acceptodds
Under review as a conference paper at ICLR 2027

Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term Memory

Abstract

Long-video generation and world models have shown strong potential for interactive entertainment and embodied simulation by predicting future observations conditioned on user actions and historical memory. However, as memory sequences grow longer and their structures become increasingly complex, managing long-range spatial context becomes increasingly challenging, calling for a more intelligent and systematic memory-management strategy. Building on the advancing spatial reasoning capabilities of multimodal large language models (MLLMs) and the broader vision of unified models, we propose Spatial Memory Intelligence (SMI), the first framework to systematically employ an understanding model for spatial-memory management in long-video world models. SMI introduces four coordinated atomic operations: spatial clustering, within-cluster sparsification, action-aware retrieval, and reliability-aware filtering. Extensive experiments across multiple baselines, benchmarks, and world-model backbones demonstrate the effectiveness and generalizability of SMI, achieving comprehensive improvements in memory sparsity, generation stability, and spatial consistency.

Primary area:
generative models
Keywords:
World Model, Spatial Memory, Multimodal Understanding

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FTHaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduce a Spatial Memory Intelligence (SMI) framework for controlling the long-horizon video context for video generation model. SMI is composed of MI four coordinated atomic operations: spatial clustering, within-cluster sparsification, action-aware retrieval, and reliability-aware filtering.

Critical strengths
  1. This paper is clear.
  2. The approach are intuitive, for example the coordinated atomic operations: spatial clustering, within-cluster sparsifica- tion, action-aware retrieval, and reliability-aware filtering.
  3. Demonstrated across two models on VBench showcase the effectiveness of SMI.
Critical weaknesses
  1. The claim of first framework to systematically employ an understanding model for spatial-memory management in long-video world models is too strong, there are many related works.
  2. The approach is very heuristic.
  3. The models tested are very old.
  4. Very limited benchmarks are evaluated.
Questions for authors
  1. How to generalize your approach to industrial scenarios?
3 comments

Thank you for your comments.

W1 Novelty Claim

Our novelty claim focuses on three aspects: (1) Task scope: Spatial-memory management in long-horizon action-conditioned video world models. (2) Comprehensiveness: A unified framework systematically addressing spatial consistency, memory sparsification, and generation stability. (3) Trainable paradigm: Dedicated training datasets and a trained understanding model for memory management, with potential for efficient and scalable deployment.

W2 Data Driven Memory Management and Scalability

Our method learns memory-management decisions from data, providing a scalable framework for memory management. In contrast, heuristic retrieval strategies rely primarily on predefined rules, such as FoV-based retrieval in Context as Memory and WorldMem, and 3D geometric retrieval in VMem.

To validate the scalability of our method, We additionally use GPT6-Astra to systematically perform all four operations for spatial memory management. We evaluate it on 20 scenes. The quantitative comparisons with Base and SMI(Qwen) are shown below, and qualitative comparisons have been added to the updated Figure 8 in the main paper. We also provide qualitative visualizations via this anonymous link: https://anonymous.4open.science/w/memory-video-comparisons/ .

MethodPSNR ↑LPIPS ↓Background Consistency ↑Motion Smoothness ↑Aesthetic Quality ↑Camera Consistency ↑Overall Visual Quality ↑Memory Bank Sparsity
Base11.40580.62580.94660.99210.665447.6150.670.00%
SMI(Qwen)12.24790.56290.95610.99260.682967.3370.2182.79%
SMI(GPT6-Astra)12.88050.51700.95900.99290.693276.5273.6786.33%

SMI(Qwen) uses the trained Qwen model, whereas SMI(GPT6-Astra) uses Astra without task-specific training.

The results show that Astra’s stronger spatial reasoning capabilities enable substantial improvements. We will also release the implementation details of SMI(GPT6-Astra) to support reproducibility.

However, Astra incurs higher inference latency, approximately 3.17 times that of Base, and a high token cost, totaling several billion tokens. Using a smaller model for memory management therefore also offers clear advantages for on-device deployment.

We also use GPT-6.1-sol to evaluate retrieval accuracy across all cases for geometry-only retrieval, geometry combined with semantics, SMI(Qwen), and SMI(GPT6-Astra), as shown below.

