PaperlayerView exact revision on arXiv

Source-linked AI summary

Dyn-3D: Unveiling and Resolving Ego-Motion Ambiguity in Vision-Language Models

Jiayu Ding, Zhuodong Liu, Lei Zhang, Manyu Xiong, Hongbo Jin, Haoran Tang, Hongbo Zhang, Changen Zhu, Wenbo Xing

arXiv:2609.01059v1cs.CV

TL;DR

VLMs need ego-motion perception to resolve monocular scale ambiguity, but current models struggle beyond smooth camera-motion patterns. Dyn-3D evaluates this with counterfactual rendering, and TempoVista applies physical supervision through Kinematic-GSPO; the authors report improved performance on Dyn-3D and generalization to VSI-Bench spatial tasks.

  • Problem

    VLMs need ego-motion perception to resolve monocular scale ambiguity, yet current models struggle with large displacements and non-smooth trajectories.

    Source: §1 · p. 3

  • Method

    Dyn-3D uses counterfactual 3D-rendered video pairs to evaluate ego-motion, while TempoVista combines kinematic-adaptive frame selection with Kinematic-GSPO using physical ground truth.

    Source: §1 · p. 3, §5 · p. 8

  • Results

    TempoVista outperforms leading open and proprietary models on Dyn-3D and generalizes to diverse spatial tasks in VSI-Bench.

    Source: §7 · p. 12

  • Takeaways & Limitations

    The authors conclude that explicit kinematic perception is essential for robust 3D spatial understanding.

    Source: §7 · p. 12

    The oracle frame-selection strategy assumes reliable camera extrinsics, such as from odometry or the 3D Gaussian Splatting pipeline.

    Source: §5.1.1 · p. 8

Abstract

from arXiv · show

As Vision-Language Models (VLMs) tackle dynamic 3D spatial reasoning, ego-motion perception becomes essential to resolve monocular scale ambiguity. However, current models often overfit to smooth trajectory priors rather than genuinely understanding physical motion. Consequently, their spatial reasoning degrades severely under large displacements, a phenomenon we term Kinematic Collapse. This failure stems from spurious visual-motion correlations in natural videos and a lack of explicit physical supervision. To evaluate this, we introduce Dyn-3D, a benchmark using counterfactual 3D rendering to rigorously decouple visual changes from true kinematic properties. Furthermore, we propose the TempoVista framework, featuring the Kinematic-GSPO algorithm. By embedding metric physical ground truth into policy optimization, TempoVista explicitly grounds visual representations in 3D space. Experiments demonstrate that our approach significantly improves both motion estimation and robust spatial reasoning by utilizing camera dynamics as an effective geometric calibration signal.

1 Introduction

Ego-motion provides the metric reference needed to distinguish scene depth from camera movement, but models relying on smooth-trajectory priors fail under large or non-smooth displacements. Dyn-3D tests this failure with counterfactual videos, while Kinematic-GSPO uses physical ground truth to improve motion perception and spatial reasoning.

Introduction

  • 25.30 percentage points: Qwen3.5-9B overall accuracy fell from 63.97% to 38.67% as inter-frame displacement increased from 0.1 to 5.0 meters.Kinematic perception accuracy fell from 69.71% to 9.66%, while implicit trajectory reasoning remained below 25% chance at all displacement magnitudes.Source: §1 · p. 3
  • Dyn-3D uses 3D Gaussian Splatting [17] to create counterfactual video pairs with identical visual paths but different kinematics, testing VLM ego-motion perception across 16,063 samples.Advanced models achieve near-chance accuracy estimating camera-motion magnitude under non-smooth or counter-intuitive trajectories.Source: §1 · p. 3
  • Kinematic-GSPO applies process-level supervision with physical ground truth to penalize unphysical motion inferences and separate true kinematics from visual appearance.The work also introduces the 33.6K-sample Dyn-3D-Instruct dataset for training and proposes Dyn-3D and Dyn-3D-Instruct through counterfactual re-rendering.Source: §1 · p. 3
  • Improved kinematic perception enhances spatial reasoning across depth, orientation, and occlusion, and boosts accuracy on 3D benchmarks including Dyn-3D and VSI-Bench.The introduction attributes spatial failure without smooth trajectory priors to kinematic ambiguity and identifies physical-ground-truth supervision as a way to improve camera-motion understanding.Source: §1 · pp. 3–4

2 Related Work

Prior MLLM benchmarks assess video and spatial reasoning without isolating ego-motion, and VLMs rarely receive direct camera-motion supervision. Dyn-3D addresses this gap through counterfactual rendering and physically consistent motion optimization.

