acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aware Memory Coordination for Spatial Understanding from Long Videos

Abstract

Spatial understanding from long videos requires integrating partial observations across viewpoints and recalling past content within a limited memory budget. Repeated visits can reveal complementary views or changes in familiar regions. Existing memory methods use predictive novelty to guide retention or enrich language inputs with geometric features. Yet novelty under viewpoint changes need not reflect distinct content, and an observation can be redundant for geometry yet retain unique content for future questions. We introduce GEOMEM, a geometry-aware framework that coordinates visuospatial and episodic memories. Visuospatial memory provides historical context for geometry encoding of new views. The resulting features enrich the visual inputs used to form episodic memory, which retains observed content and its temporal context for later questions. Shared cross-view evidence guides separate retention decisions for geometric support and content under fixed budgets. Historical content can therefore remain accessible even when its source view becomes geometrically replaceable. Experiments on VSI-Bench and VSI-SUPER show leading performance in spatial understanding, long-horizon recall, and counting. Budget analyses show improved geometry estimation and question answering at matched memory budgets. Ablations support geometry fusion and complementary memory updates. Combining geometric association with content comparison outperforms either criterion alone. Streaming measurements show stable peak GPU memory and per-frame observation time on a four-hour video. Code and scripts will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
memory coordination, visuospatial memory, episodic memory, long video understanding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YwmMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies spatial understanding from long videos under a fixed memory budget. The paper claims that a frame can be redundant for geometry estimation yet still hold content that later questions may need. GeoMem therefore keeps two separate KV memories. One stores geometry-encoder states and keeps frames by geometric coverage. The other stores language model states and drops a token only when it matches a retained token geometrically and visually. It reports state-of-the-art results on VSI-Bench and VSI-SUPER, with stable memory and per-frame time on a four-hour video.

Critical strengths
  1. The paper separates the history a geometry encoder needs from the history a language model needs, and it keeps them under independent budgets. The design is intuitively reasonable.

  2. The method adds only small peak memory and time per frame, and both stay flat over a four-hour video. This supports long-video understanding under a fixed budget.

Critical weaknesses
  1. The paper does not convincingly show that geometry is what makes the memory work. The method is presented as geometry-aware, but the evidence is confusing. On VSI-Bench, GeoMem-7B beats the strongest baselines by only about 1-2 points, and the models use different backbones. GeoMem's clearest advantages are on counting, object size, and appearance order, which are largely semantic or temporal tasks. It does not lead on absolute distance, relative direction, or route planning, where geometry should help the most.

  2. GeoMem's accuracy barely drops as videos get longer. But the episodic cache is capped at 16,384 tokens, while a four-hour video sampled at 1 FPS has 14,400 frames, so the cache can keep only a small fraction of what the model has seen. The paper does not show what is actually kept at these lengths or why that is enough to answer the questions. One possible explanation is simply recency. If the frames needed to answer happen to be near the end of the video, a small cache of recent content would give the same flat curve without any long-term memory. The benchmarks are also small, with 60 questions per recall duration and 50 per counting duration, so gaps of a few points amount to only a few questions. The headline long-horizon results are difficult to attribute to the proposed design.

Questions for authors

For the long VSI-SUPER videos, how much of the video does the episodic cache actually cover at 120 and 240 minutes (for example, tokens per frame and the time span kept)? Where in the video does the evidence needed for each question lie, and how does accuracy change when it is split by that position?

The gains are strongest on counting, object size and appearance order, and weaker on relative direction and route planning. Is this expected? What does it imply about where geometry helps?

2 comments
AuthorsOpenReview ↗

Response to Reviewer YwmM (1/2)

We appreciate your recognition of our memory design, efficiency, and stable resource usage over long streams. We address your comments point by point below.

YwmM-W1

Geometry’s contribution to GeoMem’s memory remains unclear. GeoMem-7B gains only 1–2 points over the strongest VSI-Bench baselines with different backbones. Its clearest advantages are in counting, object size, and appearance order, largely semantic or temporal tasks. It does not lead on absolute distance, relative direction, or route planning, where geometry should help most.

