GLAM: Global-Local Variation Awareness in Mamba-based World Model

Chunhui Hao; Gan Sun; Jiandong Tian; Qian He; Wenqi Liang

REVIEW 4 major objections 5 minor 47 references

GLAM: Global-Local Variation Awareness in Mamba-based World Model

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A world model that reasons about the change between states, not the states themselves, reports state-of-the-art human-normalized scores on the Atari 100k benchmark.

desk verdict A clean architectural idea with plausible but not yet statistically robust Atari results; worth reviewing, but the headline claim needs error bars and a direct S4WM comparison. read the letter →

arxiv 2501.11949 v1 pith:NLM2MHDE submitted 2025-01-21 cs.LG

Qian He , Wenqi Liang , Chunhui Hao , Gan Sun , Jiandong Tian This is my paper · ORCID

classification cs.LG

keywords worldmodelsmodel-basedreinforcementlearningstatespaceMambavariationawarenessAtari100kimaginationreasoningsampleefficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that a model-based reinforcement-learning agent can reason more effectively about its environment by predicting how state representations change from one step to the next, rather than predicting states directly. To test this, the authors build GLAM, a world model built from two parallel Mamba modules: GMamba tracks long-range patterns of state-feature differences, and LMamba uses short-range differences to guess rewards, termination signals, and the next visual state. On the Atari 100k benchmark, GLAM reports a mean human-normalized score of 130.6% and a median of 61.8%, above the previous best method's 122.3% and 58.4%. If true, this shows that explicit variation awareness is a useful inductive bias for sample-efficient imagination-based training.

What carries the argument

The central object is the state-feature difference vector, $d_t = e_t - e_{t-1}$, computed between consecutive feature representations of observations and actions. The paper's design principle is to make this difference, not the absolute state, the unit of reasoning: GMamba runs a Mamba (selective state space) layer over the whole difference sequence to capture global variation patterns, while LMamba runs a Mamba layer over short sequences to infer local variation and unknown signals. A dedicated variation loss (Eq. 10) forces GMamba's prediction to match the observed next-frame difference, and all predictions for next state, reward, and termination are produced by combining the two modules' outputs, $u^g_t$ and $u^l_t$.

What would settle it

Construct or modify an Atari-like game where the agent's reward depends on a static, absolute feature that appears nowhere as a change between consecutive frames (for example, a visible score counter that only increments on a specific action), and train GLAM on it. If its performance is at or below random, while a full-state model solves it, the central assumption is disproven, because the difference representation would be blind to that absolute information.

Watch

Extended reading notes

Core claim

GLAM establishes that a world model can be built almost entirely around first-order differences between encoded state features. Two Mamba-based modules operate in parallel: GMamba receives a long sequence (16 steps), computes the sequence of differences between consecutive features, and learns to summarize global patterns that predict the distribution of future variation; LMamba receives a short sequence (4 steps) and predicts rewards, terminations, and next-state representations from local variation. Their outputs are combined by dynamics, reward, and continuation predictors, and a dedicated variation loss explicitly trains GMamba to predict the next-frame feature difference. The authors report that GLAM achieves the highest reported human-normalized mean and median scores on the Atari 100k benchmark (130.6% and 61.8%, respectively), with ablation results indicating that each module and the variation loss contribute to the gain.

Load-bearing premise

The model assumes that the difference between two consecutive encoded states is enough to predict future states, rewards, and terminations, so that the absolute state itself can be discarded without much loss.

Editorial extensions

If this is right

If correct, world models can improve sample efficiency by modeling change rather than absolute state, since differences are often more predictable and less redundant.
The parallel two-module design shows that Mamba can serve as an efficient alternative to Transformer backbones in world-model inference.
The success on variation-heavy Atari games suggests the idea may transfer to tasks where motion and change are central signals.
The variable interaction schedule implies that matching the imagination horizon to world-model maturity improves agent training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

One natural extension is to include second-order differences (differences of differences) to handle acceleration and momentum, which the first-order design discards.
The failure mode noted in the appendix—objects blending into the background—suggests that a hybrid architecture keeping a low-frequency absolute-state channel alongside difference features might be more robust.
A testable prediction follows: a task where reward depends on an absolute screen property that is constant between frames should be hard for GLAM, because the difference representation is blind to that information.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this paper first: GLAM is a Mamba-based world model that feeds state differences into two parallel modules (GMamba for long-range variation patterns, LMamba for local variation-aware prediction of rewards, terminations, and next states). That is genuinely new in this line of work, and the idea is simple enough to be useful. The paper ships code, the ablations are honest, and the limitations section is more candid than most.