Video and spatial reasoning in MLLMs

  • Video MLLM benchmarks assess event, temporal, and commonsense reasoning, while spatial benchmarks cover object grounding, metric distance, and memory but rarely isolate ego-motion [12] [21] [24] [34] [2] [5] [32] [22].Related work also explores language-guided 3D object grounding [9] [10] [16] and tracking in dynamic 4D scenes [19]. These benchmarks primarily assess downstream performance, so high accuracy may reflect natural-video motion correlations rather than camera-dynamics reasoning.Source: §2 · p. 4

Ego-motion and geometric ambiguity

  • Classical geometry recognizes monocular scale ambiguity, and SfM/SLAM use explicit pose estimation and map consistency, whereas VLMs rarely receive direct camera-motion supervision [26].Large appearance changes can be mistaken for large translation, while stable appearance can hide substantial displacement; Dyn-3D targets this visual-kinematic ambiguity directly.Source: §2 · pp. 4–5

Controllable 3D rendering for diagnostic evaluation

  • NeRF and 3D Gaussian Splatting enable photorealistic novel-view synthesis and counterfactual controls that vary trajectory dynamics or separate rotation from translation under fixed geometry [17].Using ScanNet++ scenes and nerfstudio-based 3DGS rendering [30] [33], Dyn-3D makes ego-motion an explicit intervention variable to prevent reliance on superficial visual-motion correlations.Source: §2 · p. 5

Reasoning optimization for multimodal models

  • Kinematic-GSPO augments answer-level reward with physical-consistency supervision over path length, displacement, direction, rotation, and speed, making geometric correctness an optimization target.Prior chain-of-thought and RL methods optimize intermediate reasoning traces [8] [31], and Video-R1 extends this direction to video MLLMs [11].Source: §2 · p. 5

3 Pilot Experiments

Two pilot experiments test whether VLMs understand physical motion or rely on superficial visual changes. They reveal severe failure under large displacements and confusion between camera viewpoint changes and physical displacement.

3.1 Kinematic Collapse Under High Dynamics

  • The authors define this motion-magnitude-related degradation as Kinematic Collapse: large displacements break local visual correspondences, and implicit trajectory reasoning stays below 25% chance at every displacement.They attribute the failures to VLMs relying on continuous visual cues common in smooth training data.Source: §3.1 · p. 5

3.2 Visual-Kinematic Misalignment

  • The evaluation uses four option-shuffled versions of 835 B17 motion-type recognition pairs, yielding 3,340 instances with balanced answer positions from the existing 16,063-question benchmark.These are evaluation instances, not additional benchmark questions.Source: §3.2 · p. 5
  • Across decoupled trajectories, Qwen3.5-9B, Qwen3-VL-8B-Instruct [3], and Qwen-VL-Max average 17.22%, 13.70%, and 19.84% accuracy, respectively.Qwen3.5-9B scores 28.74% on pure rotation and 5.69% on pure translation; models misclassify pure rotation as translation plus rotation at rates of 71.26%, 99.70%, and 100.00%, respectively.Source: §3.2 · p. 5

3.1 Kinematic Collapse Under High Dynamics

  • As ∆d rises from 0.1m to 5.0m, Qwen3.5-9B overall accuracy drops from 63.97% to 38.67%, while kinematic perception falls from 69.71% to 9.66%.Spatial understanding accuracy decreases from 67.53% to 52.76%.Source: §3.1 · p. 5, Table 1 · p. 6

4 Dataset

Dyn-3D uses counterfactual videos with identical spatial paths but different dynamics to test whether VLMs understand ego-motion rather than exploit visual-motion correlations.

