MindJourney: Test-Time Scaling with World Models for Spatial Reasoning

Dear Reviewer,

We thank you for the detailed comments and questions. We are glad to hear you find our method "well-executed and reasonable" and our evaluation "rigorous". We address each concern below:

Technical Novelty

We would like to clarify that our contribution does not lie in proposing a new video diffusion model, but rather in the test-time collaboration protocol that integrates a frozen VLM with a pose‑conditioned world model to enhance 3D spatial reasoning:

VLM‑guided imagination. SpatialNavigator, to our knowledge, is the first framework in which a VLM actively controls a world model at inference time to bolster its own reasoning.

Beam‑search coupling. The VLM/world-model loop is cast as beam search, balancing diversity and evidence scoring.

We will revise the introduction sections to highlight that our novelty lies in this inference-time coordination mechanism, which builds on, but is distinct from prior works using camera-conditioned video models.

World Model Limitations

We have already included additional analysis of the quality of our trained world model and SVC, and also their failure modes in the supplementary material (Sec. B). For the former, using the out-of-domain simulator AI2Thor, we evaluated the world model's fidelity under a sequence of atomic actions defined in our work. We report the PSNR, LPIPS, and SSIM in Tab. 1 of the supplemental. The SVC and SWM models obtain comparable SSIM scores but the latter outperforms the former by larger margins of ~2 and 0.18 for PSNR and SSIM. We refer all reviewers to Sec. B.1.1 for more information. To further analyze the reliance of these world models, we evaluate the performance when extended to multiple steps(1 to 3) in the following table.

Generally, we observe consistent trends across the SVC and SWM models, where their effectiveness gradually decrease as the number of steps increases. This is expected since the uncertainty of predicting the exact visual changes is compounded over steps.

Despite the current limitations of learned world models, SpatialNavigator consistently boosts VLM reasoning performance. Ongoing advances—illustrated by lighter, higher-fidelity backbones such as Wan2.1 and the controllable ReCamMaster—provide a clear trajectory for further progress. As these models mature, we expect them to unlock even greater downstream performance with SpatialNavigator.

Expand Evaluation Scope

With regards to expanding the evaluation scope to real 3D scenes, we have also demonstrated the effectiveness of our method on SAT-Real in Tab. 1 of the main paper, where it achieves performance gains consistent with prior observations. As additional evidence of the generalizability of our approach, we also provide evaluation results on MindCube, which tasks VLMs to perform multiview spatial reasoning in real 3D scenes. We only evaluated with SVC, as it supports video generation conditioned on multiview images, unlike SWM which currently lacks this capability.

Unlike static and single-view image tasks, the multiview setting in MindCube requires to understand 3D spatial layouts and cross-view consistency. The results further show the potential of our method in real-world and multiview settings. We observe that SpatialNavigator is able to improve accuracy by approximately 4 to 7% across different VLMs. Thus, SpatialNavigator enables VLMs to generalize to multiview settings by integrating visual cues from multiple imagined perspectives. This capability highlights its potential in embodied AI and robotic tasks. Finally, the results suggest that SpatialNavigator is model-agnostic and can further improve with advances in better world models as well as VLMs. We will include these results in the next version of the paper to make our evaluations more comprehensive.

Computational Efficiency

We appreciate the reviewer’s suggestion and agree that analyzing computational trade-offs is important for practical deployment. To address this, we provide a structured analysis of the current computational efficiency of SpatialNavigator.

Also, we would like to clarify that SpatialNavigator is the first work on test-time scaling framework that uses VLM to control a video diffusion model. Thus, we mainly focused on optimizing the effectiveness instead of efficiency, we acknowledge that efficiency optimization is a promising direction for future work. To address this, we outline several acceleration techniques that make further speed-ups highly plausible.

Efficiency Analysis

We profiled computational efficiency on a single H100-80G. We evaluated using gpt4o+SN(SWM), with beam size 2 and inference step 50 on SAT-Real. The memory consumption is~40G across the experiments. As SVC cam only generates high resolution video and therefore not efficient(around x3 world model latency), we mainly use SWM in this section.