MethodCorrectIncorrectUncertainOverall Accuracy
Geometry-only Base*20619711739.62%
Geometry + Semantics19716415937.88%
SMI(Qwen)369965570.96%
SMI(GPT6-Astra)49122794.42%

These results demonstrate the framework's flexibility across understanding models and the benefits of stronger spatial reasoning for memory retrieval.

W3 Choice and Release Dates of the Evaluated Models

The HY1.5 and Wan2.2 models in WorldPlay-1.5 were open-sourced in January 2026. We began this research in February 2026, when very few open-source world models were available, and even fewer supported memory. We therefore consider these backbones a reasonable choice for evaluating our method.

W4 Evaluation Benchmarks

We combine reconstruction metrics (PSNR and LPIPS), as used in WorldPlay, Context as Memory , and WorldMem, with VBench, GPT-based assessment, and a user study. These four complementary evaluation protocols assess spatial consistency and visual quality beyond reconstruction alone. We believe this provides a comprehensive evaluation that captures the key dimensions relevant to long-horizon world models.

Q1 Generalization to Industrial Scenarios

We believe that integrating understanding and generation within a unified architecture is a promising direction for practical world models. SMI represents a step toward this goal by introducing a trainable understanding-driven memory manager. However, industrial deployment still requires further improvements in latency, efficiency, and integration with generative models. Exploring end-to-end optimization of understanding and generation is therefore an important direction for future work.

Dear Reviewer FTHa,

Thank you again for your comments. Our additional experiments demonstrate the scalability and advantages of our method, as well as its effectiveness in long-horizon generation. We hope that these results and the accompanying clarifications address your concerns, and that our work offers the community a new perspective on spatial-memory management.

We sincerely appreciate your time and effort in reviewing our paper and look forward to your further feedback.

Thank the author for their response. But as illustrated in my weakness 1. "many recent related work" are still missing for comparison.

Reviewer_1TUdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Spatial Memory Intelligence (SMI), an MLLM-driven framework for long-term spatial memory management in video world models. SMI integrates four operations: spatial clustering, within-cluster sparsification, action-aware retrieval, and reliability-aware filtering. Experiments on two world-model backbones demonstrate improved spatial consistency and generation quality while achieving over 80% memory sparsity.

Critical strengths
  1. Addressing memory management in video world models is an important and interesting research direction.

  2. The proposed method achieves substantial memory sparsification while improving spatial consistency and generation quality across two world-model backbones.

Critical weaknesses
  1. The necessity of MLLM-based memory management is not sufficiently justified. While SMI emphasizes the spatial understanding capabilities of MLLMs as a key contribution, the proposed framework also relies heavily on geometric metadata, such as camera poses, FoV overlap, and geometric coverage. However, the paper does not provide direct comparisons against simpler rule-based approaches to clustering, pruning, and retrieval that leverage the same geometric information. As a result, it remains unclear to what extent the observed improvements can be attributed to the MLLM's spatial reasoning capabilities. Although the motivation is well-founded, it is difficult to rule out the possibility that substantially simpler heuristics could achieve comparable performance.

  2. The practical benefits of improved memory sparsity remain unclear. Although SMI achieves over 80% memory sparsity, the original WorldPlay framework already uses only a limited number of historical chunks as generation context. Therefore, reducing the number of stored chunks does not necessarily translate into lower VRAM usage or computational costs, and the benefits may be largely limited to savings in relatively inexpensive external storage. Meanwhile, SMI increases end-to-end latency by 31–38%, which is a significant concern for interactive world models. Furthermore, most experiments are conducted on approximately one-minute videos, making it difficult to assess whether the benefits of memory sparsification justify the additional computational overhead over longer horizons. The proposed approach seems particularly promising for substantially longer video generation, yet the paper does not provide experiments examining its behavior or scalability in such settings.

  3. There are concerns regarding the suitability of the evaluation metrics and the consistency of the reported improvements. The VBench metrics, including background consistency, motion smoothness, and aesthetic quality, primarily measure general video quality rather than long-term spatial memory capabilities. It is also unclear whether the relatively small numerical differences from competing baselines on these metrics are practically meaningful. Similarly, reconstruction PSNR and LPIPS rely on the similarity between corresponding generated frames, which may not fully capture spatial consistency. While the GPT-based evaluation appears relatively more relevant to the paper's objectives, the results are not consistently favorable across backbones. SMI achieves the best performance on all nine evaluation dimensions for HY1.5, but only four out of nine for Wan2.2. In particular, the improvement in camera-constrained scene consistency reported in the main table is marginal for Wan2.2, with a difference of only 0.075 points compared to Deep Forcing.

