pith. sign in

arxiv: 2606.27780 · v1 · pith:36PBPHCNnew · submitted 2026-06-26 · 💻 cs.AI

Understanding Rollout Error in Graph World Models

Xinyuan Song , Zekun Cai This is my paper

Pith reviewed 2026-06-29 04:47 UTC · model grok-4.3

classification 💻 cs.AI
keywords graph world modelsrollout errordynamic graphserror-aware trainingplanning regretspectral regularizationagent planninggraph-valued bounds
0
0 comments X

The pith

Error-Aware Graph World Models prevent long-horizon rollout divergence while preserving prediction accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines how local prediction errors in graph world models can spread or amplify over long rollouts, particularly when the graph edges themselves are learned and change. It creates a unified framework covering both fixed-edge and dynamic-edge cases, derives bounds that separate errors caused by the graph's structure from errors caused by the model itself, and introduces a joint operator for handling changing edges. Guided by these bounds, the authors build Error-Aware GWM, which adds spectral regularization, rollout consistency losses, and critical-node weighting. Experiments across synthetic topologies and agent-graph environments show that this method stops divergence at extended horizons while standard models fail, and that dynamic-edge training is required whenever structure evolves. The work clarifies that graph world models are most relevant for planning tasks where the underlying network changes over time.

Core claim

The authors formulate a unified fixed-edge and dynamic-edge GWM framework with action nodes for node-, edge-, and graph-level decisions, develop graph-valued rollout bounds that separate topology-induced amplification from model-induced amplification, introduce a joint node-edge operator for dynamic-edge rollouts, and propose Error-Aware GWM combining spectral regularization, rollout consistency, and critical-node weighting; across synthetic topologies and heterogeneous agent-graph testbeds this approach prevents long-horizon divergence while preserving prediction accuracy, with rollout error and planning regret growing with horizon and dynamic-edge training required when structure evolves.

What carries the argument

Graph-valued rollout bounds that separate topology-induced amplification from model-induced amplification, together with the joint node-edge operator for dynamic-edge cases.

If this is right

  • Rollout error and planning regret increase with horizon length in both fixed and dynamic graph settings.
  • Dynamic-edge training is required whenever the underlying graph structure evolves during rollout.
  • Error-Aware GWM maintains short-term prediction accuracy while avoiding divergence at long horizons.
  • Graph world models are most useful for dynamic graph rollout and agent planning tasks.
  • Specialized graph models remain competitive on static or sparse prediction tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The error-separation bounds could be used to design hybrid models that explicitly correct for known topological amplification effects.
  • The same analysis framework might apply to planning in other structured domains such as molecular interaction graphs or transportation networks.
  • Scalability tests on larger real-world dynamic graphs would reveal whether the joint node-edge operator remains tractable.
  • Combining Error-Aware GWM with model-based reinforcement learning planners could improve sample efficiency in graph-structured environments.

Load-bearing premise

The graph-valued rollout bounds can cleanly separate topology-induced amplification from model-induced amplification for both fixed-edge and dynamic-edge cases.

What would settle it

A controlled experiment on a new dynamic graph environment in which the measured long-horizon divergence rates violate the predicted bounds or in which Error-Aware GWM shows no reduction in divergence relative to a standard GWM baseline.

Figures

Figures reproduced from arXiv: 2606.27780 by Xinyuan Song, Zekun Cai.

