Sequential World Model (SeqWM)

• SeqWM is a sequential framework that models environment dynamics for decision making in partially observable settings, LLM agents, and multi-robot cooperation.
• It leverages ordered latent representations and sequential communication to support long-term memory, decision-aware supervision, and efficient multi-agent planning.
• Empirical results show SeqWM enhances performance through improved long-horizon imagination, reduced reconstruction overhead, and scalable coordination among agents.

Sequential World Model (SeqWM) is a sequential formulation of world modeling used for sequential decision making, partially observable model-based reinforcement learning, and multi-robot cooperation. In the recent literature, the term denotes both a general SeqWM-style framing—where a policy learns a model of environment dynamics to support future actions—and a specific model-based multi-agent reinforcement learning framework that decomposes joint dynamics into sequentially structured agent-wise world models (Cai et al., 24 Jun 2026, Deng et al., 2023, Zhao et al., 16 Sep 2025). Across these uses, SeqWM is characterized by modeling dynamics over temporally ordered inputs, using that model for prediction or imagination, and aligning the learned dynamics with downstream control.

1. Terminological scope and research context

The current SeqWM literature spans at least three distinct but related formulations. One line studies SeqWM-style world models in partially observable MBRL, with emphasis on long-term memory and latent imagination. A second line studies SeqWM-style world models for LLM agents in text environments and argues that the usual next-observation target is misaligned with decision making. A third line uses Sequential World Model (SeqWM) as the proper name of a model-based MARL framework for multi-robot cooperation (Deng et al., 2023, Cai et al., 24 Jun 2026, Zhao et al., 16 Sep 2025).

Formulation	Domain	Defining property
S4WM	Partially observable MBRL	Parallelizable SSM backbone for long-range history encoding
AAWM as decision-aware SeqWM	LLM agents in text environments	Agent-authored supervision aligned with pre-action information needs
SeqWM	Multi-robot cooperation	Independent agent-wise world models with sequential communication

In the MBRL setting, world models are described as a fundamental component for performing temporally extended and consistent simulations of the future in partially observable environments. In the LLM-agent setting, the central issue is not only whether a model predicts transitions, but whether the target reflects what the policy needs before acting. In multi-robot cooperation, SeqWM addresses the complexity of joint dynamics by replacing centralized fused prediction with a sequential chain of agent-wise predictors.

A common misconception is that SeqWM names a single architecture. The literature instead uses it in a broader sequential sense. S4WM is a latent-variable world model built around a parallelizable state-space sequence backbone; AAWM is presented as a decision-aware SeqWM for textual sequential decision making; and SeqWM in robotics is an autoregressive multi-agent framework with independent per-agent world models.

2. Sequential memory and latent dynamics in partially observable MBRL

In partially observable environments, the agent cannot infer the true state from a single frame and must integrate information over time. The long-memory formulation in "Facing Off World Model Backbones: RNNs, Transformers, and S4" identifies four capabilities that a sequential world model should support: long-term imagination, context-dependent recall, reward prediction, and memory-based reasoning (Deng et al., 2023).

The paper contrasts three backbone families. RNN-based world models such as RSSM maintain a single deterministic hidden state and support efficient autoregressive generation, but have limited long-term memory and sequential training. Transformer world models improve long-range dependency modeling through self-attention and parallel training, but incur quadratic attention cost and expensive long-horizon imagination. S4, as a Structured State Space Model, is presented as combining parallel training like Transformers with recurrent single-step generation like RNNs and sub-quadratic sequence complexity.

S4WM implements this idea as a latent-variable world model:

$$p(x_{1:T} \mid x_0, \mathbf{a}_{1:T}) = \int p(z_0 \mid x_0) \prod_{t=1}^T p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) \, p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) \, dz_{0:T}.$$

The backbone consumes

$$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$

and computes a history embedding through PSSM blocks in either parallel or single-step mode:

$$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$

$$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$

The prior and observation model are then

$$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$

$$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$

Training maximizes an ELBO with factorized posterior

$$q(z_{0:T} \mid x_{0:T}, \mathbf{a}_{1:T}) = \prod_{t=0}^T q(z_t \mid x_t),$$

using KL balancing with $\alpha = 0.8$, and the latents are discrete categorical variables optimized with straight-through gradients. The key practical point is that S4WM imagines in latent space rather than pixel space. This reduces the burden on the sequence backbone, which carries compact latent summaries, while the decoder reconstructs high-dimensional image observations.