Questions for authors
  1. Could the authors provide a comparison against a simple rule-based memory management pipeline that uses the same geometric metadata (e.g., camera poses, FoV overlap, and geometric coverage) for clustering, pruning, and retrieval? Such a comparison would help clarify whether the observed improvements genuinely require MLLM-based spatial reasoning or could be achieved using simpler geometric heuristics with careful hyperparameter tuning.

  2. Could the authors provide results on substantially longer rollouts (e.g., 5–10 minutes)? It would be particularly informative to examine how memory size, generation quality, and inference latency scale with video length, and whether the benefits of memory sparsification become more pronounced over longer horizons.

4 comments

Thank you for your insightful comments.

W1 and Q1 Necessity of MLLMs and Fully Geometric Spatial Memory Management

To validate the necessity of using an MLLM, we first demonstrate the generalization and scalability.

We additionally use GPT6-Astra to systematically perform all four operations for spatial memory management. We evaluate it on 20 scenes. The quantitative comparisons with Base and SMI(Qwen) are shown below, and qualitative comparisons have been added to the updated Figure 8 in the main paper. We also provide qualitative visualizations via this anonymous link: https://anonymous.4open.science/w/memory-video-comparison/ .

Comparison of Memory Management Models

MethodPSNR ↑LPIPS ↓Background Consistency ↑Motion Smoothness ↑Aesthetic Quality ↑Camera Consistency ↑Overall Visual Quality ↑Memory Bank Sparsity
Base11.40580.62580.94660.99210.665447.6150.670.00%
SMI(Qwen)12.24790.56290.95610.99260.682967.3370.2182.79%
SMI(GPT6-Astra)12.88050.51700.95900.99290.693276.5273.6786.33%

SMI(Qwen) uses the trained Qwen model, whereas SMI(GPT6-Astra) uses Astra without task-specific training.

The results show that Astra’s stronger spatial reasoning capabilities enable substantial improvements. We will also release the implementation details of SMI(GPT6-Astra) to support reproducibility.

However, Astra incurs higher inference latency, approximately 3.17 times that of Base, and a high token cost, totaling several billion tokens. Using a smaller model for memory management therefore also offers clear advantages for on-device deployment.

In addition, we use GPT-6.1-sol to evaluate retrieval accuracy across all cases for geometry-only retrieval, geometry combined with semantics, SMI(Qwen), and SMI(GPT6-Astra), as shown below.

Retrieval Accuracy Comparison

MethodCorrectIncorrectUncertainOverall Accuracy
Base*20619711739.62%
Geometry + Semantics19716415937.88%
SMI(Qwen)369965570.96%
SMI(GPT6-Astra)49122794.42%

The results show that an understanding model can achieve higher retrieval accuracy by reasoning about spatial relationships when selecting memories. Retrieval accuracy further improves as spatial understanding becomes stronger, demonstrating the generalization and scalability of MLLM-based spatial reasoning.

We also implement a fully geometric pipeline for clustering, sparsification, and retrieval. The results are shown below.

MethodPSNR ↑LPIPS ↓Background Consistency ↑Motion Smoothness ↑Aesthetic Quality ↑Camera Consistency ↑Overall Visual Quality ↑Memory Bank Sparsity
Base11.40580.62580.94660.99210.665447.6150.670.00%
Geometry-only (90%)11.63850.61020.94310.99120.671348.6358.7165.83%
Geometry-only (50%)11.63420.61170.94290.99160.672448.2958.9367.61%
SMI(Qwen)12.24790.56290.95610.99260.682967.3370.2182.79%
SMI(GPT6-Astra)12.88050.51700.95900.99290.693276.5273.6786.33%

For within-cluster sparsification, we use two coverage thresholds: a middle memory item is removed when the coverage criterion reaches 90% or 50%, respectively. The two thresholds produce similar sparsity because, once a cluster contains enough items, coverage is usually high, so sparsification is typically triggered under either threshold.