Thank you for asking us to clarify geometry’s contribution. We agree that comparisons across backbones alone do not isolate this contribution.

Empirically, we additionally evaluate the Feature only setting from Tab. 4 (original) on VSI-Bench. We fix the GeoMem-7B checkpoint, VSM Coverage, geometry fusion, memory budgets, and reading policy, isolating the contribution of geometric correspondence to EM deduplication.

EM Dedup.Avg.Obj. Cnt.Abs. Dist.Obj. SizeRoom SizeRel. Dist.Rel. Dir.Route PlanAppr. Order
Ours69.975.253.577.272.370.479.847.982.8
Feature only68.272.352.175.971.269.079.047.279.1

Within GeoMem, geometric correspondence improves all eight subtasks, including absolute distance, relative direction, and route planning. These gains characterize its contribution to EM retention within the same model. Rankings against other systems also involve differences in backbone, training, and components. Fig. 3 (original) and Tab. 3 (original) further support the benefits of VSM Coverage.

We have added this comparison to §4.4 of the revised paper.

YwmM-W2-1

GeoMem’s accuracy barely drops with video length, although its 16,384-token cache can retain only a small fraction of 14,400 frames over four hours at 1 FPS. The paper does not show what is kept or why it suffices. If question evidence is near the end, recent content alone could explain this flat curve without long-term memory.

Good question. Empirically, we additionally evaluate GeoMem-7B using only the final 10 minutes of each VSR-240 video. We reset both memories and all other streaming states, then answer the same 60 questions. All settings below share the checkpoint, geometry fusion, VSM Coverage, memory budgets, and reading policy.

EM UpdateVideo InputVSR-240 Acc.
OursFull 240 min43.3
w/o EM Dedup. (FIFO)Full 240 min35.0
OursLast 10 min only26.7

The input window comparison supports the contribution of earlier observations to the full system. The comparison with EM FIFO separately demonstrates the benefit of our EM update strategy with the VSM policy fixed. These experiments do not isolate the respective contributions of EM and VSM to the benefit of earlier input. We report the actual cache coverage in Q1-1 and extend the same input window comparison in Q1-2.

We have added the comparison to §4.3.

YwmM-W2-2

The benchmarks are also small, with 60 questions per recall duration and 50 per counting duration, so gaps of a few points amount to only a few questions. The headline long-horizon results are difficult to attribute to the proposed design.

We agree that the limited number of questions warrants caution when interpreting small score differences. For example, on VSR-240, our method answers 26 of 60 questions correctly, compared with 21 of 60 for EM FIFO.

The controlled comparisons in W1 and W2-1 support the contributions of geometric correspondence and the EM update policy, respectively. Q1-2 examines the full system’s dependence on earlier input. Together with the component ablations in Tab. 3 (original), these analyses help explain the observed gains, but the uncertainty associated with the small VSI-SUPER evaluation sets remains.

AuthorsOpenReview ↗

Response to Reviewer YwmM (2/2)

YwmM-Q1-1

For the long VSI-SUPER videos, how much of the video does the episodic cache actually cover at 120 and 240 minutes (for example, tokens per frame and the time span kept)?

At both durations, EM retains an average of 182.04 visual tokens per source frame represented in the cache, counting only frames with retained entries. The mean earliest and latest source timestamps are 118.50 and 119.97 minutes for the 120-minute setting, and 238.50 and 239.97 minutes for the 240-minute setting.

These timestamps span approximately 1.47 minutes on average at both durations, indicating that the retained entries originate mainly from the recent portion of each video.

YwmM-Q1-2

Where in the video does the evidence needed for each question lie, and how does accuracy change when it is split by that position?

The available VSR annotations do not include event timestamps, and the questions concern the order of multiple appearances. We therefore assess the full system’s dependence on earlier observations through controlled input windows.

Empirically, we compare three input windows on the same 60 VSR-240 questions, reusing the Full 240 min and Last 10 min results from W2-1. Each setting starts with empty memories and reset streaming states, while keeping the checkpoint, sampling rate, memory budgets, reading policy, and decoding settings fixed.