Figure 3
Figure 3. Figure 3: FE versus DE rollout regimes. FE models keep edges fixed and reduce error propagation to node￾only dynamics, while DE models predict edges and activate the full joint operator B. fixed during rollout, node features and adjacency matrices are bounded, and FE models are recovered when the edge-update map is absent. Assumption 1 (Compact graph state space). At each time t, the graph state is Gt = (Vt , Et , X… view at source ↗
Figure 2
Figure 2. Figure 2: GWM accumulative-error framework. A GWM rolls out graph states autoregressively; one-step errors accumulate through the joint node-edge operator B and affect long-horizon planning. 4 Error Amplification Theory Framework overview. Figures 1–3 summarize the graph rollout-error framework [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Over-smoothing and GWM error amplifica￾tion use different operators. The over-smoothing op￾erator remains near the normalized propagation domain, while GEAF( \ ℓ) grows with message-passing depth and graph spectral radius. This separates representation col￾lapse from autoregressive rollout amplification. 7.2.2 Rollout amplification is distinct from over-smoothing We next compare GWM rollout amplification w… view at source ↗
Figure 8
Figure 8. Figure 8: reports the resulting spectral radius and NodeMSE@32. As edge density increases, ρ(A) increases. The 0.025 0.050 0.075 0.100 0.125 0.150 0.175 0.200 ER edge probability p 3 4 5 6 7 8 9 10 ρ(A) sp ectral ra dius (a) ρ(A) vs edge density B2 GCN B5 ActionNode B6 ErrorAware 0.025 0.050 0.075 0.100 0.125 0.150 0.175 0.200 ER edge probability p 10 −1 10 4 10 9 10 14 10 19 10 24 10 29 NodeMSE@32 (log) (b) NodeMSE… view at source ↗
Figure 7
Figure 7. Figure 7: Rollout length scaling. NodeMSE@H is plotted as a function of rollout horizon across graph topologies. Shaded regions indicate variation across baselines and seeds. 7.2.4 Topology effects strengthen with rollout horizon and edge density We study how rollout error changes with horizon. The experiment evaluates H ∈ {1, 2, 4, 8, 16, 32, 48} across six baselines, seven topologies, and three seeds [PITH_FULL_I… view at source ↗
Figure 9
Figure 9. Figure 9: Planning-regret growth increases with GEAF. Regret GrowthSlope from horizon 4 to hori￾zon 32 is plotted against log10 GEAF \. Higher-GEAF runs tend to accumulate planning regret faster, linking the rollout-error operator to downstream decision qual￾ity. 12 4 8 16 32 Rollout horizon H 0 10 20 30 40 50 NodeMSE@H Scale-Free GCN Error-Aware 12 4 8 16 32 Rollout horizon H 0 10 20 30 40 50 NodeMSE@H Star GCN Err… view at source ↗
Figure 10
Figure 10. Figure 10: Error-Aware GWM prevents long-horizon divergence. NodeMSE@H is shown for the vanilla GCN world model and Error-Aware GWM on scale￾free and star topologies. Error-Aware GWM stays near the low-error floor, while the vanilla model exhibits horizon-dependent error explosion. 7.4.1 Error-Aware GWM prevents divergence while preserving accuracy [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Graph rewiring comparison. NodeMSE@H = 20 is reported for six rewiring methods on the scale-free topology across three world-model variants and three seeds. Identity denotes the original graph without rewiring. Error-Aware GWM remains stable across rewiring operations, supporting the role of spectral regularization under topology perturbation. topologies, six baselines, and three seeds. Fig￾ure 13 reports… view at source ↗
Figure 13
Figure 13. Figure 13: Graph size scaling. Left: GEAF \ as a function of graph size N. Right: NodeMSE@32 as a function of graph size N. Curves report means across seeds; shaded regions indicate variation across seeds or runs. The low-error regime of ActionNode GWM, GPS, and Error-Aware GWM persists as graph size grows, while the vanilla GCN baseline remains vulnerable on high-risk cells. 7.5.2 DE models activate node-edge coupl… view at source ↗
Figure 14
Figure 14. Figure 14: FE-regime contraction and GEAF re￾duction. Left: injected perturbations decay rapidly to the numerical floor in the FE regime, indicating dor￾mant node-edge coupling. Right: Error-Aware GWM achieves substantially lower GEAF than ActionNode GWM across topologies, showing how spectral regular￾ization reduces the amplification ceiling. Equation (7.1) therefore cannot activate [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 15
Figure 15. Figure 15: Perturbation decomposition in the FE regime. Node-only perturbations are absorbed by con￾tractive dynamics, edge-only perturbations remain near zero growth, and joint perturbations closely follow edge￾only behavior. This supports the interpretation that node￾edge cross-coupling is dormant in the FE regime where MX = 0. tently negative slopes, indicating self-correction under contractive FE dynamics. Edge-… view at source ↗
Figure 17
Figure 17. Figure 17: compares R-GCN predictions with simu￾lator outputs after the sigmoid-head fix. On the agent calling-tree testbed, the R-GCN mean is 0.442 versus ground-truth mean 0.437, with Pear￾son r = 0.311 and MAE = 0.051. On the skill￾graph testbed, the R-GCN mean is 0.387 versus ground-truth mean 0.406, with Pearson r = 0.231 and MAE = 0.029. These results show that the R-GCN models are reasonably calibrated in mea… view at source ↗
Figure 18
Figure 18. Figure 18: Critical subgraph masking on the agent calling-tree testbed. Left: success-rate drop after mask￾ing different subgraphs. Right: NodeMSE@T under the same masking conditions. Full masking gives the largest degradation, while leaf masking has minimal effect. trajectory-level NodeMSE [PITH_FULL_IMAGE:figures/full_fig_p017_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Agent-role injection analysis. Left: failure propagation depth distributions by injection kind. Right: number of error nodes at the final timestep by injection kind. counts are broadly similar across injection types, but the distributions vary by role. These results show that failure propagation depends on func￾tional role, not only graph position. 7.6.3 Correction policies We also test whether a local GE… view at source ↗
Figure 21
Figure 21. Figure 21: Node correction policy comparison. NodeMSE reduction is reported under a 10% correction budget. Oracle correction provides the upper bound, uncertainty gives the strongest practical improvement, and GEAF-proxy correction matches degree-based cor￾rection. Thus, GEAF is useful as a non-random structural heuristic, but in this FE correction setting it does not provide a distinct node-ranking signal beyond de… view at source ↗
Figure 22
Figure 22. Figure 22: Agent-level correction policy compari￾son. Error-flag reduction is reported for random, degree, GEAF-proxy, and oracle correction under a 10% node￾correction budget [PITH_FULL_IMAGE:figures/full_fig_p019_22.png] view at source ↗
Figure 25
Figure 25. Figure 25: Temporal link prediction on Bitcoin￾Alpha. FE-GWM obtains the best AP, Static GCN obtains the best AUC, and JODIE-style obtains the best MRR. DE-GWM does not improve over FE-GWM on this sparse real-world temporal graph, indicating that dynamic-edge modeling is not automatically beneficial outside controlled DE simulators. predictive signal over the rollout horizon. 8 Additional Experiments This section pr… view at source ↗
Figure 26
Figure 26. Figure 26: Seed stability across topologies. Left: seed￾level distributions of NodeMSE@32. Right: seed-level distributions of GrowthSlope. 2 4 6 8 10 Hop distance from injection node 10 −12 10 −9 10 −6 10 −3 10 0 10 3 10 6 10 9 10 12 NodeMSE@H output node (log) Multi-hop Error Decay — Hop Distance vs NodeMSE Topology Chain Scale-Free [PITH_FULL_IMAGE:figures/full_fig_p021_26.png] view at source ↗
Figure 28
Figure 28. Figure 28: shows that cross-topology transfer causes a large increase in NodeMSE@H = 20. Grid Small-World Scale-Free Star Test topology Chain Tree Grid Small-World Scale-Free Train topology 0.25 0.27 2284.81 0.33 865661718.12 OOD NodeMSE@20 Heatmap 1 2 3 4 5 6 7 8 NodeMSE@H=20 (OOD mean) 1e8 [PITH_FULL_IMAGE:figures/full_fig_p021_28.png] view at source ↗
Figure 27
Figure 27. Figure 27: Multi-hop error propagation under per￾turbation injection. NodeMSE@H = 20 is plotted as a function of hop distance from the injection node. Chain graphs remain contractive with near-zero error at all hop distances [PITH_FULL_IMAGE:figures/full_fig_p021_27.png] view at source ↗
Figure 31
Figure 31. Figure 31: Temporal memory variants. NodeMSE@H is reported across rollout hori￾zons for recurrent, transformer, and retrieval-augmented memory variants. Shaded regions indicate variation across topologies and seeds. Markov variant, with most of the gain appearing from the last-2 context. In the extended compari￾son, retrieval-augmented memory achieves lower long-horizon NodeMSE than recurrent and trans￾former varian… view at source ↗
Figure 30
Figure 30. Figure 30: Directed and undirected graph variants. GrowthSlope8→32 is reported for directed and undi￾rected variants of chain, scale-free, and tree graphs. Bars show means and error bars show variation across runs. variants, covering three baselines and three seeds [PITH_FULL_IMAGE:figures/full_fig_p022_30.png] view at source ↗
read the original abstract