In our main experiments, we focused on optimizing the effectiveness so the default setting is not the most cost-effective setting. To explore more cost-optimized settings, we further investigate reducing the number of inference step in the table below:

The results demonstrate that our model can maintain a >95% of the accuracy gain at ≤33% of the latency, with a latency of about 23s. We tested o1 on the SAT-Real with latency in average 19.6s, so SpatialNavigator in cost-optimized setting is in the same latency band as widely used test-time-scaling pipelines.

We further added the ablation study on beam size in the table below:

Scalability & Acceleration Pathways

Backbone Model Recent released Wan2.1-1.3B provides videos with quality higher than our backbone CogVideoX-5B in ≈8GB VRAM. Wan2.1 achieved ≈1.7×lower (8.7s vs.14.5s) latency and ≈½ the memory footprint relative our current backbone tested under identical frame counts and resolution,

Step Distillation Latent Consistency Models produce high-fidelity outputs in 2–4 inference steps, yielding 10–20× speed-ups over 50-step schedulers. Video-specific extensions such as Distribution-Matching Distillation report similar benefits on text-to-video transformers.

Low-Bit Quantization Post-training methods like SVDQuant quantize both weights and activations to 4-bit while keeping FID within ±2 points of full precision. Empirical results report ≈3× speedups and up to 3.6× memory reduction on DiTs, enabling real-time sampling on a 16GB laptop-class GPU.

All three techniques are open-source and multually compatible, which indicates that improving computational cost is both practical and well-supported by the community.

Missing Baselines

For additional baselines, we report results for (i)a recent 3D-aware VLM VLM3R [1] and (ii)end-to-end fine-tuning on SAT-Real/SAT-Synth.

3D-aware VLMs VLM3R was evaluated under the same zero-shot protocol used for our method (SpatialNavigator, SN).

Despite being explicitly designed for 3D spatial reasoning, VLM3R lags behind both the GPT-4o baseline and our test-time scaling scheme, underscoring the effectiveness of coupling a strong VLM with a controllable world model at inference time.

End-to-End Fine-Tuning Fine-tuning requires task-specific training data, whereas our approach is strictly zero-shot, which makes the comparison unfair. Nevertheless, for context we refer the finetuning baselines from the SAT paper with our numbers:

The absolute gains delivered by test-time scaling(+8.7%) are on par with supervised fine-tuning(+9~13%), while preserving the zero-shot assumption and avoiding additional training costs. [1] Z, Fan., et al. VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction. arXiv preprint, 2025.

Other Questions

World Model quality in outdoor scenes Assessing world-model fidelity outdoors would require using a dedicated outdoor simulator to render GT image sequences for atomic actions, which require engineering efforts to set up. As the outdoor environments are inherently unstructured (e.g., vegetation, variable lighting), generalizing to outdoor scene will be comparatively challenging. Although SpatialNavigator also boost performance on SAT-Real with outdoor scenes, this study targets more on structured indoor domain.

Top 3 Failure Mode We manually inspected part of the rollouts and the top 3 failure modes are B.1.1 Inaccurate Forward Movement, B.1.4 Visual Artifacts, and B.1.5 Scene Misinterpretation mentioned in the supplimentary material.

Human Performance Gap Humans score 92.8% on SAT-Synth.—≈15% above our best model—signaling significant headroom for future work.

Dear Reviewer Pkry,

We appreciate your detailed comments and insightful suggestions. We are glad to hear you find our idea "thought-provoking" and our experiments "comprehensive". Here are the responses to your questions.

Computational Efficiency

We appreciate the reviewer’s suggestion and agree that analyzing computational trade-offs is important for practical deployment. To address this, we provide a structured analysis of the current computational efficiency of SpatialNavigator.

Also, we would like to clarify that SpatialNavigator is the first work on test-time scaling framework that uses VLM to control a video diffusion model. Thus, we mainly focused on optimizing the effectiveness instead of efficiency, we acknowledge that efficiency optimization is a promising direction for future work. To address this, we outline several acceleration techniques that make further speed-ups highly plausible.

Efficiency Analysis

We profiled computational efficiency on a single H100-80G. We evaluated using gpt4o+SN(SWM), with beam size 2 and inference step 50 on SAT-Real. The memory consumption is~40G across the experiments. As SVC cam only generates high resolution video and therefore not efficient(around x3 world model latency), we mainly use SWM in this section.