What it does well: the architecture is clearly motivated, the ablations isolate the main modules (though on only four games), and the variable imagination horizon is a reasonable add-on. The Atari 100k numbers are plausible, and the median gain of roughly 3 points over STORM is the kind of modest, believable improvement you'd expect from a good architectural trick.

Now the soft spots. The headline claim that GLAM 'outperforms existing methods' is not backed by any seed-wise error bars or statistical tests. Per-game, GLAM is worse than STORM on 16 of 26 games; the mean gain is driven by a few big wins (Kung Fu Master, Krull, Road Runner). Without variance, that gap could easily be seed noise. The ablation is also single-run, on four games, and it never compares against S4WM, which is directly relevant as an SSM-based world model and is cited in the paper. The claim to be 'first' at parallel multi-module inference is overbold and unnecessary.

One thing in the stress-test note that I'd push back on: the 'ours w/o G&L' Pong score of 8 is not meant to be a reproduction of STORM's 11—it's a Mamba-backbone baseline, so the mismatch is expected. The real issue is the general absence of variance, not that particular number.

Bottom line: this is a serious paper with a clear, testable idea. It deserves peer review. I'd ask the authors for seed-wise stats, a paired test or at least per-game error bars, an S4WM baseline, and a tempering of the 'first' claims. For someone working on sample-efficient world models, it's worth a read. I'd probably cite it once the code runs cleanly and the numbers survive scrutiny.

Referee Report

4 major / 5 minor

Summary. This paper proposes GLAM, a Mamba-based world model for sample-efficient model-based RL. The model computes differences between consecutive latent state features and uses two parallel modules: LMamba, which conditions prediction of rewards, terminations, and next-state features on local (short-window) variations, and GMamba, which learns global patterns of variation over longer sequences and predicts future variation. The outputs of the two modules are combined by MLP predictors, and a phased imagination-horizon schedule is introduced. The method is evaluated on 26 Atari 100k games against SimPLe, TWM, IRIS, DreamerV3, Hieros, and STORM, reporting a human-normalized mean of 130.6% and median of 61.8%, with ablations on four games.

Significance. If the empirical claims hold, GLAM makes a modest but useful contribution: explicit difference-based reasoning is a sensible inductive bias for world models, and the local/global split is clearly motivated. The manuscript is generally reproducible: it releases code, gives algorithmic pseudocode, and reports hyperparameters. The benchmark evaluation is external, so the auxiliary variation loss (Eq. 10) does not create circularity. The main weakness is experimental: the headline improvement over STORM is small, per-game results are mixed, and no uncertainty quantification is provided. The paper's own qualitative results (Fig. 7) confirm that first-order differences are imperfect for objects blending into the background, but the authors correctly note that final agent scores do not require pixel-perfect imagination; I therefore do not treat the difference-feature concern as fatal.

major comments (4)

[Results on Atari 100k, Table 1] The headline claim that GLAM outperforms all baselines rests on an aggregate mean gap of 8.3 human-normalized points and a median gap of 3.4 points over STORM, but the paper reports only average final-checkpoint scores, with no seed-wise results, standard deviations, or significance tests despite stating that 5 seeds and 20 evaluation episodes were used. Per-game, GLAM is lower than STORM on many games, e.g., Assault 646 vs 801, Asterix 504 vs 1028, Gopher 3149 vs 8240, and Private Eye 100 vs 7781. Without variance information and a paired test, the aggregate gap could be seed noise.
[Ablation Studies, Table 2] The ablation baseline labeled 'ours w/o G&L' is described as STORM, but its Pong score is 8, whereas STORM's Pong score in Table 1 is 11. Since this row is the controlled comparison for the ablation claims, the inconsistency makes the ablation table internally unreliable; the authors should either use the identical protocol as Table 1 or explain the discrepancy, and they should provide variance information for the four ablation games as well.
[Variable Number of Interaction and Number of layers in the Mamba, Fig. 4] The default hyperparameters that are used for all 26 games, including the variable-horizon schedule (nmax, ni, delta_n) and the GMamba/LMamba layer counts, are selected based on experiments on a single game, KungFuMaster. Since the phased imagination-horizon schedule is one of the stated contributions, the authors need to justify why choices tuned on one game transfer to the other 25 games, or provide cross-game validation; otherwise the final benchmark scores may partly reflect per-game hyperparameter selection.
[Method, Eq. (10) and Eq. (5)] The variation loss in Eq. (10) is written as KL(delta_o_t || gD_phi(u^g_t)), but the dynamics predictor in Eq. (5) is defined as gD_phi(u^g_t, u^l_t) over next-state features. The loss also compares a point estimate of the feature difference, delta_o_t = sg(z_{t+1} - z_t), with what appears to be a distribution over next states. This overloading and semantic mismatch need to be specified precisely, because the variation loss is central to the training of GMamba.