World models are often used for planning by rolling learned dynamics forward. Many planning environments, however, are not vectors or images; they are graphs of agents, tools, skills, routes, and dependencies. In these settings, a local prediction error may stay local or spread through the graph, and the failure mode changes again when edges are predicted rather than fixed. This paper studies long-horizon rollout error in Graph World Models (GWMs). We formulate a unified fixed-edge and dynamic-edge GWM framework with action nodes for node-, edge-, and graph-level decisions. We develop graph-valued rollout bounds that separate topology-induced amplification from model-induced amplification, and we introduce a joint node-edge operator for dynamic-edge rollouts. Guided by the analysis, we propose Error-Aware GWM, which combines spectral regularization, rollout consistency, and critical-node weighting. Across synthetic topologies and heterogeneous agent-graph testbeds, rollout error and planning regret grow with horizon, dynamic-edge training is needed when structure evolves, and Error-Aware GWM prevents long-horizon divergence while preserving prediction accuracy. Real-world graph benchmarks clarify the scope of GWMs: they are most useful for dynamic graph rollout and agent planning, while specialized graph models remain strong on static or sparse prediction tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims to study long-horizon rollout error in Graph World Models (GWMs) by formulating a unified framework for fixed-edge and dynamic-edge cases with action nodes. It develops graph-valued rollout bounds separating topology-induced from model-induced amplification, introduces a joint node-edge operator for dynamic-edge rollouts, and proposes Error-Aware GWM using spectral regularization, rollout consistency, and critical-node weighting. Experiments across synthetic topologies and heterogeneous agent-graph testbeds indicate that rollout error grows with horizon, dynamic-edge training is necessary for evolving structures, and Error-Aware GWM prevents divergence while maintaining prediction accuracy. Real-world benchmarks suggest GWMs are useful for dynamic graph rollout and agent planning.