Although the purely geometric method achieves some sparsification, it provides limited improvements in consistency and stability. The main reason is the limited generalization of this rule-based method, which still makes many retrieval errors.

We also consider geometric information indispensable in world models. In future work, we plan to explore feeding geometric information directly into the understanding model so that it can perform coarse spatial localization itself.

W2 and Q2 Practical Benefits of Memory Sparsity and Longer Rollouts

Thank you for this insightful suggestion. As you anticipated, our method has clear advantages for substantially longer generation.

We compare Base and SMI on ten sets of six-minute rollouts, reporting memory size, generation latency, and aesthetic quality for each minute, as shown below.

Method / MetricMinute 1Minute 2Minute 3Minute 4Minute 5Minute 6
Base (Memory Size, Chunks)90180270360450540
SMI (Memory Size, Chunks)12.412.212.212.212.212.2
Base (Latency)1.00×1.42×1.83×2.27×2.69×3.11×
SMI (Latency)1.36×1.27×1.27×1.26×1.26×1.26×
Base (Aesthetic Score)74.6865.0248.8345.1544.4439.97
SMI (Aesthetic Score)78.4475.5273.4768.0264.4762.59

As generation proceeds, Base’s memory bank grows without bound. At each retrieval step, it compares against all memory chunks, causing latency to increase linearly. By the sixth minute, latency reaches 3.11 times that of the first minute. SMI remains relatively stable and is even slightly faster than in its first minute.

SMI also maintains relatively stable generation quality over the six-minute rollouts, with a much smaller decline than Base.

These results demonstrate the significant benefits of sparsification for substantially longer video generation. We thank the reviewer again for this insightful suggestion.

W3 Suitability of the Evaluation Metrics and Consistency of the Improvements

We combine reconstruction metrics (PSNR and LPIPS), as used in WorldPlay, Context as Memory , and WorldMem, with VBench, GPT-based assessment, and a user study. These four complementary evaluation protocols assess spatial consistency and visual quality beyond reconstruction alone. We believe this provides a comprehensive evaluation that captures the key dimensions relevant to long-horizon world models.

Regarding Wan2.2, the base model itself has weaker consistency and is highly unstable during long-video generation, making it difficult to achieve the best results on every metric.

Nevertheless, our method improves consistency and stability while maintaining substantial memory sparsification, demonstrating the generalization of the framework.

Dear Reviewer 1TUd,

Thank you once again for your constructive comments. We would like to know if our responses have adequately addressed your concerns. We are also eager to hear any additional concerns or suggestions from you that could help us further strengthen our paper. Your feedback would be invaluable in making this work even more compelling.

As the rebuttal window is closing soon, we genuinely appreciate your feedback. We hope our response has addressed your concerns.

Reviewer_1TUdOpenReview ↗

Follow up

Dear Authors,

Thank you for the detailed response and for the additional experiments on Q1 and Q2. After reading the response and the revision, I will keep my rating. My main remaining concern is W3, which the response did not address with new analysis.

  • The VBench differences are reported without variance, so I cannot judge whether they are meaningful. On Wan2.2, SMI trails both Deep Forcing and MoC on long-range identity and on overall scene consistency, and its background consistency is below Base (0.917 vs. 0.930). A weaker backbone does not explain this, because all methods use the same backbone.
  • The new evidence relies on the same kinds of metrics. The six-minute study (Table 7) measures quality only with the aesthetic score, which does not reflect spatial memory. The retrieval accuracy in Table 6 comes from an LLM judge, with no definition of its labels and no validation against human judgments.

The new results point in a favorable direction. However, they do not resolve my W1 and W2 concerns, because the metrics behind them do not directly measure long-term spatial consistency. I encourage the authors to strengthen the evaluation in a future version.

Reviewer_pVZnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper utilizes the spatial reasoning capabilities of MLLMs and proposes SMI to employ an MLLM for spatial-memory management in long-video world models. SMI introduces four operations: spatial clustering, within-cluster sparsification, action-aware retrieval, and reliability-aware filtering. Their experimental results across multiple baselines and benchmarks demonstrate the effectiveness and generalizability of SMI.