The broader implication for SeqWM-style modeling is architectural rather than domain-specific: a sequential world model benefits from compact history encoding, support for both parallel learning and stepwise imagination, and a memory mechanism that does not collapse long histories into a weak recurrent summary.

3. Decision-aware supervision for LLM-agent world models

For LLM agents in partially observed text environments, SeqWM-style modeling is explicitly reframed as a problem of target selection in "Beyond Next-Observation Prediction: Agent-Authored World Modeling for Sequential Decision Making" (Cai et al., 24 Jun 2026). The paper studies a standard sequential decision setting in which the policy receives a textual decision context $o_t$, samples an action $a_t \sim \pi_\theta(\cdot \mid o_t)$, and observes $$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$0. Trajectories are

$$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$1

with one-step transition records

$$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$2

The paper’s central claim is that prediction quality and decision quality are not the same objective. Standard next-observation prediction is written as

$$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$3

where $$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$4 is derived from the next observation. This target is environment-driven: it reflects what the transition exposes, not what the policy needs to decide. The paper argues that such supervision can miss persistent latent state, action preconditions, valid/invalid action constraints, mechanism-level dynamics, and information spread across multiple trajectories.

AAWM replaces this target with agent-authored supervision. It maintains a Global transition pool $$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$5 containing all transition records from all trajectories and an Immediate transition set $$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$6 containing transitions from the same context $$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$7, including the logged transition and additional transitions obtained by executing alternative actions from that state. At each state, the policy is prompted to externalize its internal model into

$$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$8

where $$\mathbf{g}_t = \text{MLP}(\text{concat}[z_{t-1}, \mathbf{a}_t])$$9 are confirmed patterns and $$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$0 are open questions. In the main setup, the paper uses $$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$1 propositions per category.

Evidence is then retrieved across trajectories:

$$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$2

Each proposition is treated as an independent query; transitions are embedded with Qwen3-Embedding-0.6B; indexed with FAISS inner-product search over normalized embeddings; and selected with maximal marginal relevance (MMR). The reported hyperparameters are recall pool $$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$3, final hits per query $$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$4, and MMR trade-off $$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$5. Same-episode overlapping transitions are excluded when querying from a given state.

The final target is synthesized by an external instruction-following model:

$$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$6

The synthesis is instructed to correct unsupported beliefs, answer open questions when possible, summarize dynamics relevant to action selection, avoid merely repeating observations, and not propose actions or plans. The resulting dataset is

$$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$7

with fine-tuning objective

$$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$8

Within the SeqWM framing, this is a shift from environment-driven reconstruction to agent-driven dynamics modeling. The target is authored from the policy’s beliefs and uncertainties rather than mechanically derived from $$(\text{parallel}) \quad \mathbf{h}_{1:T}, \mathbf{s}_T = \text{PSSM\_Blocks}(\mathbf{g}_{1:T}, \mathbf{s}_0),$$9. The paper therefore presents AAWM as a decision-aware SeqWM: it still learns a world model over transitions, but the supervision is aligned with pre-action information needs.

4. Sequential communication and autoregressive decomposition in multi-robot cooperation

In cooperative robotics, SeqWM is introduced as a model-based MARL framework that integrates the distributed sequential paradigm into world modeling and planning (Zhao et al., 16 Sep 2025). The motivation is stated against two prior families. Early decentralized world models learned independent world models per agent but typically ignored inter-agent couplings. Centralized world models modeled joint dynamics in a single large model, often with transformers or GNNs and assumptions of full observability or unrestricted communication, but were argued to suffer from modeling complexity, communication overhead, and deployment difficulty.

SeqWM adopts the distributed sequential paradigm, in which agents communicate and act in a fixed order and each agent conditions on its predecessors. The paper explicitly emphasizes that this turns multi-robot cooperation into an autoregressive process. Each agent maintains its own encoder $$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$0, dynamics model $$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$1, reward predictor $$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$2, critic $$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$3, and actor $$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$4, with no parameter sharing:

$$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$5

At each time step, each agent’s observation-action pair $$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$6 is treated as a token, producing the sequence

$$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$7

from which the model predicts

$$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$8

Prediction is autoregressive across agents: agent $$(\text{single step}) \quad \mathbf{h}_t, \mathbf{s}_t = \text{PSSM\_Blocks}(\mathbf{g}_t, \mathbf{s}_{t-1}).$$9 predicts first, then $$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$0 predicts conditioned on $$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$1’s prediction, and so on.