Significance. Should the derived bounds rigorously separate the two sources of amplification and the experimental results demonstrate clear improvements with statistical significance, this work would offer valuable theoretical and practical guidance for designing robust world models in graph-structured environments, particularly for long-horizon planning where error propagation is a key concern. The unified treatment of fixed and dynamic edges addresses an important extension beyond standard vector or image-based world models.

major comments (2)
  1. Abstract: The central claim that the graph-valued rollout bounds separate topology-induced amplification from model-induced amplification, including for the joint node-edge operator in dynamic-edge cases, lacks any derivation steps, explicit assumptions, or proof sketches. This makes it impossible to assess whether the separation remains valid when edge predictions depend on prior node states, as the operator may couple the sources.
  2. Abstract: No quantitative experimental results, error bars, specific metrics (e.g., rollout error values, planning regret), or comparisons to baselines are provided, undermining the ability to verify claims about Error-Aware GWM preventing long-horizon divergence while preserving accuracy across testbeds.
minor comments (1)
  1. The abstract mentions 'real-world graph benchmarks' but does not specify which ones or how they clarify the scope of GWMs relative to specialized graph models.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed feedback on the abstract. Both comments correctly identify that the abstract is too concise to fully convey the derivations or quantitative results. We will revise the abstract to incorporate a brief outline of the key assumptions, separation result, and sample metrics while remaining within length limits. The full manuscript already contains the supporting material in Sections 3 and 5; the revisions will make this more immediately accessible from the abstract. We respond to each major comment below.