Video InputVSR-240 Acc.
Full 240 min43.3
Last 120 min only31.7
Last 10 min only26.7

Accuracy decreases when the first 120 minutes are omitted and falls further when only the final 10 minutes remain. These results support the contribution of earlier input to the full system’s answers. This comparison does not identify individual questions’ evidence positions or provide accuracy grouped by those positions.

YwmM-Q2

The gains are strongest on counting, object size and appearance order, and weaker on relative direction and route planning. Is this expected? What does it imply about where geometry helps?

GeoMem uses geometry to improve the encoding and retention of historical evidence. Geometric correspondence helps distinguish repeated observations from visually similar content at different locations, which can benefit counting and appearance order. Object size also involves geometric estimation. Our memory updates do not introduce a dedicated procedure for direction reasoning or route planning, so gains need not be uniform across tasks.

Empirically, the controlled comparison in W1 shows larger gains for counting and appearance order, while direction and planning also improve. Both settings retain geometry fusion and VSM Coverage, so this comparison specifically isolates the contribution of geometric correspondence to EM retention.

We have clarified this interpretation in §4.4.

Thank you again for your constructive feedback. We have incorporated the indicated revisions into the paper and provided further analyses in this rebuttal.

Reviewer_Dr2qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

GeoMem introduces a dual-memory framework built on a geometry encoder and a VLM, maintaining visuospatial and episodic memories using shared cross-view evidence. Experiments demonstrate competitive performance on spatial understanding, long-horizon recall, and continual counting on VSI-Bench and VSI-SUPER. Ablations support the benefits of geometry fusion and the complementary roles of the two memory update strategies.

Critical strengths
  1. The paper distinguishes the memory requirements of continued geometry encoding from those of answering questions about past observations. It uses shared cross-view evidence to update the two memories according to geometric coverage and content redundancy, respectively.
  2. GeoMem achieves competitive results on VSI-Bench and VSI-SUPER, covering spatial understanding, long-horizon recall, and continual counting.
Critical weaknesses
  1. GeoMem's dual-memory architecture, in which the geometry encoder and the VLM maintain their respective KV caches, has precedents in prior work such as JanusVLN.
  2. The claimed "separate retention decisions" should be distinguished from behavior already enabled by independently managed caches. Different retention policies or budgets can naturally lead the two caches to retain different historical observations. Therefore, the fact that episodic content remains accessible after its source frame is removed from the geometry cache is, by itself, insufficient to establish the novelty of the method.
Questions for authors
  1. Could the authors clarify the specific methodological advances and measurable benefits of geometric coverage selection and content deduplication based on shared cross-view evidence, relative to existing dual-memory management mechanisms?
  2. The main comparisons involve different backbones, parameter counts, and training configurations, making it difficult to attribute the performance differences solely to the proposed memory mechanism. For example, the comparison between GeoMem-8B and VLM²-7B involves differences in both backbone and model size. While the paper includes ablations using the same model, could the authors provide comparisons under a common backbone to better isolate the benefits of the proposed mechanism?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Dr2q (1/2)

We appreciate your positive assessment of our memory design and experimental results. We address your comments point by point below.

Dr2q-W1

GeoMem's dual-memory architecture, in which the geometry encoder and the VLM maintain their respective KV caches, has precedents in prior work such as JanusVLN.

Thank you for highlighting this relevant connection. We acknowledge JanusVLN as a precedent for maintaining geometric and semantic KV memories. GeoMem’s contribution lies in using shared cross-view evidence to coordinate retention according to the different requirements of continued geometry encoding and future content retrieval.

JanusVLN retains KV from initial frames and a recent sliding window. GeoMem also protects reference and recent frames, while selecting additional VSM views by geometric coverage, assessed through support from other selected views. EM combines geometric correspondence and visual similarity to identify repeated content, omitting current tokens only when their historical representatives remain in the updated cache. These distinct criteria guide frame retention for continued geometry encoding and token retention for future recall.

We have clarified this contribution in §1 and the comparison with JanusVLN in §2 of the revised paper.

Dr2q-W2

Independent caches can naturally retain different histories through different retention policies or budgets. Please clarify what makes “separate retention decisions” novel, since episodic content remaining accessible after its source frame leaves the geometry cache is insufficient evidence by itself.