minor comments (5)

[Variable Number of Interaction, Eq. (11)] The notation around Eq. (11) is inconsistent: the formula uses ni and delta_n, while the text introduces nf instead of ni and calls delta_n the increment, although it appears in the denominator as an interval; please align the symbols with Table 3.
[Loss Function, Eq. (8) and Appendix Eq. (12)] The loss in Eq. (8) omits the weights beta_1, beta_2, beta_3 that appear in Eq. (12) and Table 3; the main text should use the weighted form consistently.
[Algorithm 1] Algorithm 1 contains apparent typos: the line 'u^g_{1:t} = SiLU(g_{1:t})' should presumably be 'SiLU(u^g_{1:t})', and the module used for the global variation inference is called 'LMamba' even though the text says GMamba handles the global sequence.
[Table 1 caption] The caption states that 'model results with score fluctuations within 5% are indicated in bold font', but no fluctuation statistics or confidence intervals are reported anywhere, so this convention is not interpretable.
[Related Work and Baselines] The related work credits S4WM as the first universal SSM-based world model, but S4WM is not included in the Atari 100k comparison; the authors should add this baseline or explicitly justify its exclusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: GLAM's variation-prediction losses are standard supervised objectives and the headline claim is tested against an external benchmark.

full rationale

The derivation chain in GLAM is not circular. The core modules GMamba and LMamba are trained with explicit self-supervised losses (Eqs. 8-10): L_VAR minimizes KL(Delta_ot || gD(u_g_t)) with Delta_ot = sg(z_{t+1} - z_t), i.e., the model is trained to match observed next-frame feature differences. This is a standard supervised learning objective, not a fitted parameter later renamed as a prediction; the same trained model is then rolled out in imagination and evaluated against external Atari 100k scores and baselines (Table 1). The paper's self-citations (e.g., Liang et al. 2024; Sun et al. 2024) appear only in related-work context and are not load-bearing for the method or result. No uniqueness theorem or prior-work ansatz is imported to force the architecture choice; Mamba is adopted with a rationale about causal selective scanning, and the LMamba/GMamba split is an architectural proposal rather than a derived result. The supervised L_VAR and the KL dynamics losses do mean the model's 'predictions' are trained on ground-truth differences, but this is the normal world-model training loop and does not make the downstream benchmark comparison circular. The reader's cited concern about per-game variability and missing error bars is a statistical robustness issue, not circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

GLAM's empirical gains are conditional on several hand-chosen hyperparameters and on the domain assumption that state-feature differences are informative. No new theoretical entities are introduced.

free parameters (7)

GMamba sequence length l = 16
Fixed input length for global module; chosen without a stated principle; results may depend on it.
LMamba sequence length s = 4
Fixed input length for local module; similarly hand-chosen.
Initial imagination horizon n0 = 16
Variable interaction schedule starts at 16 imagined steps.
Horizon increment delta_n = 8
Number of steps added to imagination horizon per update.
Horizon increase frequency ni = 25000
Every 25000 env steps the horizon grows; Eq. 11.
Maximum imagination horizon nmax = 32
Cap on number of imagined interactions; ablated on KungFuMaster only.
Loss weights beta1, beta2, beta3 = 0.5, 0.1, 0.1
Weights for dynamics, representation, and variation losses; hand-chosen.