read point-by-point responses

  1. Referee: Abstract: The central claim that the graph-valued rollout bounds separate topology-induced amplification from model-induced amplification, including for the joint node-edge operator in dynamic-edge cases, lacks any derivation steps, explicit assumptions, or proof sketches. This makes it impossible to assess whether the separation remains valid when edge predictions depend on prior node states, as the operator may couple the sources.

    Authors: We agree that the abstract omits derivation steps and proof sketches. The manuscript derives the bounds in Section 3 by decomposing the rollout operator into a topology-dependent amplification term (controlled by the spectral radius of the graph operator) and an additive model-error term (bounded by the predictor's Lipschitz constant). For the dynamic-edge case the joint node-edge operator is defined so that edge predictions are applied after the current node state, allowing the two error sources to be bounded separately; the proof sketch appears in Appendix A under the stated assumptions of bounded node features and Lipschitz-continuous predictors. We will revise the abstract to state the main assumptions and the separation result explicitly, thereby addressing the concern about potential coupling. revision: yes

  2. Referee: Abstract: No quantitative experimental results, error bars, specific metrics (e.g., rollout error values, planning regret), or comparisons to baselines are provided, undermining the ability to verify claims about Error-Aware GWM preventing long-horizon divergence while preserving accuracy across testbeds.

    Authors: We agree that the abstract presents only qualitative summaries of the experimental findings. The full manuscript reports quantitative results in Section 5, including rollout error curves with error bars across horizons, planning regret values, and statistical comparisons (paired t-tests) against standard GNN and RNN baselines on both synthetic and real-world testbeds. To allow immediate verification of the claims about divergence prevention and accuracy preservation, we will revise the abstract to include one or two representative quantitative highlights drawn from those tables and figures. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation chain self-contained with no reductions to fitted inputs or self-citations shown

full rationale

The abstract formulates a unified GWM framework, develops graph-valued rollout bounds separating topology-induced from model-induced amplification, introduces a joint node-edge operator, and proposes Error-Aware GWM combining spectral regularization, rollout consistency, and critical-node weighting. No equations, self-citations, or fitted parameters are quoted that reduce any prediction or bound to its own inputs by construction. The reader's assessment notes absence of indication that bounds or method reduce to fitted quantities. Without load-bearing steps that collapse to self-definition or fitted renaming, the analysis remains independent of its outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the framework is described at the level of operators and regularization terms without numerical or definitional detail.

pith-pipeline@v0.9.1-grok · 5744 in / 1110 out tokens · 47410 ms · 2026-06-29T04:47:19.148908+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

73 extracted references · 13 canonical work pages · 10 internal anchors

  1. [1]

    International Conference on Machine Learning (ICML) , year =

    Lipschitz Continuity in Model-based Reinforcement Learning , author =. International Conference on Machine Learning (ICML) , year =

  2. [2]

    Science , volume=

    Emergence of Scaling in Random Networks , author=. Science , volume=. 1999 , publisher=

    1999

  3. [3]

    Nature , volume=

    Collective Dynamics of `Small-World' Networks , author=. Nature , volume=. 1998 , publisher=

    1998

  4. [4]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    When to Trust Your Model: Model-Based Policy Optimization , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  5. [5]

    2024 , eprint=

    A Note on Loss Functions and Error Compounding in Model-based Reinforcement Learning , author=. 2024 , eprint=

    2024

  6. [6]