The communication mechanism is likewise sequential:

$$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$2

In the planner appendix, the message is

$$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$3

Thus agent $$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$4 conditions on the predicted latent states and planned actions of all predecessors. The paper calls this explicit intention sharing.

For agent $$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$5, the world-model loss is

$$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$6

where the communication channel is not backpropagated through. The actor provides initial action estimates and is trained via HASAC, while planning uses a sequential MPPI-style procedure. The rollout value is

$$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$7

and the planner iteratively updates its action distribution from elite trajectories. The reported default planner settings are rollout horizon $$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$8, planning iterations $$p(z_t \mid z_{<t}, \mathbf{a}_{\le t}) = \text{MLP}(\mathbf{h}_t),$$9, temperature $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$0, sampled actions $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$1, elites $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$2, and actor samples $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$3.

SeqWM also removes explicit raw-observation reconstruction and predicts in latent space only. Auxiliary components include SEM Norm with $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$4, two-hot distributional reward/Q regression, symlog/symexp, and percentile scaling of the critic. Communication cost is reported as linear in the number of agents: centralized or fully connected communication is contrasted as $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$5, whereas the SeqWM chain structure is $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$6. The introductory comparison gives MAT: $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$7, MAZero: $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$8, MARIE: $$p(x_t \mid z_{\le t}, \mathbf{a}_{\le t}) = \mathcal{N}(\hat{x}_t, \mathbf{1}), \qquad \hat{x}_t = \text{Decoder}(\text{concat}[\mathbf{h}_t, z_t]).$$9, SeqComm: $$q(z_{0:T} \mid x_{0:T}, \mathbf{a}_{1:T}) = \prod_{t=0}^T q(z_t \mid x_t),$$0, and SeqWM: $$q(z_{0:T} \mid x_{0:T}, \mathbf{a}_{1:T}) = \prod_{t=0}^T q(z_t \mid x_t),$$1.

5. Empirical performance, behavior, and deployment

The empirical record for SeqWM-style models is domain-specific but internally consistent: sequential structure improves either memory, decision alignment, or cooperation, depending on the setting.

In partially observable MBRL, S4WM is reported to be strongest on long-term imagination. On Four Rooms, generation MSE is 219.4 for RSSM-TBTT, 224.4 for TSSM-XL, and 44.0 for S4WM. On Two Rooms, the corresponding values are 62.2, 62.9, and 27.3. On Ten Rooms, they are 323.1, 360.4, and 224.4. On Distracting Memory, S4WM achieves 100% inference and 100% imagination accuracy across widths 100, 200, and 400. On Multi Doors Keys, S4WM is reported to be dramatically better; for Three Keys, generation MSE is 5.16 for RSSM-TBTT, 1.27 for TSSM-XL, and 0.04 for S4WM (Deng et al., 2023).

For LLM agents, AAWM is evaluated on ALFWorld and WebShop with Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct, using Implicit World Modeling (IWM) as the main baseline, full-parameter supervised fine-tuning for world modeling, lightweight imitation learning on 32 trajectories, and GRPO for RL rows. With imitation learning plus RL at 1.5B, the reported numbers are ALFWorld: Base 74.0, IWM 76.8, AAWM 83.1 and WebShop: Base 59.4, IWM 67.2, AAWM 73.4. At 7B, they are ALFWorld: Base 84.6, IWM 86.7, AAWM 90.1 and WebShop: Base 70.6, IWM 73.2, AAWM 76.6. The paper highlights gains of up to +6.3 success points on one environment setting and +6.2 on the other. Harder ALFWorld tasks requiring multi-step preconditions and persistent state reasoning—Cool, Look, and Pick2—benefit most. In AgentGym across TextCraft, SciWorld, WebShop, and ALFWorld, AAWM is the only world-modeling initialization that improves over Base in every environment, with aggregate weighted scores Base 61.1, IWM 59.6, IWM-Summary 61.4, and AAWM 69.3. Judge-based analysis reports decision helpfulness of 0.96 on ALFWorld and 1.00 on WebShop, with open questions resolved at 0.96 and 0.99 respectively (Cai et al., 24 Jun 2026).