4.1 Data Acquisition and Generation

  • The pipeline reconstructs 443 eligible ScanNet++ scenes and renders five motion characteristics per scene, including controlled rotation-only and translation trajectories.Each video receives metadata covering camera motion and visible-object geometry, with coordinates restored to the world frame.Source: §4.1 · p. 6

4.2 Annotation Pipeline and Quality Control

  • Questions are deterministically generated from 3D metadata as 19 four-option types covering kinematics, spatial understanding, and implicit trajectory reasoning.Evaluation questions undergo geometry-based answer recomputation, physical-validity filtering, and manual review for clarity and agreement with rendered evidence.Source: §4.2 · p. 7

4.3 Dataset Statistics and Splits

  • The benchmark contains 16,063 four-option questions from 167 scenes and 835 videos, spanning kinematic perception, spatial understanding, and implicit trajectory reasoning.Questions are balanced across five trajectory types, with 168 diagnostic rotation and translation traps; answers are independently recomputed from 3D metadata.Source: §4.3 · p. 7

4.4 Comparison with Existing Benchmarks

  • Dyn-3D isolates rotation, translation, and speed with counterfactual videos, addressing benchmarks that do not control these dynamics in 3D trajectories.The benchmark targets ego-motion perception and models’ vulnerability to Kinematic Collapse.Source: §4.4 · p. 7, Table 2 · p. 7

5 Method

TempoVista addresses motion ambiguity with kinematic-adaptive frame selection and Kinematic-GSPO, which uses explicit policy optimization to ground VLM motion reasoning in physical kinematics.

5.1 Kinematic-Adaptive Frame Selection

  • Kinematic-adaptive selection allocates a limited frame budget by camera-motion variation rather than timestamp spacing, targeting intervals where motion changes rapidly.Uniform sampling may underrepresent sharp turns or fast translations.Source: §5.1 · p. 8

5.1.1 Oracle Selection via SE(3) FPS

  • With reliable camera extrinsics, oracle selection uses farthest-point sampling in SE(3) to cover trajectory diversity, selecting more frames in rapidly changing regions and fewer in near-static ones.Its distance balances normalized translation and rotation; initializing with the first and last frames preserves temporal boundaries.Source: §5.1.1 · pp. 8–9

5.2 Kinematic-GSPO: Physics-Guided Group Sequence Policy Optimization

  • Kinematic-GSPO trains models to produce traces of five camera-motion quantities and rewards consistency with scene-derived kinematic annotations alongside answer correctness.The trace covers path length, displacement magnitude and vector, accumulated rotation, and average speed; vector consistency checks both magnitude and direction.Source: §5.2 · p. 9
  • The policy is optimized with group-relative advantages and a clipped GSPO objective, combining answer supervision with kinematic process rewards.The objective includes KL regularization toward a reference policy.Source: §5.2 · p. 10

6 Experiments

TempoVista improves Dyn-3D motion-reasoning accuracy across two base VLMs and generally transfers to selected VSI-Bench spatial tasks. Ablations show additional gains from kinematic rewards and oracle SE(3)-based frame sampling.

6.1 Dyn-3D-Instruct Dataset

  • Dyn-3D-Instruct comprises 9,600 SFT QA samples and 24,000 RL samples; RL spans 24 task types and pairs each item with five-field kinematic metadata for Kinematic-GSPO.Source: §6.1 · p. 10

6.2 Experimental Setup

  • Evaluation uses unseen Dyn-3D scenes and trajectories plus selected VSI-Bench tasks, comparing proprietary and open-source VLMs and TempoVista on InternVL-3.5-8B and Qwen3-VL-8B-Instruct.Source: §6.2 · pp. 10–11

6.4 Spatial Understanding Evaluation

  • On VSI-Bench, TempoVista lifts Qwen3-VL-8B-Instruct overall accuracy from 51.3% to 55.6%; InternVL-3.5-8B + TempoVista leads route planning at 34.5%, showing model- and task-dependent transfer.Source: Table 4 · p. 11, §6.4 · p. 11

6.3 Kinematic Perception Evaluation

  • TempoVista raises InternVL-3.5-8B Overall accuracy from 50.6% to 57.1% (+6.5 points) and Qwen3-VL-8B-Instruct from 50.3% to 61.2% (+10.9 points), exceeding Qwen-VL-Max by 11.4 points overall.Source: §6.3 · p. 11

6.5 Ablation Study

  • Kinematic-GSPO reaches 57.1% for InternVL-3.5-8B and 61.2% for Qwen3-VL-8B-Instruct, beating Base GSPO by 1.6 and 1.1 points, respectively; policy optimization supplies most improvement.Source: Table 5 · p. 12, §6.5 · p. 12
  • Oracle SE(3) sampling yields 57.1% for InternVL-3.5-8B + TempoVista and 61.2% for Qwen3-VL-8B-Instruct + TempoVista, 1.5 points above uniform for both; flow-proxy trails uniform by 0.5 and 0.7 points.Source: Table 6 · p. 12, §6.5 · p. 12

7 Conclusion · Appendix

TempoVista addresses kinematic collapse by grounding visual cues in metric 3D motion, and it outperforms leading open and proprietary models on Dyn-3D while generalizing to spatial tasks in VSI-Bench.