    International Conference on Machine Learning (ICML) , year =

    Model-Based Reinforcement Learning with Value-Targeted Regression , author =. International Conference on Machine Learning (ICML) , year =

  7. [7]

    2019 , eprint=

    Learning to Combat Compounding-Error in Model-Based Reinforcement Learning , author=. 2019 , eprint=

    2019

  8. [8]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Recurrent World Models Facilitate Policy Evolution , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  9. [9]

    World Models

    World Models , author=. arXiv preprint arXiv:1803.10122 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Proceedings of the 36th International Conference on Machine Learning , pages=

    Learning Latent Dynamics for Planning from Pixels , author=. Proceedings of the 36th International Conference on Machine Learning , pages=

  11. [11]

    International Conference on Learning Representations , year=

    Dream to Control: Learning Behaviors by Latent Imagination , author=. International Conference on Learning Representations , year=

  12. [12]

    International Conference on Learning Representations , year=

    Mastering Atari with Discrete World Models , author=. International Conference on Learning Representations , year=

  13. [13]

    International Conference on Learning Representations , year =

    Semi-Supervised Classification with Graph Convolutional Networks , author =. International Conference on Learning Representations , year =

  14. [14]

    Proceedings of the 34th International Conference on Machine Learning , pages =

    Neural Message Passing for Quantum Chemistry , author =. Proceedings of the 34th International Conference on Machine Learning , pages =. 2017 , series =

    2017

  15. [15]

    Advances in Neural Information Processing Systems , volume =

    Recipe for a General, Powerful, Scalable Graph Transformer , author =. Advances in Neural Information Processing Systems , volume =

  16. [16]

    International Conference on Machine Learning , year =

    Graph World Model , author =. International Conference on Machine Learning , year =

  17. [17]

    International Conference on Learning Representations , year=

    Transformers are Sample-Efficient World Models , author=. International Conference on Learning Representations , year=

  18. [18]

    Proceedings of the 6th Conference on Robot Learning , pages=

    DayDreamer: World Models for Physical Robot Learning , author=. Proceedings of the 6th Conference on Robot Learning , pages=

  19. [19]

    Hansen, Nicklas and Su, Hao and Wang, Xiaolong , booktitle=

  20. [20]

    Learning Latent Dynamics for Planning from Pixels

    Learning Latent Dynamics for Planning from Pixels , author =. arXiv preprint arXiv:1811.04551 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    International Conference on Learning Representations (ICLR) , year =

    Dream to Control: Learning Behaviors by Latent Imagination , author =. International Conference on Learning Representations (ICLR) , year =

  22. [22]

    Mastering Diverse Domains through World Models

    Mastering Diverse Domains through World Models , author =. arXiv preprint arXiv:2301.04104 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Nature , year =

    Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model , author =. Nature , year =

  24. [24]

    International Conference on Machine Learning (ICML) , year =

    Graph World Model , author =. International Conference on Machine Learning (ICML) , year =

  25. [25]

    Graph World Models: Concepts, Taxonomy, and Future Directions

    Graph World Models: Concepts, Taxonomy, and Future Directions , author =. arXiv preprint arXiv:2604.27895 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    International Conference on Machine Learning (ICML) , year =

    World Model as a Graph: Learning Latent Landmarks for Planning , author =. International Conference on Machine Learning (ICML) , year =

  27. [27]

    International Conference on Learning Representations (ICLR) , year =

    Value Memory Graph: A Graph-Structured World Model for Offline Reinforcement Learning , author =. International Conference on Learning Representations (ICLR) , year =

  28. [28]

    2025 , eprint=

    SPARTAN: A Sparse Transformer World Model Attending to What Matters , author=. 2025 , eprint=

    2025

  29. [29]

    2025 , eprint=

    Dyn-O: Building Structured World Models with Object-Centric Representations , author=. 2025 , eprint=

    2025

  30. [30]