Good Question. We agree that independent caches can naturally retain different histories, so continued access to episodic content after its source frame leaves VSM does not alone establish novelty. Our contribution lies in using shared cross-view evidence to assess geometric coverage for VSM and content redundancy for EM. The latter requires both geometric correspondence and visual similarity, with the historical representative retained after the update.

Empirically, Tab. 3 (original) keeps both caches active and fixes the checkpoint, budgets, and reading policy, isolating the benefits of these update rules.

We have clarified this distinction in §1 and §3.3.

AuthorsOpenReview ↗

Response to Reviewer Dr2q (2/2)

Dr2q-Q1

Could the authors clarify the specific methodological advances and measurable benefits of geometric coverage selection and content deduplication based on shared cross-view evidence, relative to existing dual-memory management mechanisms?

Thank you for suggesting a more direct comparison. Empirically, we additionally evaluate an adaptation of JanusVLN’s initial and recent window retention policy within GeoMem-7B. We fix the checkpoint, geometry fusion, memory budgets, and reading policy, comparing retention rules within the same trained model.

For VSM, we use the FIFO baseline described in §4.3 (original). For EM, we protect all visual tokens from the first frame within the total token budget. The remaining capacity stores recent content, with individual tokens evicted in FIFO order.

Memory Update PolicyVSI-Bench Avg.VSR-120 Acc.VSC-120 MRA
Ours69.943.338.8
Initial + Recent (adapted)65.038.334.2

Our updates improve all three metrics under the same budgets. Tab. 3 (original) further shows that VSM Coverage and EM Dedup. each contribute, while Tab. 4 (original) shows that combining geometric correspondence with visual similarity outperforms either criterion alone. These results support using distinct geometric and content criteria to update the two memories.

We have added this comparison to §4.4.

Dr2q-Q2

Main comparisons differ in backbone, model size, and training configuration, including GeoMem-8B versus VLM²-7B, making it difficult to isolate the benefits of the memory mechanism. Beyond the existing ablations within the same model, could the authors provide comparisons under a common backbone?

The comparison between GeoMem-8B and VLM²-7B reflects overall system performance, with differences in backbone, model size, and training configuration. Empirically, GeoMem-2B and Spatial-TTT-2B share the Qwen3-VL-2B backbone, achieving 66.4 and 64.4 on VSI-Bench, respectively, in Tab. 1 (original). This provides a comparison under a common backbone, although training configurations and other components differ.

To isolate the effects of memory updates, the additional comparison in our response to Q1 fixes the GeoMem-7B checkpoint, geometry fusion, memory budgets, and reading policy. Our updates outperform the adapted JanusVLN retention policy across all three metrics, supporting their benefits within the same trained model.

We have clarified the scope in §4.2 and added the comparison to §4.4.

Thank you again for your careful review and constructive suggestions. Your comments helped us clarify our contribution and strengthen the experimental comparisons in the revised version.

Reviewer_821aOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Long-video spatial understanding with a fixed memory budget. The authors keep two caches—one for continued geometry encoding, one for content retrieval—and use a cross-view check (reprojected pose matching + token similarity) to decide, per cache, whether a new frame is redundant or worth keeping. Trained with VGGT distillation; evaluated on VSI-Bench and VSI-SUPER with ablations and a 4-hour streaming run.

Critical strengths

Clean and useful split, geometric redundancy and content redundancy are not the same thing, and existing memory policies conflate them. Operationalizing the split with cross-view evidence is a sensible contribution.

Critical weaknesses

Incremental. VLM-2 / VLM-3R / Spatial-TTT already do geometry-aware episodic memory. The paper argues against a predictive-novelty strawman but never spells out what the two-cache design actually buys over those systems.

Questions for authors

How does dedup accuracy degrade as VGGT pose/depth drifts over 120–240 min clips? Why does 2B beat 7B on VSI-Bench—backbone effect?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 821a (1/2)

We sincerely thank you for recognizing the value of our method. We address your comments point by point below.

