectural novelty is Equation (10/12): at each timestep, the latent state is simultaneously propagated across graph edges via the normalized Laplacian , mixed along a memory axis via a learnable matrix , temporally modulated via HiPPO-initialized decay , and injected with input-dependent features via . This contrasts with prior approaches that either (a) apply spatial and temporal updates in separate stages, or (b) serialize graph nodes into artificial sequences for SSM processing (e.g., GraphMamba). The effective time-varying propagator can be interpreted as an input-conditioned Koopman operator on graphs, where evolution is linear in the lifted latent state but the operator adapts to changing dynamics.
2. Methodological Rigor
Theoretical analysis is a notable strength. The paper provides: (i) a formal Koopman interpretation (Theorem C.2) showing linearity in the latent state with input-dependent operators; (ii) stability guarantees via spectral analysis of (Lemma C.3) and bounded multi-step Jacobians (Proposition C.10); (iii) a graph ARMA representation (Theorem 4.1/C.12) showing the unrolled dynamics admit an absolutely convergent moving-average form with time-varying coefficients; and (iv) permutation equivariance (Proposition C.4). These results are mathematically sound and provide meaningful guarantees — particularly the stability analysis connecting HiPPO initialization with graph spectral properties.
Experimental evaluation is comprehensive, spanning six distinct benchmarks: N-body simulations (3 variants), robotics (3 environments), motion capture (2 sequences), MD17 molecular dynamics (8 molecules), and protein dynamics. The paper compares against 10+ baselines spanning GNNs, equivariant networks, Koopman methods, diffusion models, and graph-SSM approaches. Error bars from 5 runs are reported throughout. Results are consistently strong: GraMO achieves lowest error across nearly all settings, with particularly large gains on long-horizon prediction (e.g., 67.3% FMSE reduction on MoCap vs. NS-EGNN).
Ablation studies systematically isolate contributions of temporal modeling, bidirectional SSM, selectivity mechanism, and HiPPO initialization. The temporal extrapolation experiment (Figure 3) is particularly compelling, showing GraMO maintains stable predictions well beyond the training horizon while Koopman baselines collapse.
Potential concerns: The zero-shot generalization test (training on N∈{5,7,8}, testing on N=9) is limited to a single setting. The paper would benefit from more extensive generalization experiments across different system sizes and interaction types. The claim of "largest gains in long-horizon prediction" could be more rigorously quantified with standardized relative improvement metrics.
3. Potential Impact
The method addresses a genuine bottleneck in learned simulators: the decoupling of spatial interactions and temporal memory leading to compounding errors over long rollouts. The unified formulation has broad applicability:
The architectural insight — coupling graph Laplacian propagation with SSM recurrence — is modular and could be integrated into other frameworks. The connection to graph ARMA processes and time-varying Koopman theory may inspire theoretical work bridging dynamical systems theory and graph neural networks.
4. Timeliness & Relevance
The paper is well-positioned at the intersection of two active research threads: (1) the rapid adoption of SSMs/Mamba architectures across domains, and (2) the push toward learned simulators for physical systems. The observation that serializing graph nodes for SSM processing (as in GraphMamba) destroys topology is timely given the proliferation of such approaches. The empirical demonstration that topology-preserving Laplacian propagation outperforms artificial node orderings (Tables 1-2) provides important guidance for the community.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Overall Assessment: GraMO makes a solid contribution by identifying and addressing a real architectural limitation in learned simulators for interacting systems. The theoretical analysis is above average for this type of paper, and the experimental validation is thorough. The main risk to long-term impact is scalability — the benchmarks are all small-scale, and it remains unclear whether the approach will maintain its advantages on systems with thousands of interacting entities.
Generated Jun 9, 2026
Comparison History (20)
Paper 2 introduces a concrete, novel architecture (GraMO) that integrates state-space models with graph-based interaction learning in a unified recurrence, addressing a well-defined and important problem (long-horizon prediction of interacting dynamical systems). It demonstrates state-of-the-art results across multiple benchmarks with clear practical applications in physics simulation, robotics, and motion capture. Paper 1 provides valuable empirical analysis and theoretical insights about linear structures in neural networks, but is more analytical/diagnostic in nature, characterizing existing phenomena rather than introducing a transformative new method. Paper 2's broader applicability across scientific domains and its methodological contribution give it higher impact potential.
