acceptodds
Under review as a conference paper at ICLR 2027

DynThinker: Unifying Static 3D and Dynamic 4D Spatial Reasoning

Abstract

Achieving general spatial intelligence requires a unified understanding of both static 3D and dynamic 4D. However, we find that existing models struggle to excel at both: 3D post-training can weaken dynamic understanding, while 4D-specialized models often lack strong static 3D reasoning. Our analysis reveals that this spatial–dynamic interference is highly non-uniform across parameter groups, with the balance between static and dynamic capabilities particularly sensitive to changes in late-layer MLPs. To address this interference, we propose DynThinker, a unified framework for jointly strengthening static and dynamic spatial reasoning. Decoupled Exponential Moving Average (D-EMA) applies configurable group-wise averaging to constrain drift in interference-sensitive pretrained parameters while preserving flexibility for capability enhancement. Spatial-Dynamic Fusion (SDF) complements D-EMA by selectively integrating spatial and temporal priors from a dynamic geometry encoder into intermediate VLM representations under semantic guidance. Across ten metrics spanning nine static 3D and dynamic 4D benchmarks, DynThinker achieves the highest average score among all evaluated models. Notably, DynThinker achieves 56.0 on ReVSI, 67.2 on DSR-Bench, and 51.2 and 40.1 under the sample-wise and group-wise evaluation protocols of DSI-Bench, respectively, setting new state-of-the-art results among the compared open-source models across all four evaluations within a unified model.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Spatial Intelligence, Multimodal Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_H8VvOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work studies spatial–dynamic interference in current VLMs, where post-training for static 3D and dynamic 4D reasoning can degrade the other capabilities. To address this issue, this paper introduces DynThinker, a framework for jointly enhancing 3D static and 4D dynamic spatial reasoning by Decoupled Exponential Moving Average (D-EMA) and Spatial-Dynamic Fusion (SDF). D-EMA maintains averaged weights for pretrained parameter groups with configurable averaging horizons, while excluding newly introduced modules from averaging. SDF integrates spatial and temporal features through importance-guided gating. Trained on approximately 2.68 million samples comprising 3D spatial reasoning, 4D temporal reasoning, and general video reasoning data, DynThinker achieves an average score of 60.5 across the evaluated static and dynamic reasoning metrics, outperforming the compared baselines.

Critical strengths
  1. The paper is well motivated and clearly written. It investigates an important limitation of existing 3D and 4D specialized VLMs: improving spatial/dynamic reasoning capability can come at the expense of the other. These findings motivate preserving pretrained capabilities while learning complementary spatial and dynamic knowledge with D-EMA and SDF.

  2. The proposed components are simple and effective for unified spatial reasoning. SDF integrates spatial and temporal features through importance-gated addition and residual injection, and D-EMA supports capability preservation through configurable group-wise temporal averaging. Combined with training on a mixture of static spatial and dynamic reasoning data, DynThinker achieves strong performance across 3D and 4D reasoning benchmarks.

Critical weaknesses
  1. Component ablations under the main training regime. Tables 3 and Table 10 provide useful component ablations and data scaling, but these experiments do not use the 2.68M-sample, 32-frame configuration underlying the main results. Given the substantial benefits of large-scale 3D spatial and 4D dynamic training data, it is better to evaluate the contributions of D-EMA and SDF under the main training regime.
Questions for authors

Please see the weaknesses above.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer H8Vv

We thank the reviewer for the positive and constructive feedback. We appreciate your recognition of our motivation, the effectiveness of D-EMA and SDF, and DynThinker's strong performance across static 3D and dynamic 4D benchmarks. To address your concern, we provide additional component ablations under the full training regime.


W1 (Ablations under the main training regime)
"...it is better to evaluate the contributions of D-EMA and SDF under the main training regime."

We conduct component ablations under identical training settings (2.68M samples, 32 frames), evaluating the contribution of D-EMA and SDF separately. As shown in Table A, removing D-EMA or SDF individually lowers the ten-metric average from 60.3 to 59.4 and 57.2, respectively.