821a-W1

VLM-2 / VLM-3R / Spatial-TTT already do geometry-aware episodic memory. The paper argues against a predictive-novelty strawman but never spells out what the two-cache design actually buys over those systems.

Good question! Our contribution is to coordinate memory retention according to the different demands of continued geometry encoding and future content retrieval.

VLM² [1] maintains recent and older fused representations in working and episodic memories, with feature similarity guiding episodic replacement. VLM-3R [2] retains geometric history through CUT3R’s recurrent state, while Spatial-TTT [3] accumulates history in fast weights. GeoMem uses shared cross-view evidence to select frame KV for continued geometry encoding and token KV for episodic retrieval under separate budgets. VSM selects views by geometric coverage, while EM combines geometric correspondence and content similarity to omit repeated current tokens only when their historical representatives remain in the updated cache.

Empirically, the ablation below, reproduced from Tab. 3 (original), keeps both caches active and fixes other settings. Each ablated update is replaced with FIFO.

SettingVSI-Bench Avg.VSR-120 Acc.VSC-120 MRA
Ours69.943.338.8
w/o VSM Coverage66.740.035.8
w/o EM Dedup.67.638.334.8
w/o VSM & EM Updates64.633.330.8

Both updates contribute, and combining them performs best across spatial understanding, recall, and counting. These results support the complementary benefits of the proposed retention rules within the same model.

We have clarified the methodological differences in §1 and §2 of the revised paper.

[1] Zuntao Liu et al. Vision-Language Memory for Spatial Reasoning. ECCV, 2026.

[2] Zhiwen Fan et al. VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction. CVPR, 2026.

[3] Fangfu Liu et al. Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training. ECCV, 2026.

821a-Q1

How does dedup accuracy degrade as VGGT pose/depth drifts over 120–240 min clips?

Thank you for highlighting this concern. GeoMem omits tokens only when both geometric and visual checks pass (§3.2–§3.3).

Empirically, we additionally conduct two analyses using GeoMem-7B on the 60 VSR-240 videos with default memory budgets. First, we sample 100 deduplication pairs per hour per video, with ten sampled without replacement in each six-minute interval. Each pair comprises an omitted current token and its retained historical representative, and is grouped by the current token’s timestamp. Offline VGGT and DA3 each jointly process the two source frames. Both references evaluate all sampled pairs without confidence filtering. Agreement is the percentage whose reference reprojection reaches the selected historical token’s cell and satisfies the 5% relative depth tolerance.

Time Interval (min)VGGT Agreement (%)DA3 Agreement (%)
0–6087.483.2
60–12085.682.0
120–18084.381.2
180–24083.880.9

Both references show a gradual decline in agreement over stream time. These values measure consistency with predicted geometry and serve as a proxy for geometric matching quality.

Second, we perturb only the geometry used for EM matching. Rotation and translation receive Gaussian noise, with translation RMS normalized by median valid depth. Depth receives one lognormal multiplier per frame with mean 1 and the reported standard deviation. Perturbations remain fixed for each frame, while VSM geometry and fused features remain unchanged.

SettingRotation Std. (°)Translation RMS (%)Depth Scale Std. (%)VSR-240 Acc. (%)
Original00043.3
Mild11241.7
Strong33538.3

Recall decreases slightly under mild perturbations and more under stronger perturbations. This experiment measures sensitivity to geometric errors in EM matching, rather than simulating accumulated drift.

We have added both analyses and their settings to §4.3.

AuthorsOpenReview ↗

Response to Reviewer 821a (2/2)

821a-Q2

Why does 2B beat 7B on VSI-Bench—backbone effect?

In Tab. 1 (original), GeoMem-2B achieves 66.4, below GeoMem-7B at 69.9. They use Qwen3-VL-2B and Qwen2.5-VL-7B, respectively. Backbone differences can affect comparisons with other 7B methods. For example, the original Qwen3-VL-2B already outperforms Qwen2.5-VL-7B, scoring 50.3 versus 31.4. We therefore assess the benefits of our memory updates through the controlled ablations in Tab. 3 (original), which fix the checkpoint, geometry fusion, memory budgets, and reading policy.

We have clarified this comparison in §4.2.