Paper 2 introduces a novel architectural combination (Graph Mamba Operator) that addresses fundamental challenges in modeling interacting dynamical systems, such as error accumulation over long horizons. Its methodological innovation has broad applicability across multiple fields, including physics simulation, robotics, and motion capture. In contrast, Paper 1, while highly relevant to industrial applications, focuses primarily on semiconductor manufacturing, which may result in a narrower breadth of impact compared to a foundational machine learning architecture.
Paper 1 offers a highly practical and scalable solution to a major bottleneck in reinforcement learning: the instability of training expressive continuous control policies like diffusion and flow models. By shifting policy optimization to test-time gradient guidance, it bypasses complex training dynamics while maintaining competitive performance and lowering computational costs. This has immediate, broad applications in robotics and continuous control. While Paper 2 presents an innovative fusion of state-space models and GNNs, Paper 1 provides a foundational algorithmic advancement likely to see widespread adoption across the rapidly growing field of expressive policy RL.
Paper 1 introduces a new, principled evaluation concept—epistemic calibration—that is strictly stronger than classical calibration, comes with theoretical results (impossibility theorem) and consistent estimators (EECE/TECE), and applies broadly to many uncertainty-quantification methods and high-stakes domains. This combination of conceptual novelty, methodological rigor, and cross-domain relevance suggests wider, longer-lasting impact. Paper 2 is a strong modeling contribution for dynamical systems simulation, but its impact is more application- and benchmark-dependent and may be superseded by rapid architectural turnover.
Paper 2 likely has higher scientific impact: it proposes a new modeling operator (GraMO) that unifies graph interactions with state-space temporal recurrence, targeting a broad class of dynamical systems. This is novel at the algorithmic level, applicable across physics simulation, robotics, and time-series on graphs, and timely given interest in state-space/Mamba-like models. Paper 1 is highly valuable engineering (efficient GMM kernel) with clear practical gains in ANN, but its impact is narrower (GPU kernel + specific clustering/IVF use) and more incremental relative to existing acceleration work.
Paper 1 introduces a fundamental shift in representation learning by integrating symbolic structures (HRRs) into neural networks for disentanglement, backed by rigorous information-theoretic proofs. This neuro-symbolic approach offers broad theoretical implications and novel insights into representation robustness. While Paper 2 presents a strong architectural advancement for dynamical systems by combining GNNs and state-space models, Paper 1's foundational contribution to the longstanding challenge of disentanglement likely yields a deeper and broader impact across multiple subfields of artificial intelligence.
Paper 2 is a comprehensive review that proposes a unifying framework for a rapidly expanding, high-impact field (data-driven physics). Its breadth across multiple scientific domains and its ability to shape future research directions give it a higher potential scientific impact than Paper 1, which presents a specialized, albeit novel, architectural improvement for specific dynamical systems.
Paper 2 (GraMO) has higher likely impact due to strong novelty in coupling graph interactions with state-space/Mamba-style recurrence for long-horizon simulation, clear methodological rigor via broad benchmark wins, and wide applicability to physical simulation, robotics, and any interacting dynamical system. Its operator-style latent simulator targets a timely, high-demand problem (stable long-range dynamics) with immediate practical benefits. Paper 1 is valuable for scientific causal reasoning under structural uncertainty, but appears less broadly validated (domain-limited evaluation) and the integrated two-stage causal discovery + dynamic inference pipeline is a more incremental extension of existing neural causal modeling trends.
Paper 1 introduces a novel integration of state-space models (Mamba) and GNNs, addressing the critical issue of error accumulation in long-horizon predictions. Its demonstrated applications across diverse domains (N-body systems, motion capture, robotics) suggest broad scientific impact. In contrast, Paper 2 presents a valuable but narrower framework for molecular force prediction, with evaluations primarily focused on a specific minimal testbed (NaCl system), limiting its immediate generalizability compared to Paper 1.
Paper 1 addresses a critical bottleneck (fine-tuning trainability) in the rapidly emerging field of Large Time Series Models. Its proposed solution is elegant, methodologically sound, and extensively validated across eight different foundation models. This broad applicability and relevance to foundation model adaptation give it a higher potential for widespread impact compared to Paper 2's domain-specific architectural contribution for dynamical systems.