In our main experiments, we focused on optimizing the effectiveness so the default setting is not the most cost-effective setting. To explore more cost-optimized settings, we further investigate reducing the number of inference step in the table below:

The results demonstrate that our model can maintain a >95% of the accuracy gain at ≤33% of the latency, with a latency of about 23s. We tested o1 on the SAT-Real with latency in average 19.6s, so SpatialNavigator in cost-optimized setting is in the same latency band as widely used test-time-scaling pipelines.

We further added the ablation study on beam size in the table below:

Scalability & Acceleration Pathways

Backbone Model Recent released Wan2.1-1.3B provides videos with quality higher than our backbone CogVideoX-5B in ≈8GB VRAM. Wan2.1 achieved ≈1.7×lower (8.7s vs.14.5s) latency and ≈½ the memory footprint relative our current backbone tested under identical frame counts and resolution.

Step Distillation Latent Consistency Models produce high-fidelity outputs in 2–4 inference steps, yielding 10–20× speed-ups over 50-step schedulers. Video-specific extensions such as Distribution-Matching Distillation report similar benefits on text-to-video transformers.

Low-Bit Quantization Post-training methods like SVDQuant quantize both weights and activations to 4-bit while keeping FID within ±2 points of full precision. Empirical results report ≈3× speedups and up to 3.6× memory reduction on DiTs, enabling real-time sampling on a 16GB laptop-class GPU.

All three techniques are open-source and multually compatible, which indicates that improving computational cost is both practical and well-supported by the community.

Multiview Evaluation

As additional evidence of the generalizability of our approach, we provide evaluation results on the MindCube benchmark, which evaluates VLMs' capability to perform multiview spatial reasoning in real 3D scenes. We used the Stable Virtual Camera (SVC) backbone because SVC can synthesize videos conditioned on multiple input views, whereas our current SWM lacks multiview generation support. MindCube supplies an initial frame for every question, enabling us to plug our SpatialNavigator into the multiview setting with only minor modifications.

Unlike static and single-view image datasets, the multiview setting in MindCube requires the ability to understand 3D spatial layouts and cross-view consistency. The results further demonstrate the potential of our method in real-world and multiview settings. We observe that SpatialNavigator is able to improve average accuracy by approximately 4 to 7% across different VLMs. Thus, SpatialNavigator enables VLMs to generalize to multiview settings by integrating visual cues from multiple imagined perspectives. This capability highlights its potential in embodied AI and robotic tasks. Last but not least, the results suggest that SpatialNavigator is model-agnostic and can further improve with advances in more capable world models as well as VLMs. We will include these results in the next version of the paper to make our evaluations more comprehensive.

World Model Limitations

We included additional analysis of the quality of SWM and SVC, and also their failure modes in the supplementary material, which helps the community to quantify the world model quality in our application and understand the weakness of current world models for further improvement (Sec B).

For the former, using the out-of-domain simulator AI2Thor, we evaluated the world model's fidelity under a sequence of atomic actions (turn-left/right, move-forward) defined in our work. We report the PSNR, LPIPS, and SSIM in Table 1 of the supplemental. The SVC and SWM models obtain comparable SSIM scores but the latter outperforms the former by larger margins of ~2 and 0.18 for the PSNR and SSIM scores. We refer all reviewers to Section B.1.1 of our supplemental for more information.

Quanitifying World Model Quality vs. Downstream Performance

To gauge how visual quality translates into reasoning, we ran SpatialNavigator with the original SWM and two deliberately degraded checkpoints.

As PSNR/SSIM drop and LPIPS rises, SAT accuracy falls proportionally, showing that the three metrics, especially PSNR and LPIPS provide a practical heuristic for evaluating world models before downstream deployment.

More Ablation Studies and Experiments

Experiments

We added the experiments on gpt-4.1 + SVC and InterVL3 + SVC on the SAT-Real as suggested by the reviewer

Action Space Ablations

To disentangle the role of translation versus rotation in viewpoint acquisition, we evaluated two reduced-action variants of our policy: (i) No-Move—all move-forward commands were removed; (ii) No-Turn—all turn-left / turn-right commands were removed.

According to the ablation results, the performance of turn_only is close to the complete action set. Rotations are potentially more important than forward for the tasks in SAT.

Threshold Ablations

Section 4.3 varies the pruning threshold τ controlling which roll-outs are retained. As illustrated in Fig. 4, very low thresholds (τ ≤ 5) admit numerous low-fidelity frames; accumulated error then grows with trajectory length. Performance stabilises once τ ≥ 7—the curves for τ = 7 and τ = 9 are similar.