7 Conclusion

  • Kinematic collapse is a spatial cognition failure under nonsmooth trajectories, caused by VLM reliance on 2D visual shortcuts.The conclusion identifies this as the problem the work addresses.Source: §7 · p. 12
  • TempoVista outperforms leading open and proprietary models on Dyn-3D and generalizes to diverse spatial tasks in VSI-Bench.The paper argues these findings show explicit kinematic perception is essential for robust 3D spatial understanding.Source: §7 · p. 12
  • Dyn-3D evaluates this failure using counterfactual rendering to decouple visual changes from physical motion.Source: §7 · p. 12
  • TempoVista combines kinematic-adaptive frame selection with the Kinematic-GSPO algorithm.Source: §7 · p. 12
  • Kinematic-GSPO embeds metric motion constraints into policy optimization to ground visual cues in 3D space.Source: §7 · p. 12
  • The conclusion presents explicit kinematic perception as essential for robust 3D spatial understanding.Source: §7 · p. 12

A Dataset Construction Details

Dyn-3D is constructed from quality-filtered ScanNet++ scenes, controlled-trajectory renders, and metric camera and object metadata to separate physical motion from visual change.

Scene Filtering

  • Of 499 raw ScanNet++ scenes, 451 with complete segments_anno.json annotations are retained to support deterministic semantic and spatial question generation.The annotations provide object semantics, 3D bounding boxes, visibility, and object-level spatial relations.Source: A Dataset Construction Details · p. 17
  • After quality filtering and validation, 443 scenes form the eligible pool: 263 train, 167 held out, and 13 reserved, with no train–evaluation overlap.Manual inspection removes scenes with failed or distorted reconstructions, broken trajectory videos, or inadequate frame quality; dataset-level QA, media, and geometry validation removes four more.Source: A Dataset Construction Details · p. 17

Reconstruction and Rendering

  • For retained scenes, fish-eye images are undistorted, resolution standardized to 1752×1168, and 3D Gaussian Splatting models trained for 30,000 iterations before trajectory rendering.Each scene averages 261 valid images; training uses nerfstudio’s splatfacto implementation on dual RTX 4090 GPUs.Source: A Dataset Construction Details · p. 17

Counterfactual Trajectory Design

  • Five trajectory types are rendered at 30 FPS; fast, smooth, and slow videos share one spatial path but vary temporal sampling across 5, 15, and 30 seconds.The shared path is generated by clustering camera poses and applying cubic spline interpolation; the videos contain 150, 450, and 900 frames, respectively.Source: A Dataset Construction Details · p. 17
  • Two diagnostic trajectories contrast a 360-degree in-place rotation with strong visual change but almost zero displacement against translation with clear displacement and a relatively stable viewpoint.They test whether models mistake pixel-level variation for true metric motion.Source: A Dataset Construction Details · p. 17

3D Metadata Extraction

  • Metadata maps reconstructed scenes back to metric coordinates and records camera motion, object geometry and semantics, visibility, egocentric relations, and frame-wise poses.Visibility is estimated by projecting 3D bounding-box vertices into the camera frustum; 23 non-interactive background categories are removed.Source: A Dataset Construction Details · p. 17

B Benchmark Details

Dyn-3D defines 19 deterministic, four-option question types across metric kinematic perception, object-level spatial understanding, and trajectory reasoning. Its rotation and translation traps test whether models infer physical camera motion rather than rely on visual change alone.