    2025 , eprint=

    BAgger: Backwards Aggregation for Mitigating Drift in Autoregressive Video Diffusion Models , author=. 2025 , eprint=

    2025

  31. [31]

    International Joint Conference on Artificial Intelligence (IJCAI) , year =

    AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents , author =. International Joint Conference on Artificial Intelligence (IJCAI) , year =

  32. [32]

    International Conference on Learning Representations (ICLR) , year =

    Contrastive Learning of Structured World Models , author =. International Conference on Learning Representations (ICLR) , year =

  33. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Learning Knowledge Graph-based World Models of Textual Environments , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  34. [34]

    International Conference on Learning Representations (ICLR) , year =

    Graph Neural Networks Exponentially Lose Expressive Power For Node Classification , author =. International Conference on Learning Representations (ICLR) , year =

  35. [35]

    2023 , eprint=

    A Survey on Oversmoothing in Graph Neural Networks , author=. 2023 , eprint=

    2023

  36. [36]

    Relational inductive biases, deep learning, and graph networks

    Relational Inductive Biases, Deep Learning, and Graph Networks , author =. arXiv preprint arXiv:1806.01261 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    VRAG: Learning World Models for Interactive Video Generation

    Learning World Models for Interactive Video Generation , author=. arXiv preprint arXiv:2505.21996 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [38]

    Hierarchical Planning with Latent World Models

    Hierarchical Planning with Latent World Models , author=. arXiv preprint arXiv:2604.03208 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    Reinforcement Learning Journal , volume=

    Long-Horizon Planning with Predictable Skills , author=. Reinforcement Learning Journal , volume=

  40. [40]

    arXiv preprint arXiv:2511.23465 , year=

    SmallWorlds: Assessing Dynamics Understanding of World Models in Isolated Environments , author=. arXiv preprint arXiv:2511.23465 , year=

    work page arXiv

  41. [41]

    2026 , eprint=

    GraSP: Graph-Structured Skill Compositions for LLM Agents , author=. 2026 , eprint=

    2026

  42. [42]

    arXiv preprint arXiv:2507.08800 , year=

    NeuralOS: Towards Simulating Operating Systems via Neural Generative Models , author=. arXiv preprint arXiv:2507.08800 , year=

    work page arXiv

  43. [43]

    Prismatic World Model: Learning Compositional Dynamics for Planning in Hybrid Systems

    Prismatic World Model: Learning Compositional Dynamics for Planning in Hybrid Systems , author=. arXiv preprint arXiv:2512.08411 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [44]

    Long-Horizon Model-Based Offline Reinforcement Learning Without Explicit Conservatism

    Long-Horizon Model-Based Offline Reinforcement Learning Without Conservatism , author=. arXiv preprint arXiv:2512.04341 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  45. [45]

    ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year =

    ROLAND: Graph Learning Framework for Dynamic Graphs , author =. ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year =

  46. [46]

    International Conference on Machine Learning (ICML) , year =

    GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models , author =. International Conference on Machine Learning (ICML) , year =

  47. [47]

    International Conference on Learning Representations (ICLR) , year =

    DiGress: Discrete Denoising Diffusion for Graph Generation , author =. International Conference on Learning Representations (ICLR) , year =

  48. [48]

    From Spark to Fire: Modeling and Mitigating Error Cascades in LLM-Based Multi-Agent Collaboration

    From Spark to Fire: Modeling and Mitigating Error Cascades in LLM-Based Multi-Agent Collaboration , author =. arXiv preprint arXiv:2603.04474 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    2026 , eprint=

    Evaluating Tool-Using Language Agents: Judge Reliability, Propagation Cascades, and Runtime Mitigation in AgentProp-Bench , author=. 2026 , eprint=

    2026

  50. [50]

    2024 , eprint=

    The Effectiveness of Curvature-Based Rewiring and the Role of Hyperparameters in GNNs Revisited , author=. 2024 , eprint=

    2024

  51. [51]