For multi-robot cooperation, SeqWM is evaluated on Bi-DexHands and Multi-Quad, and compared with HASAC, MARIE, MAT, and MAPPO. The paper reports that SeqWM outperforms all baselines in both final performance and sample efficiency. In Bi-DexHands, it reaches near-optimal performance in roughly 2–4 million steps on tasks such as Over, CatchOver2Underarm, and Scissors, and is the only method that successfully solves BottleCap. In Multi-Quad, it reaches near-100% success in Gate and Shepherd early in training and is the only method that solves PushBox. Qualitative behaviors include predictive adaptation in CatchOver2Underarm, temporal alignment in Pen, role division in BottleCap and PushBox, yielding and queue formation in Gate, and predictive coordination in Shepherd (Zhao et al., 16 Sep 2025).

SeqWM is also reported to transfer to physical robots. The deployment uses two Unitree Go2-W quadruped robots in an indoor arena of size $$q(z_{0:T} \mid x_{0:T}, \mathbf{a}_{1:T}) = \prod_{t=0}^T q(z_t \mid x_t),$$2, with 8 Mars motion-capture cameras and a NOKOV 3D motion capture system. Real-world tasks include PushBox, Gate, and Shepherd. Sim-to-real enhancements include observation transformation into ego-centric coordinates, domain randomization for initial poses, target positions, and friction, and sensor and actuation perturbations via noise and delays. The observed real-world behaviors mirror simulation: one robot pushes while the other steers in PushBox; robots yield smoothly in Gate; and one robot advances while the other adapts in Shepherd.

6. Interpretive issues, misconceptions, and future directions

The literature supports several clarifications. First, SeqWM should not be reduced to next-observation prediction. The LLM-agent work makes this point explicitly: decision-aware world-model targets provide a more effective learning signal than next-observation prediction, and world model targets should be decision-aware, not just predictive (Cai et al., 24 Jun 2026). This is an objective-level critique rather than an architectural one.

Second, SeqWM should not be identified with raw observation reconstruction. S4WM performs latent imagination for high-dimensional image sequences, and the multi-robot SeqWM removes explicit raw-observation reconstruction and predicts in latent space only (Deng et al., 2023, Zhao et al., 16 Sep 2025). AAWM goes further by making the target a natural-language paragraph that captures decision-relevant dynamics rather than observation reconstruction.

Third, SeqWM is not synonymous with centralized joint modeling. In cooperative robotics, sequential prediction is reported to recover much of the benefit of centralized modeling while avoiding the same communication burden: the ablation comparing sequential, centralized, and decentralized world models finds that sequential and centralized models achieve similarly low prediction errors, and both clearly outperform the decentralized model (Zhao et al., 16 Sep 2025).

Fourth, SeqWM is not tied to a single sequential backbone. The S4WM paper defines a general PSSM interface and explicitly instantiates the framework with S5 as S5WM, which performs even better than S4WM on long-term imagination (Deng et al., 2023). This indicates that the sequential world-model formulation is backbone-agnostic within the class of parallelizable state-space models.

Several forward-looking directions are stated in the source material. One is to re-run self-probing after each training round so that the AAWM target evolves with the policy’s understanding, turning AAWM into a curriculum-like adaptive SeqWM (Cai et al., 24 Jun 2026). Another is the use of a full-history posterior

$$q(z_{0:T} \mid x_{0:T}, \mathbf{a}_{1:T}) = \prod_{t=0}^T q(z_t \mid x_t),$$3

which can improve performance on longer sequences at higher computational cost (Deng et al., 2023). In multi-robot planning, the paper reports that KL-based early stopping reduces execution time by about 27.3% with only about 5.9% performance loss at threshold 0.5, suggesting a practical efficiency-accuracy trade-off for deployment (Zhao et al., 16 Sep 2025).

Taken together, these works define SeqWM as a family of sequentially organized world-modeling strategies rather than a single recipe. In MBRL, its central problem is long-horizon memory; in LLM agents, it is supervision aligned with pre-action information needs; and in cooperative robotics, it is sequential decomposition of joint dynamics with explicit intention sharing.