Table A. Ablations under the main Qwen3-VL-8B training regime.

ConfigurationAVGVSIMMSIMindCubeViewSp.SITECVVLM4DDSRDSI-SDSI-G
DynThinker60.369.331.590.348.059.487.160.067.750.539.1
w/o D-EMA59.470.632.494.247.456.085.554.475.944.932.4
w/o SDF57.269.127.881.047.659.686.758.872.642.826.6

Removing D-EMA reduces VLM4D by 5.6, DSI-S by 5.6, and DSI-G by 6.7 points. SDF contributes to both static and dynamic reasoning, with its removal reducing MindCube by 9.3, MMSI by 3.7, DSI-S by 7.7, and DSI-G by 12.5 points.

Overall, the full model achieves the highest average, supporting the complementary contributions of D-EMA and SDF under the full training regime. We have added these results to Appendix D.2 (highlighted in red).

Thank you again for your positive assessment of our work and for suggesting this comparison. We hope the new ablations address your concern, and we would be happy to clarify any remaining questions.

Reviewer_H8VvOpenReview ↗

Reply

Thank you to the authors for the reply. Most of my concerns are addressed. I recommend that the authors incorporate these experiments into the final version, and I maintain my positive score.

AuthorsOpenReview ↗

Thank You for Your Positive Assessment

We thank the reviewer for the positive assessment and continued support. We are glad that the additional experiments address most of your concerns. Following your suggestion, we will incorporate these experiments into the final version. We appreciate your constructive feedback, which has helped strengthen our paper.

Reviewer_oVmeOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the loss of dynamic spatial reasoning during 3D-focused VLM post-training. It proposes D-EMA for selective parameter averaging and SDF for integrating spatial and temporal features from VDPM. Evaluation covers static and dynamic spatial benchmarks, with component ablations, data-scaling experiments, and a LIBERO transfer study.

Critical strengths
  1. The interference analysis is useful. Checkpoint interpolation and parameter restoration provide concrete evidence that spatial specialization can hurt dynamic reasoning, with much of the recoverable degradation concentrated in later MLP blocks.

  2. The evaluation covers both static and dynamic capabilities. The model achieves a better overall balance than the compared specialized models, and the controlled ablations show benefits from both parameter averaging and geometry fusion.

Critical weaknesses
  1. The evidence for group-specific averaging is mixed. Table 9 favors a shared averaging budget for Qwen3-VL, and the main model also uses a shared decay for pretrained parameters. This weakens the connection between the late-MLP sensitivity analysis and the method used for the main results.

  2. The main comparison does not fully separate method gains from training-data gains. DynThinker adds substantial dynamic and general-video data relative to GeoThinker. The smaller controlled experiments are helpful, but do not establish how much of the final improvement comes from D-EMA and SDF under the full training recipe.

Questions for authors

/

3 comments
AuthorsOpenReview ↗

Reply to Reviewer oVme

Thank you for your thoughtful and encouraging review. We appreciate your recognition of our spatial–dynamic interference analysis, particularly the role of late-layer MLPs, and DynThinker's balanced performance across static and dynamic benchmarks, supported by controlled ablations of D-EMA and SDF. To address your concerns, we provide additional budget-allocation results and same-data comparisons with the baseline.


W1 (Sensitivity analysis to shared budgets on Qwen3-VL)
"...This weakens the connection between the late-MLP sensitivity analysis and the method used for the main results."

Thank you for pointing out this connection. Our sensitivity analysis motivated the parameter grouping, which guided the design of budget ablations to select suitable decay settings for different backbones and target benchmarks. For Qwen3-VL-8B, we selected a shared decay based on its competitive performance on VSI-Bench and VLM4D, our primary benchmarks for configuration selection.

Table A. Qwen3-VL-8B budget allocations. DSR (another 4D benchmark) results are added here.