Thank you again for your time and constructive feedback. We have incorporated these clarifications and the additional experiments into the revised version.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Responses to all Reviewers

Dear Chairs and Reviewers,

Thank you for your constructive and detailed feedback.
This note addresses four concerns shared across the reviews.

1. Contribution beyond Existing Memory Designs (821a, Dr2q)
GeoMem introduces distinct retention rules for VSM and EM based on shared cross-view evidence. VSM keeps views that provide geometric coverage for encoding new frames. EM keeps content for later questions, removing repeated tokens only when geometric and visual checks pass and their historical representatives remain stored. We clarified these rules and their differences from prior designs in §1, §2, and §3.3 of the revised paper.

2. Backbone Effects and Performance Comparisons (821a, Dr2q, YwmM)
The revised discussion in §4.2 explains how backbone differences affect our comparisons. GeoMem-2B and Spatial-TTT-2B share Qwen3-VL-2B, although their training and other components differ. Empirically, Tabs. 5 and 8 (revised) compare memory updates using the same GeoMem-7B checkpoint and memory budgets. Our updates outperform FIFO and the adapted Initial + Recent policy, supporting their contribution within the same model.

3. Geometry’s Contribution to Memory Updates (Dr2q, YwmM)
Empirically, Tab. 9 (revised) compares our EM update with the Feature only setting. Both retain VSM Coverage and geometry fusion. Geometric correspondence improves all eight VSI-Bench subtasks, including direction and route planning. The larger gains in counting and appearance order are consistent with better association of observations across views. We expanded the discussion in §4.4 to explain how geometry supports evidence retention and why benefits vary across tasks.

4. Memory Behavior over Long Streams (821a, YwmM)
The revised §4.3 examines geometric matching and the contribution of earlier input. Tabs. 3–4 (revised) show gradually declining agreement with VGGT and DA3 references and greater recall loss under stronger geometric perturbations. Tab. 7 (revised) shows higher recall with full videos than with only the final 10 minutes. This supports the benefit of earlier input to the full system. EM Dedup. also outperforms FIFO with the same VSM policy. We additionally report the source time span of the final EM entries in §4.3.

We hope these clarifications address the shared concerns and help explain our design and results.
Thank you again for your time and thoughtful comments.

Sincerely,
The Authors

AuthorsOpenReview ↗

Summary of Revisions to the Paper

Dear Chairs and Reviewers,

We appreciate your constructive feedback and the opportunity to revise our paper.
This document summarizes the main changes made in the revised version.
Separate responses will provide global clarifications and point-by-point responses to individual comments.

Below we list the key revisions:

  1. We refined the abstract to describe the two memory update criteria more clearly.
  2. We clarified GeoMem’s methodological contribution in §1.
  3. We expanded the comparison with VLM², VLM-3R, Spatial-TTT, and JanusVLN in §1 and §2.
  4. We clarified the distinct VSM and EM retention criteria in §3.3.
  5. We clarified the backbone differences between GeoMem-2B and GeoMem-7B in §4.2.
  6. We expanded the discussion of GeoMem and Spatial-TTT using the same backbone in §4.2.
  7. We added an EM Dedup. versus EM FIFO comparison on VSR-240 in Tab. 7 in §4.3.
  8. We added comparisons using full videos and shorter input windows in §4.3.
  9. We reported tokens per retained source frame and the final EM’s temporal coverage in §4.3.
  10. We analyzed geometric matching agreement over time using VGGT and DA3 in Tab. 3 in §4.3.
  11. We added a geometric perturbation analysis for EM matching in Tab. 4 in §4.3.
  12. We added a comparison with an Initial + Recent policy adapted from JanusVLN in Tab. 8 in §4.4.
  13. We specified the adapted policy’s initial frame retention and FIFO settings in §4.4.
  14. We added a Feature-only comparison across eight VSI-Bench subtasks in Tab. 9 in §4.4.
  15. We expanded the discussion of how geometric correspondence benefits different spatial tasks in §4.4.

We hope these revisions help clarify our method and address the main concerns raised in the reviews.
Thank you again for your time and thoughtful comments.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.