Benchmark Details

  • The benchmark separates 19 question types into metric kinematic perception (B1–B6, B19), object-level spatial understanding (B7–B16), and trajectory-pattern inference (B17–B18).Each type has four options and is generated deterministically from extracted 3D metadata. B4 and B5 diagnose Kinematic Collapse by testing displacement overestimation under large visual changes and underestimation under weak visual changes, respectively.Source: B Benchmark Details · pp. 18–19
  • B1–B6 assess displacement magnitude, camera-motion direction, rotation angle, rotation-versus-translation traps, and average camera-speed category.The traps test whether models mistake rotation-driven appearance changes for displacement or miss real translation when appearance remains stable.Source: B Benchmark Details · p. 18
  • B7–B12 test object directions at video start and end, distance trends, nearest-object identification, coordinate re-anchoring, and relative distances.These tasks evaluate spatial judgments about queried objects under egocentric camera motion or a redefined coordinate frame.Source: B Benchmark Details · p. 18
  • B13–B16 cover conditional object counting, physical-size comparisons, cross-frame causal direction, and visibility after accounting for depth and occlusion.Object size is judged using 3D bounding-box metadata.Source: B Benchmark Details · p. 18
  • B17–B18 require recognizing whether a trajectory is rotation-, translation-, or mixed-motion-dominated and identifying its geometric shape.Shape examples include straight, curved, and circular paths.Source: B Benchmark Details · p. 18
  • B19 measures total camera travel accumulated along the trajectory, complementing the benchmark’s other metric kinematic-perception tasks.Unlike straight-line displacement, this question targets total path length.Source: B Benchmark Details · p. 18

C Method Details

Kinematic-GSPO computes a kinematic reward from five ground-truth motion quantities, then combines it with option-answer and formatting rewards. Field-level scoring uses quantized scalar and vector metrics, with explicit handling for missing outputs.

Reward formulation

  • Kinematic-GSPO averages five field-level rewards, assigns −1 to missing or unparsable fields, and uses γ = 0.1 and λ = 0.1; base GSPO sets λ = 0.The five quantities are total path length, displacement magnitude, displacement vector, accumulated rotation angle, and average speed.Source: C Method Details · p. 19

Scalar Metrics

  • Scalar metrics use four bins with thresholds {0.5, 4, 9}m for path length, {0.5, 1.5, 3}m for displacement magnitude, {15°, 360°, 600°} for rotation, and {0.15, 0.4, 0.8}m/s for speed.Exact-bin matches receive +1, one-bin errors receive 0, and larger deviations or missing outputs receive −1.Source: C Method Details · p. 19

Vector Metric

  • The displacement-vector reward averages a magnitude score using displacement bins and a direction score based on cosine similarity.Direction scores are +1 for cos ≥ 0.866, 0 for 0.5 ≤ cos < 0.866, and −1 for cos < 0.5; near-zero vectors score +1 only if both norms are below 0.5m.Source: C Method Details · p. 19

Formatting Weight

  • The formatting reward Rfmt uses a scaling factor γ = 0.1 in the total reward formulation.Kinematic-GSPO also uses γ = 0.1 for its reward setting.Source: C Method Details · p. 19

C.1 Implementation Details

The implementation fixes frame budgets and adaptation protocols, then specifies Kinematic-GSPO optimization, reward design, and evaluation and parsing procedures.

Frame Selection and Input Budget

  • Dyn-3D uses 8 oracle SE(3)-FPS frames per video, α = 0.7, and Dmax set to the 95th-percentile camera-center distance plus 10−6; VSI-MCQ uses 32 cached frames.The translation term is clipped at one; SFT and RL examples also contain eight frames, and VSI-MCQ results use a 2,490-question test set.Source: §C.1 · p. 19

SFT and Parameter-Efficient Adaptation

  • Both model families receive one-epoch answer-only SFT on 9,600 eight-frame examples, and both RL variants start from the same merged SFT checkpoint.SFT uses effective batch size 16 and learning rate 2 × 10−5; LoRA rank 32 and scaling 64 are applied to language-model projections while vision encoders remain frozen.Source: §C.1 · p. 19

GSPO Optimization

  • Each GSPO rollout samples 4 responses for each of 12 prompts (48 total), with temperature 1.0, top-p 0.95, and a 192-token response cap.Training uses two GPUs, one update micro-batch per GPU, AdamW at 10−5 learning rate and 10−2 weight decay, cosine scheduling with 150 warmup steps, and at most 3,000 optimization steps.Source: §C.1 · pp. 19–20
  • GSPO uses sequence-averaged gspo_token loss with a sequence-level importance ratio and token-level gradient correction; the instructed XML trace requests explicit kinematic fields.The prompt requests a <reasoning> block with truth, cot, and answer lines; the truth line lists path length, displacement magnitude and vector, accumulated rotation angle, and average speed with units.Source: §C.1 · p. 20