Dear Reviewer,

We thank the reviewer for the detailed reviews and insightful suggestions. We are glad to hear that you find our method "systematic and well-grounded" and our paper well motivated and well-written. Here are the responses to your questions.

World Model Limitations

We acknowledge that the performance of our method depends on the fidelity of the world models. Therefore, we provide metrics and visualizations to analyze this dependency both quantitatively and qualitatively. We included additional analysis of the quality of SWM and SVC, and also their failure modes in the supplementary material, which helps the community to quantify the world model quality in our application and understand the weakness of current world models for further improvement (Sec B).

For the former, using the out-of-domain simulator AI2Thor, we evaluated the world model's fidelity under a sequence of atomic actions (turn-left/right, move-forward) defined in our work. We report the PSNR, LPIPS, and SSIM in Table 1 of the supplemental. The SVC and SWM models obtain comparable SSIM scores but the latter outperforms the former by larger margins of ~2 and 0.18 for the PSNR and SSIM scores. We refer all reviewers to Section B.1.1 of our supplemental for more information.

Quanitifying World Model Quality vs. Downstream Performance

To gauge how visual quality translates into reasoning, we ran SpatialNavigator with the original SWM and two deliberately degraded checkpoints.

As PSNR/SSIM drop and LPIPS rises, SAT accuracy falls proportionally, showing that the three metrics, especially PSNR and LPIPS provide a practical heuristic for evaluating world models before downstream deployment.

Future Development of World Models

Despite the current limitations of learned world models, SpatialNavigator consistently boosts VLM reasoning performance. Ongoing advances—illustrated by lighter, higher-fidelity backbones such as Wan2.1 and the controllable ReCamMaster—provide a clear trajectory for further progress. As these models improve, we expect them to unlock even greater downstream performance with SpatialNavigator.

Computational Efficiency

We appreciate the reviewer’s suggestion and agree that analyzing computational trade-offs is important for practical deployment. To address this, we provide a structured analysis of the current computational efficiency of SpatialNavigator.

Also, we would like to clarify that SpatialNavigator is the first work on test-time scaling framework that uses VLM to control a video diffusion model. Thus, we mainly focused on optimizing the effectiveness instead of efficiency, we acknowledge that efficiency optimization is a promising direction for future work. To address this, we outline several acceleration techniques that make further speed-ups highly plausible.

Efficiency Analysis

We profiled computational efficiency on a single H100-80G. We evaluated using gpt4o+SN(SWM), with beam size 2 and inference step 50 on SAT-Real. The memory consumption is~40G across the experiments. As SVC cam only generates high resolution video and therefore not efficient(around x3 world model latency), we mainly use SWM in this section.

In our main experiments, we focused on optimizing the effectiveness so the default setting is not the most cost-effective setting. To explore more cost-optimized settings, we further investigate reducing the number of inference step in the table below:

The results demonstrate that our model can maintain a >95% of the accuracy gain at ≤33% of the latency, with a latency of about 23s. We tested o1 on the SAT-Real with latency in average 19.6s, so SpatialNavigator in cost-optimized setting is in the same latency band as widely used test-time-scaling pipelines.

We further added the ablation study on beam size in the table below:

Scalability & Acceleration Pathways

Backbone Model Recent released Wan2.1-1.3B provides videos with quality higher than our backbone CogVideoX-5B in ≈8GB VRAM. Wan2.1 achieved ≈1.7×lower (8.7s vs.14.5s) latency and ≈½ the memory footprint relative our current backbone tested under identical frame counts and resolution.

Step Distillation Latent Consistency Models produce high-fidelity outputs in 2–4 inference steps, yielding 10–20× speed-ups over 50-step schedulers. Video-specific extensions such as Distribution-Matching Distillation report similar benefits on text-to-video transformers.

Low-Bit Quantization Post-training methods like SVDQuant quantize both weights and activations to 4-bit while keeping FID within ±2 points of full precision. Empirical results report ≈3× speedups and up to 3.6× memory reduction on DiTs, enabling real-time sampling on a 16GB laptop-class GPU.

All three techniques are open-source and multually compatible, which indicates that improving computational cost is both practical and well-supported by the community.