assumptions (4)

domain assumption First-order state-feature differences are a sufficient signal for predicting future states, rewards, and terminations in Atari.
Core premise of GLAM (Eq. 5-7, L_VAR Eq. 10); no formal or empirical justification beyond final scores on 4 games.
domain assumption Mamba's selective scanning aligns better than Transformer self-attention with the causal structure of variation-driven environment changes.
Stated in Method/Preliminary without comparative evidence; claimed but not tested.
domain assumption Atari 100k human-normalized mean is an appropriate aggregate metric for sample efficiency.
Standard in the field, but the mean is outlier-dominated (e.g., Krull contributes 10.35 normalized).
domain assumption The imagination horizon schedule (variable interaction) improves training without hurting world model accuracy.
Ablated only on KungFuMaster (Fig. 4); not validated across the benchmark.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GLAM: Global-Local Variation Awareness in Mamba-based World Model." pith.science (2026). https://pith.science/paper/NLM2MHDE

@misc{pith2026250111949,
  author       = {Pith},
  title        = {Pith review of: GLAM: Global-Local Variation Awareness in Mamba-based World Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NLM2MHDE}},
  note         = {Machine review of arXiv:2501.11949}
}

read the original abstract

Mimicking the real interaction trajectory in the inference of the world model has been shown to improve the sample efficiency of model-based reinforcement learning (MBRL) algorithms. Many methods directly use known state sequences for reasoning. However, this approach fails to enhance the quality of reasoning by capturing the subtle variation between states. Much like how humans infer trends in event development from this variation, in this work, we introduce Global-Local variation Awareness Mamba-based world model (GLAM) that improves reasoning quality by perceiving and predicting variation between states. GLAM comprises two Mambabased parallel reasoning modules, GMamba and LMamba, which focus on perceiving variation from global and local perspectives, respectively, during the reasoning process. GMamba focuses on identifying patterns of variation between states in the input sequence and leverages these patterns to enhance the prediction of future state variation. LMamba emphasizes reasoning about unknown information, such as rewards, termination signals, and visual representations, by perceiving variation in adjacent states. By integrating the strengths of the two modules, GLAM accounts for highervalue variation in environmental changes, providing the agent with more efficient imagination-based training. We demonstrate that our method outperforms existing methods in normalized human scores on the Atari 100k benchmark.

Figures

Figures reproduced from arXiv: 2501.11949 by the authors.

**Figure 2.** Overview of the Global-Local variation Awareness Mamba-based world model (GLAM). Based on the known sequence, GLAM infers future information to train the agent in imagination. In each inference, GLAM captures the awareness of global and local variation from long and short sequences and integrates both for prediction. sequences. The latest SSM, Mamba (Gu and Dao 2023), introduces selective scanning and hardware-awa… view at source ↗

**Figure 3.** Comparison of the training results between GLAM and STORM. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Ablation study on variable imagine steps (left) and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 6.** Figure 6: The comparison between ours w/o G&L and ours w/o G on Pong, Battle Zone, Boxing and Kung Fu Master. Atari video predictions In [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: The comparison between the imagined trajectories generated by GLAM and the actual game frames. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 24 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Alonso, E.; Jelley, A.; Micheli, V.; Kanervisto, A.; Storkey, A.; Pearce, T.; and Fleuret, F. 2024. Diffusion for World Modeling: Visual Details Matter in Atari. arXiv preprint arXiv:2405.12399

arXiv 2024
[4]

L.; Kiros, J

Ba, J. L.; Kiros, J. R.; and Hinton, G. E. 2016. Layer normalization. arXiv preprint arXiv:1607.06450

arXiv 2016
[5]

G.; Naddaf, Y.; Veness, J.; and Bowling, M

Bellemare, M. G.; Naddaf, Y.; Veness, J.; and Bowling, M. 2013. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47: 253--279

work page 2013
[6]

Chen, C.; Wu, Y.-F.; Yoon, J.; and Ahn, S. 2022. Transdreamer: Reinforcement learning with transformer world models. arXiv preprint arXiv:2202.09481

arXiv 2022
[7]

Deng, F.; Park, J.; and Ahn, S. 2024. Facing off world model backbones: Rnns, transformers, and S4. 36

work page 2024
[8]

Ding, Z.; Zhang, A.; Tian, Y.; and Zheng, Q. 2024. Diffusion world model. arXiv preprint arXiv:2402.03570

arXiv 2024

Show all 47 references

[9]

Dong, J.; Liang, W.; Cong, Y.; and Sun, G. 2023. Heterogeneous forgetting compensation for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 11742--11751

2023
[10]

Dosovitskiy, A. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

2020 arXiv
[11]

Elfwing, S.; Uchibe, E.; and Doya, K. 2018. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural networks, 107: 3--11