Reward and Missing Fields

  • Kinematic-GSPO combines format, kinematic-field, and answer rewards, while Base GSPO sets λ = 0 and keeps the other settings unchanged.Kinematic-GSPO uses γ = 0.1 and λ = 0.1; missing reasoning or kinematic fields receive −1, and answer reward is 1 for a correct option and 0 otherwise.Source: §C.1 · p. 20

Evaluation Prompt and Parsing

  • Evaluation uses greedy generation and ordered answer-letter parsing; VSI-MCQ additionally matches normalized free-form responses to a unique option text.Dyn-3D parse rates refer to the answer-letter parser, not strict XML-format compliance.Source: §C.1 · p. 20

D Error Analysis

TempoVista improves Dyn-3D performance and transfers positively but unevenly to VSI-Bench, while failures persist in egocentric depth, coordinate re-anchoring, and fine-grained motion estimation.

Dyn-3D results

  • 62.6% kinematic-perception accuracy for Qwen3-VL-8B-Instruct (up from 51.0%) anchors TempoVista’s gains across all three Dyn-3D dimensions and both evaluated base models.For Qwen3-VL-8B-Instruct, spatial-understanding accuracy rose from 47.6% to 58.3% and trajectory-reasoning accuracy from 71.4% to 81.8%; for InternVL-3.5-8B, the corresponding gains were 52.0% to 64.5%, 47.8% to 52.7%, and 70.4% to 77.7%. All compared Dyn-3D models had a 100% answer parse rate, so the reported failures concern reasoning rather than response formatting.Source: D Error Analysis · p. 20

Egocentric Depth-Axis Inversion

  • The model sometimes reverses front–rear relations while preserving left–right, suggesting egocentric depth-axis confusion rather than object misidentification.For single-frame questions, this is better interpreted as depth-axis confusion than temporal updating failure; across frames, incorrect propagation of the camera-centered coordinate frame may also contribute.Source: D Error Analysis · p. 20

Coordinate Re-Anchoring

  • Coordinate re-anchoring remains unreliable: after defining a bed-centered frame with the bed-to-sofa direction as positive Y, the model predicts Rear-Left instead of Front-Right for a charger.Solving the example requires grounding the referenced objects, constructing the requested frame, and transforming the queried object’s position; these operations are not reliably composed even when objects are visually identifiable.Source: D Error Analysis · pp. 20–21

Fine-Grained Motion Estimation

  • Fine-grained motion errors persist: the model misses lateral displacement in a camera-motion example and assigns Moderate speed to a near-stationary trajectory.These cases show that visible image variation can still be confused with actual camera displacement or physical speed, despite aggregate gains from kinematic supervision.Source: D Error Analysis · p. 21

Transfer to Real-World Spatial Reasoning

  • 55.6% overall VSI-Bench accuracy for Qwen3-VL-8B-Instruct (up from 51.3%) indicates positive transfer, with the largest gain in appearance-order reasoning.Relative-distance accuracy rose from 53.5% to 56.2%, relative-direction accuracy from 47.9% to 51.7%, appearance-order accuracy from 60.2% to 67.8% (a 7.6-point gain), and route-planning accuracy from 32.0% to 34.0%. Qualitative examples still show a sofa-direction error and an incorrect initial route turn propagating into later decisions.Source: D Error Analysis · p. 21
  • 50.7% overall VSI-Bench accuracy for InternVL-3.5-8B (up from 50.1%) shows modest, task-dependent transfer, including a slight decline in relative-direction accuracy.Relative-distance accuracy rose from 52.4% to 54.4%, appearance-order accuracy from 55.1% to 55.7%, and route-planning accuracy from 32.9% to 34.5%, while relative-direction accuracy fell from 48.6% to 48.1%.Source: D Error Analysis · p. 21

Remaining limitations

  • Improved motion grounding does not resolve every spatial operation; egocentric depth disambiguation, explicit coordinate re-anchoring, and fine-grained metric motion estimation remain open limitations.The quantitative and qualitative findings show that stronger spatial reasoning can co-occur with improved motion grounding, but kinematic supervision alone is insufficient.Source: D Error Analysis · p. 21

Related papers

Loading 2609.01059v1…