Critical strengths

The paper presents a novel direction by systematically using an MLLM for spatial-memory management in long-video world models. The proposed framework is straightforward and integrates several memory management operations. The experiments are relatively comprehensive, showing promising improvements in spatial consistency and memory efficiency across different world model backbones.

Critical weaknesses
  1. The paper mainly motivates the use of MLLMs by their spatial reasoning capabilities, but only evaluates Qwen3.5-4B. It would be interesting to see how different MLLMs perform and whether stronger spatial reasoning capabilities actually lead to better memory management and generation consistency.

  2. I think more discussion and ablation studies are needed to understand the advantages of MLLM-based memory management over traditional methods. Different approaches may have their own strengths and weaknesses, and it is still unclear when MLLMs are more useful than geometric or semantic methods. Would combining these approaches be more effective than simply replacing traditional methods?

  3. The role of the MLLM is not very clear to me. From the current description, it seems that the MLLM mainly performs visual matching, ranking, and filtering. What specific spatial reasoning capabilities are needed for these operations? I also found Figure 2 a bit difficult to follow, especially how the MLLM contributes to more consistent generation.

  4. Most experiments seem to focus on camera movement in relatively static environments. It would be interesting to see how the method handles more dynamic scenarios, where objects move or their states change over time. This could introduce additional challenges for memory management and consistency.

Questions for authors

see my questions in weakness.

4 comments

Thank you for your insightful comments.

W1 Different MLLMs with Stronger Spatial Reasoning Capabilities

Thank you for your suggestion. We use GPT6-Astra to systematically perform all four operations for spatial memory management. We evaluate it on 20 scenes. The quantitative comparisons with Base and SMI(Qwen) are shown below, and qualitative comparisons have been added to the updated Figure 8 in the main paper. We also provide qualitative visualizations via this anonymous link: https://anonymous.4open.science/w/memory-video-comparisons/ .

MethodPSNR ↑LPIPS ↓Background Consistency ↑Motion Smoothness ↑Aesthetic Quality ↑Camera Consistency ↑Overall Visual Quality ↑Memory Bank Sparsity
Base11.40580.62580.94660.99210.665447.6150.670.00%
SMI(Qwen)12.24790.56290.95610.99260.682967.3370.2182.79%
SMI(GPT6-Astra)12.88050.51700.95900.99290.693276.5273.6786.33%

SMI(Qwen) uses the trained Qwen model, whereas SMI(GPT6-Astra) uses Astra without task-specific training.

The results show that Astra’s stronger spatial reasoning capabilities enable substantial improvements. We will also release the implementation details of SMI(GPT6-Astra) to support reproducibility.

However, Astra incurs higher inference latency, approximately 3.17 times that of Base, and a high token cost, totaling several billion tokens. Using a smaller model for memory management therefore also offers clear advantages for on-device deployment.

W2 Advantages of MLLMs and the Combination of Geometry and Semantics

We agree that different memory-management approaches have distinct strengths and weaknesses. However, the geometric and semantic baselines considered here rely on predefined retrieval criteria, which may not fully capture complex spatial relationships and semantic changes, potentially resulting in retrieval errors. MLLMs, by contrast, use visual and spatial understanding to manage and retrieve memories, and can learn these strategies from data, offering the potential for better generalization and scalability.

We use GPT-6.1-sol to evaluate retrieval accuracy across all cases for geometry-only retrieval, geometry combined with semantics, SMI(Qwen), and SMI(GPT6-Astra), as shown below.

Retrieval Accuracy Comparison

MethodCorrectIncorrectUncertainOverall Accuracy
Geometry-only Base*20619711739.62%
Geometry + Semantics19716415937.88%
SMI(Qwen)369965570.96%
SMI(GPT6-Astra)49122794.42%

The results show that an understanding model can achieve better retrieval accuracy by reasoning about spatial relationships when selecting memories. Retrieval accuracy further improves as spatial understanding becomes stronger, demonstrating the generalization and scalability of MLLM-based spatial reasoning.

We also provide an overall performance comparison with retrieval that combines geometry and semantics.