    2025 , eprint=

    PIORF: Physics-Informed Ollivier-Ricci Flow for Long-Range Interactions in Mesh Graph Neural Networks , author=. 2025 , eprint=

    2025

  52. [52]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  53. [53]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Recipe for a General, Powerful, Scalable Graph Transformer , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  54. [54]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Do Transformers Really Perform Bad for Graph Representation? , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  55. [55]

    arXiv preprint arXiv:2512.23760 , year =

    Audited Skill-Graph Self-Improvement for Agentic LLMs via Verifiable Rewards, Experience Synthesis, and Continual Memory , author =. arXiv preprint arXiv:2512.23760 , year =

    work page arXiv

  56. [56]

    2026 , eprint=

    CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution , author=. 2026 , eprint=

    2026

  57. [57]

    2026 , eprint=

    Reinforcement Learning for Self-Improving Agent with Skill Library , author=. 2026 , eprint=

    2026

  58. [58]

    International Conference on Learning Representations (ICLR) , year =

    Semi-Supervised Classification with Graph Convolutional Networks , author =. International Conference on Learning Representations (ICLR) , year =

  59. [59]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Inductive Representation Learning on Large Graphs , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  60. [60]

    International Conference on Learning Representations (ICLR) , year =

    Graph Attention Networks , author =. International Conference on Learning Representations (ICLR) , year =

  61. [61]

    International Conference on Machine Learning (ICML) , year =

    Neural Message Passing for Quantum Chemistry , author =. International Conference on Machine Learning (ICML) , year =

  62. [62]

    International Conference on Machine Learning (ICML) , year =

    Approximately Optimal Approximate Reinforcement Learning , author =. International Conference on Machine Learning (ICML) , year =

  63. [63]

    Proceedings of the 41st International Conference on Machine Learning , year=

    Genie: Generative Interactive Environments , author=. Proceedings of the 41st International Conference on Machine Learning , year=

  64. [64]

    Wu, Jialong and Hu, Shaofeng and Guo, Ningya and Guo, Lanqing and Du, Jian and Yan, Junwei and Hao, Jianye and Huang, Minlie , journal=

  65. [65]

    The Twelfth International Conference on Learning Representations , year=

    One for All: Towards Training One Graph Model for All Classification Tasks , author=. The Twelfth International Conference on Learning Representations , year=

  66. [66]

    2024 , publisher=

    Chen, Runjin and Zhao, Tong and Jaiswal, Ajay and Shah, Neil and Wang, Zhangyang , booktitle=. 2024 , publisher=

    2024

  67. [67]

    International Conference on Learning Representations , year=

    Graph Attention Networks , author=. International Conference on Learning Representations , year=

  68. [68]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

    Predicting Dynamic Embedding Trajectory in Temporal Interaction Networks , author=. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

  69. [69]

    Proceedings of the 33rd International Conference on Machine Learning , year=

    Revisiting Semi-Supervised Learning with Graph Embeddings , author=. Proceedings of the 33rd International Conference on Machine Learning , year=

  70. [70]

    Advances in Neural Information Processing Systems , volume=

    Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models , author=. Advances in Neural Information Processing Systems , volume=

  71. [71]

    Advances in Neural Information Processing Systems , volume=

    Interaction Networks for Learning about Objects, Relations and Physics , author=. Advances in Neural Information Processing Systems , volume=

  72. [72]

    Proceedings of the 35th International Conference on Machine Learning , pages=

    Graph Networks as Learnable Physics Engines for Inference and Control , author=. Proceedings of the 35th International Conference on Machine Learning , pages=. 2018 , series=

    2018

  73. [73]

    and Bronstein, Michael and Webb, Stefan and Rossi, Emanuele , booktitle=

    Chamberlain, Ben and Rowbottom, James and Gorinova, Maria I. and Bronstein, Michael and Webb, Stefan and Rossi, Emanuele , booktitle=. 2021 , series=

    2021