Test-Time Scaling Baselines

As suggested by the reviewer, we implemented two classic test-time scaling baselines for further comparisons.

1. Self-Consistency: The model is queried n times to give answers and rationales; the majority answer is returned.
2. Re-Ranking: The model generates n rationales and then scores them itself; the top-scoring answer is chosen.

We evaluated the two methods using gpt-4o as the base model and experimented with different n values, 10 and 30.

According the tables, although the classic test-time scaling methods generally improve from the baselines, they are not as effective as SpatialNavigators for the spatial reasoning tasks.

Thank you for the thoughtful feedback. However, several premises in your comments are inaccurate. We address each point below and provide concrete evidence that directly rebuts the misconceptions.

Training Data Size

It is very difficult to train a reliable and generalizable world model with only 50K HM3D samples and 20K real videos.

This claim is not consistent with recent controllable video generation model literature.

Comparison to recent works. Recent studies show our data regime is well within the range required for high-quality models. Stable Virtual Camera (SVC)[1] is trained on ≈47 K videos, 20 K of which (DL3DV-10K + RealEstate10K) overlap with our corpus. ReCamMaster[2], another camera-controlled video diffusion model, uses 136K videos. Our 70K-video training set therefore lies between these well-established scales.

Task difficulty. Unlike ReCamMaster, which must model temporal dynamics, and SVC, which needs to simulate any defined camera trajectories, our Search World Model (SWM) produces next-view predictions in a static scene for a restricted action space, {move-forward, turn-left, turn-right}, specialized for SpatialNavigator. This narrower scope further lowers the data requirement and training difficulty.

Strong Backbone Model. As mentioned at Sec. 3.4, SWM is fine-tuned based on CogVideoX-5B, a powerful video diffusion model. Fine-tuning from a stong backbone model also lowers the data requirement and training difficulty.

Thus, it is feasible to obtain a world model that is reliable enough for our applications with our dataset.

Additional Benchmarks

Why didn’t you consider alternative standard 3D spatial QA datasets？

We evaluated our method on a subset of a new real-world 3D spatial reasoning benchmark MindCube[3]. The results demonstrate that our method generalizes beyond SAT. (Which also addresses the concerns of reviewer Pkry and w3Uk)

Realism of the Dataset

Also, I feel that this dataset is quite unrealistic.

SAT-Real. We have demonstrated the effectiveness of our method on the real-world split SAT-Real in Table 1.

MindCube. We also demonstrated that our method is effective on a new real-world 3D spatial reasoning benchmark MindCube.

“Zoom-in” vs. True Spatial Reasoning

The improvement appears more like a zoom-in effect, rather than a meaningful enhancement in spatial reasoning.

Our ablation studies decisively refute this explanation.

We evaluated our method on two reduced action space: i) Turn-Only: move-forward is removed from the action space; ii) Forward-Only: turn-left/right are removed. The "Zoom-in Effect" is similar to the Forward-Only setting.

Forward-only—closest to a mere “zoom”—is 6–7% worse than full SpatialNavigator. It demonstrate that the combination of translation + rotation is important for optimal spatial reasoning. Therefore, the improvement is not caused by the "zoom-in effect" the reviewer mentioned.

Computational Cost

Given the substantial cost of training and using a world model, I am not yet convinced that this is a promising method.

Training Cost. As we did not train any VLMs, the training cost of SWM is comparable to training VLMs in other related works. Moreover, pre-trained controllable models such as SVC can be used zero-shot, reducing training cost to zero for adopters.

Inference Cost. Other than the detailed cost analysis we provided before, we would like to highlight that the recently released controllable world model, Genie 3 [4], reports real-time rollout, which illustrates a promising future of our method.

Novelty Clarification

We understand the general concerns of the reviewer on the quality of world models. Although current SWM and SVC are already good enough to provide substantial performance gain, there is indeed room for further improvement, as we also illustrated in the failure modes (Sec. B)

However, our main contribution is not improving the controllable world model, but rather in the test-time collaboration protocol that integrates a VLM with a pose‑conditioned world model to enhance 3D spatial reasoning.

Therefore, our method is orthogonal to advances in controllable video diffusion, and the new advances like ReCamMaster[2] and Genie 3[4] can further elevate SpatialNavigator.