MethodPSNR ↑LPIPS ↓Background Consistency ↑Motion Smoothness ↑Aesthetic Quality ↑Camera Consistency ↑Overall Visual Quality ↑Memory Bank Sparsity
Geometry-only Base*11.40580.62580.94660.99210.665447.6150.670.00%
Geometry + Semantics11.21280.63310.94880.99150.669846.0248.280.00%
SMI(Qwen)12.24790.56290.95610.99260.682967.3370.2182.79%
SMI(GPT6-Astra)12.88050.51700.95900.99290.693276.5273.6786.33%

W3 Spatial Reasoning Required by the Four Operations

We clarify these roles below, with visual examples provided in Appendix A.

Retrieval requires correctly identifying which candidates best match the spatial relationships in the current scene structure, such as whether they correspond to the left, right, or rear, and selecting the appropriate candidates accordingly.

Clustering and sparsification require the MLLM to understand whether the current spatial structure provides unique spatial evidence relative to the cluster as a whole, or whether it is already covered by other members of the same cluster. Filtering requires understanding changes in spatial semantics and identifying whether the currently generated frames exhibit degradation.

Together, these four operations require the MLLM to understand spatial relationships between different structures and changes in spatial semantics.

W4 More Dynamic Scenarios

We agree that handling dynamic scenes is an interesting and important topic in spatial-memory management. We show several dynamic scenarios in the appendix and video demos. More complex dynamic scenarios remain challenging because the current base model also lacks the ability to handle dynamic consistency. Addressing this may require exploring whether a unified multimodal model (UMM) can learn from data to model such dynamic consistency as scenes evolve.

Dear Reviewer pVZn,

Thank you once again for your constructive comments. We would like to know if our responses have adequately addressed your concerns. We are also eager to hear any additional concerns or suggestions from you that could help us further strengthen our paper. Your feedback would be invaluable in making this work even more compelling.

As the rebuttal window is closing soon, we genuinely appreciate your feedback. We hope our response has addressed your concerns.

Thank you for the detailed responses and additional experiments. I appreciate the authors' efforts in addressing my concerns, and the new results strengthen the paper.

My remaining concern is that it is still unclear to what extent the improvements come specifically from spatial reasoning rather than general visual understanding and matching capabilities. I consider this an important question given the paper's central motivation, and I am not fully convinced that it has been adequately addressed.

Overall, while the rebuttal has improved my assessment, I am not sufficiently convinced to raise my rating to 3. I have no further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the reviewers for their constructive feedback and summarize our main responses below. New additions to the revised manuscript are highlighted in blue\color{blue}{\text{blue}}.

1. Model Scalability and Advantages of MLLM-Based Spatial Memory Management (pVZn and 1TUd).

To validate the Model Scalability of SMI, we additionally use GPT6-Astra and show that stronger spatial reasoning leads to better spatial memory management (Section 3.5.3, Table 5 and Figure 8). We also compare retrieval accuracy across understanding-driven and alternative strategies (Appendix C.1, Table 6). This analysis, together with overall performance comparisons against geometry-plus-semantics retrieval and a fully geometric pipeline, further demonstrates the advantages of MLLM-based spatial memory management.

2. Long-Horizon Scalability and Benefits of Memory Sparsification (1TUd).

To validate the benefits of sparsification, we compare SMI and Base on ten sets of six-minute rollouts (Appendix C.2, Table 7). SMI maintains stable per-minute generation time and more stable visual quality. By the sixth minute, generation time is 1.26× for SMI versus 3.11× for Base, both normalized to Base’s first-minute generation time.

3. Comprehensive Evaluation (1TUd and FTHa).

To comprehensively evaluate SMI, we combine reconstruction metrics (PSNR and LPIPS), as used in WorldPlay [1], Context as Memory [2], and WorldMem [3], with VBench, GPT-based assessment, and a user study. These four complementary protocols assess spatial consistency and visual quality beyond reconstruction alone (Section 3.2; Appendices E and F).

[1] Wenqiang Sun et al. WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling. ICML 2026.

[2] Jiwen Yu et al. Context as Memory: Scene-Consistent Interactive Long Video Generation with Memory Retrieval. SIGGRAPH Asia 2025.

[3] Zeqi Xiao et al. WorldMem: Long-term Consistent World Simulation with Memory. NeurIPS 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.