2018
[12]

Gao, Z.; Mu, Y.; Chen, C.; Duan, J.; Luo, P.; Lu, Y.; and Li, S. E. 2024. Enhance Sample Efficiency and Robustness of End-to-End Urban Autonomous Driving via Semantic Masked World Model. IEEE Transactions on Intelligent Transportation Systems, 1--13

2024
[13]

Gu, A.; and Dao, T. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

2023 arXiv
[14]

Gu, A.; Goel, K.; and R \'e , C. 2021. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396

2021 arXiv
[15]

Ha, D.; and Schmidhuber, J. 2018. Recurrent world models facilitate policy evolution. 31

2018
[16]

Ha, J.; Kim, K.; and Kim, Y. 2023. Dream to generalize: zero-shot model-based reinforcement learning for unseen visual distractions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 7802--7810

2023
[17]

Hafner, D.; Lillicrap, T.; Ba, J.; and Norouzi, M. 2019 a . Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603

2019 arXiv
[18]

Hafner, D.; Lillicrap, T.; Fischer, I.; Villegas, R.; Ha, D.; Lee, H.; and Davidson, J. 2019 b . Learning latent dynamics for planning from pixels. In International conference on machine learning, 2555--2565. PMLR

2019
[19]

Hafner, D.; Lillicrap, T.; Norouzi, M.; and Ba, J. 2020. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193

2020 arXiv
[20]

Hafner, D.; Pasukonis, J.; Ba, J.; and Lillicrap, T. 2023. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104

2023 arXiv
[21]

M.; and Bertasius, G

Islam, M. M.; and Bertasius, G. 2022. Long movie clip classification with state-space video models. In European Conference on Computer Vision, 87--104. Springer

2022
[22]

H.; Czechowski, K.; Erhan, D.; Finn, C.; Kozakowski, P.; Levine, S.; et al

Kaiser, L.; Babaeizadeh, M.; Milos, P.; Osinski, B.; Campbell, R. H.; Czechowski, K.; Erhan, D.; Finn, C.; Kozakowski, P.; Levine, S.; et al. 2019. Model-based reinforcement learning for atari. arXiv preprint arXiv:1903.00374

2019 arXiv
[23]

E.; et al

Kalman, R. E.; et al. 1960. A new approach to linear filtering and prediction problems [J]. Journal of basic Engineering, 82(1): 35--45

1960
[24]

S.; Henderson, D.; Howard, R

LeCun, Y.; Boser, B.; Denker, J. S.; Henderson, D.; Howard, R. E.; Hubbard, W.; and Jackel, L. D. 1989. Backpropagation applied to handwritten zip code recognition. Neural computation, 1(4): 541--551

1989
[25]

Liang, W.; Sun, G.; He, Q.; Ren, Y.; Dong, J.; and Cong, Y. 2024. Never-Ending Embodied Robot Learning. arXiv preprint arXiv:2403.00336

2024
[26]

Mattes, P.; Schlosser, R.; and Herbrich, R. 2023. Hieros: Hierarchical Imagination on Structured State Space Sequence World Models. arXiv preprint arXiv:2310.05167

2023 arXiv
[27]

Micheli, V.; Alonso, E.; and Fleuret, F. 2022. Transformers are sample-efficient world models. arXiv preprint arXiv:2209.00588

2022 arXiv
[28]

Min, C.; Zhao, D.; Xiao, L.; Zhao, J.; Xu, X.; Zhu, Z.; Jin, L.; Li, J.; Guo, Y.; Xing, J.; Jing, L.; Nie, Y.; and Dai, B. 2024. DriveWorld: 4D Pre-trained Scene Understanding via World Models for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision...

2024
[29]

Nguyen, E.; Goel, K.; Gu, A.; Downs, G.; Shah, P.; Dao, T.; Baccus, S.; and R \'e , C. 2022. S4nd: Modeling images and videos as multidimensional signals with state spaces. 35: 2846--2861

2022
[30]

Pan, M.; Zhu, X.; Zheng, Y.; Wang, Y.; and Yang, X. 2023. Model-Based Reinforcement Learning with Isolated Imaginations. IEEE Transactions on Pattern Analysis and Machine Intelligence

2023
[31]

Robine, J.; H \"o ftmann, M.; Uelwer, T.; and Harmeling, S. 2023. Transformer-based world models are happy with 100k interactions. arXiv preprint arXiv:2303.07109