[1] Stable Virtual Camera: Generative View Synthesis with Diffusion Models. https://stable-virtual-camera.github.io/

[2] ReCamMaster: Camera-Controlled Generative Rendering from A Single Video. https://jianhongbai.github.io/ReCamMaster/

[3] MindCube: Spatial Mental Modeling from Limited Views. https://mind-cube.github.io/

[4] Genie 3: A new frontier for world models. https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/

Dear Reviewer,

We appreciate the detailed comments and insightful suggestions. We are glad that you find our method both high-intuitive and effective. Here are the responses to your questions.

Computational Efficiency

We appreciate the reviewer’s suggestion and agree that analyzing computational trade-offs is important for practical deployment. To address this, we provide a structured analysis of the current computational efficiency of SpatialNavigator.

Also, we would like to clarify that SpatialNavigator is the first work on test-time scaling framework that uses VLM to control a video diffusion model. Thus, we mainly focused on optimizing the effectiveness instead of efficiency, we acknowledge that efficiency optimization is a promising direction for future work. To address this, we outline several acceleration techniques that make further speed-ups highly plausible.

Efficiency Analysis

We profiled computational efficiency on a single H100-80G. We evaluated using gpt4o+SN(SWM), with beam size 2 and inference step 50 on SAT-Real. The memory consumption is~40G across the experiments. As SVC cam only generates high resolution video and therefore not efficient(around x3 world model latency), we mainly use SWM in this section.

In our main experiments, we focused on optimizing the effectiveness so the default setting is not the most cost-effective setting. To explore more cost-optimized settings, we further investigate reducing the number of inference step in the table below:

The results demonstrate that our model can maintain a >95% of the accuracy gain at ≤33% of the latency, with a latency of about 23s. We tested o1 on the SAT-Real with latency in average 19.6s, so SpatialNavigator in cost-optimized setting is in the same latency band as widely used test-time-scaling pipelines.

We further added the ablation study on beam size in the table below:

Scalability & Acceleration Pathways

Backbone Model Recent released Wan2.1-1.3B provides videos with quality higher than our backbone CogVideoX-5B in ≈8GB VRAM. Wan2.1 achieved ≈1.7×lower (8.7s vs.14.5s) latency and ≈½ the memory footprint relative our current backbone tested under identical frame counts and resolution.

Step Distillation Latent Consistency Models produce high-fidelity outputs in 2–4 inference steps, yielding 10–20× speed-ups over 50-step schedulers. Video-specific extensions such as Distribution-Matching Distillation report similar benefits on text-to-video transformers.

Low-Bit Quantization Post-training methods like SVDQuant quantize both weights and activations to 4-bit while keeping FID within ±2 points of full precision. Empirical results report ≈3× speedups and up to 3.6× memory reduction on DiTs, enabling real-time sampling on a 16GB laptop-class GPU.

All three techniques are open-source and multually compatible, which indicates that improving computational cost is both practical and well-supported by the community.

Constrained scaling the Spatial Beam Search

One fundamental constraint is error accumulation. With greater search depths, there is an increased likelihood of compounding errors in both the imagined views synthesized by the world model and the subsequent reasoning steps by the VLM. The lower fidelity of the subsequent generated views may hurt VLMs' ability to perform spatial reasoning. We further evaluated the world models' fidelity at different steps to confirm it.

Using the out-of-domain simulator AI2Thor, we evaluated the world model's fidelity under a sequence of atomic actions defined in our work. We report the PSNR, LPIPS, and SSIM in Tab. 1 of the supplemental. The SVC and SWM models obtain comparable SSIM scores but the latter outperforms the former by larger margins of ~2 and 0.18 for PSNR and SSIM. We refer all reviewers to Sec. B.1.1 for more information. To further analyze the reliance of these world models, we evaluate the performance when extended to multiple steps(1 to 3) in the following table.

Generally, we observe consistent trends across the SVC and SWM models, where their effectiveness gradually decrease as the number of steps increases. This is expected since the uncertainty of predicting the exact visual changes is compounded over steps.

Critical advancements required

Given our observation that the fidelity of the world model can deteriorate quickly over extended imagined trajectories, we believe that addressing this challenge is important. Some possibilities include but are not limited to a long-term memory module that is capable of storing a mental model of a 3D scene as it is explored, as well as learning to generate environments based on scene-graphs. The latter is aimed at improving generalizability to spatial layouts unseen during training time.