2023 arXiv
[32]

Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

2022
[33]

Schrittwieser, J.; Antonoglou, I.; Hubert, T.; Simonyan, K.; Sifre, L.; Schmitt, S.; Guez, A.; Lockhart, E.; Hassabis, D.; Graepel, T.; et al. 2020. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588(7839): 604--609

2020
[34]

L.; and Abbeel, P

Seo, Y.; Lee, K.; James, S. L.; and Abbeel, P. 2022. Reinforcement learning with action-free pre-training from videos. In International Conference on Machine Learning, 19561--19579. PMLR

2022
[35]

Shi, Y.; Dong, M.; and Xu, C. 2024. Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model. arXiv preprint arXiv:2405.14174

2024 arXiv
[36]

T.; Warrington, A.; and Linderman, S

Smith, J. T.; Warrington, A.; and Linderman, S. W. 2022. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933

2022 arXiv
[37]

Sun, G.; Liang, W.; Dong, J.; Li, J.; Ding, Z.; and Cong, Y. 2024. Create your world: Lifelong text-to-image diffusion. IEEE Transactions on Pattern Analysis and Machine Intelligence

2024
[38]

Sutton, R. S. 1991. Dyna, an integrated architecture for learning, planning, and reacting. ACM Sigart Bulletin, 2(4): 160--163

1991
[39]

N.; Kaiser, .; and Polosukhin, I

Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. 30

2017
[40]

M.; Mathieu, M.; Dudzik, A.; Chung, J.; Choi, D

Vinyals, O.; Babuschkin, I.; Czarnecki, W. M.; Mathieu, M.; Dudzik, A.; Chung, J.; Choi, D. H.; Powell, R.; Ewalds, T.; Georgiev, P.; et al. 2019. Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782): 350--354

2019
[41]

Wang, J.; Zhu, W.; Wang, P.; Yu, X.; Liu, L.; Omar, M.; and Hamid, R. 2023. Selective Structured State-Spaces for Long-Form Video Understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 6387--6397

2023
[42]

Wang, Y.; He, J.; Fan, L.; Li, H.; Chen, Y.; and Zhang, Z. 2024. Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 14749--14759

2024
[43]

Wu, J.; Ma, H.; Deng, C.; and Long, M. 2024. Pre-training contextualized world models with in-the-wild videos for reinforcement learning. 36

2024
[44]

N.; Gu, J.; and Rush, A

Yan, J. N.; Gu, J.; and Rush, A. M. 2024. Diffusion Models Without Attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 8239--8249

2024
[45]

Ye, W.; Liu, S.; Kurutach, T.; Abbeel, P.; and Gao, Y. 2021. Mastering atari games with limited data. 34: 25476--25488

2021
[46]

Zhang, W.; Wang, G.; Sun, J.; Yuan, Y.; and Huang, G. 2024. STORM: Efficient stochastic transformer based world models for reinforcement learning. 36

2024
[47]

Zhu, L.; Liao, B.; Zhang, Q.; Wang, X.; Liu, W.; and Wang, X. 2024. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417

2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Alonso, E.; Jelley, A.; Micheli, V.; Kanervisto, A.; Storkey, A.; Pearce, T.; and Fleuret, F. 2024. Diffusion for World Modeling: Visual Details Matter in Atari. arXiv preprint arXiv:2405.12399

arXiv 2024

[4] [4]

L.; Kiros, J

Ba, J. L.; Kiros, J. R.; and Hinton, G. E. 2016. Layer normalization. arXiv preprint arXiv:1607.06450

arXiv 2016

[5] [5]

G.; Naddaf, Y.; Veness, J.; and Bowling, M

Bellemare, M. G.; Naddaf, Y.; Veness, J.; and Bowling, M. 2013. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47: 253--279

work page 2013

[6] [6]

Chen, C.; Wu, Y.-F.; Yoon, J.; and Ahn, S. 2022. Transdreamer: Reinforcement learning with transformer world models. arXiv preprint arXiv:2202.09481

arXiv 2022

[7] [7]

Deng, F.; Park, J.; and Ahn, S. 2024. Facing off world model backbones: Rnns, transformers, and S4. 36

work page 2024

[8] [8]

Ding, Z.; Zhang, A.; Tian, Y.; and Zheng, Q. 2024. Diffusion world model. arXiv preprint arXiv:2402.03570

arXiv 2024