Late MLPsOthersDSRVSI-BenchVLM4D
12.5%12.5%29.1658.1661.78
12.5%50%31.4058.0656.05

1. Group-specific budgets for pretrained weights.

We selected the shared 12.5%/12.5% budget for Qwen3-VL-8B based on its performance on VSI-Bench and VLM4D (Appendix F.2). Different evaluation priorities may favor different allocations. For example, the non-uniform 12.5%/50% setting improves DSR by 2.24 points and the average of DSR and VSI-Bench from 43.66 to 44.73 (Table A), illustrating the trade-off between capabilities.

2. Shared decay in the main model.

Given the results on VSI-Bench and VLM4D and the high cost of large-scale training (2.68M samples, ~3,072 GPU-hours per run), we carried the shared-decay strategy into large-scale training. A finer search over the decay rates for late MLPs and the remaining weights could further improve performance across a wider range of target benchmarks.

3. Non-uniform decays on Qwen2.5-VL-7B.

As shown in Appendix F.2, non-uniform allocation (late MLPs: 12.5%, others: 100%) achieves a better VSI-Bench/VLM4D balance than shared budgets. Compared with the shared 100% budget, it improves both VSI-Bench (49.65 → 50.37) and VLM4D (46.64 → 48.56), while the shared 12.5% budget limits VSI-Bench to 34.44. These results suggest that sensitivity analysis helps identify parameter groups worth tuning separately, providing flexibility to balance capabilities across backbones and target benchmarks.


W2 (Method gains vs. data gains)
"The main comparison does not fully separate method gains from training-data gains..."

Thank you for this suggestion. We separate method gains from training-data gains through two controlled comparisons: DynThinker improves AVG from 49.1 to 50.6 using GeoThinker's data, while removing D-EMA or SDF under the full training recipe reduces performance.

Same data.

Table B compares the two models on GeoThinker's training data. DynThinker improves six metrics, including VLM4D (+8.9) and DSI-S (+5.9), and trails GeoThinker on VSI, MMSI, MindCube, and DSR.

Table B. Same-data comparison with GeoThinker (Qwen3-VL-8B). Both methods are trained on GeoThinker's training data.

ModelAVGVSIMMSIMindCubeViewSp.SITECVVLM4DDSRDSI-SDSI-G
GeoThinker49.172.630.983.045.955.985.142.430.933.011.0
DynThinker50.670.029.882.647.657.786.951.327.238.913.7

Full training setup.

We conduct component ablations under identical training settings (2.68M samples, 32 frames), removing one component at a time. Removing D-EMA lowers the average from 60.3 to 59.4, including drops on VLM4D and both DSI metrics. Removing SDF lowers the average to 57.2 (Table C).

Table C. Component ablations under the main Qwen3-VL-8B training setup.

ConfigurationAVGVSIMMSIMindCubeViewSp.SITECVVLM4DDSRDSI-SDSI-G
DynThinker60.369.331.590.348.059.487.160.067.750.539.1
w/o D-EMA59.470.632.494.247.456.085.554.475.944.932.4
w/o SDF57.269.127.881.047.659.686.758.872.642.826.6

We have added the same-data comparison in Appendix D.4 and the component ablations in Appendix D.2, highlighted in red.

Thank you again for recognizing our contributions and for your helpful suggestions. We would be happy to provide any further clarification.

Good work. I have no further questions

AuthorsOpenReview ↗

Thank You for Your Positive Feedback

We sincerely thank the reviewer for their positive feedback and continued support. We are glad that our additional experiments have addressed your concerns, and will incorporate them into the final version. We greatly appreciate your constructive suggestions for improving our work.

Reviewer_83fQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates spatial–dynamic interference in VLMs, where static 3D specialization may degrade dynamic 4D reasoning. The authors propose DynThinker, combining Decoupled Exponential Moving Average (D-EMA) for capability preservation and Spatial-Dynamic Fusion (SDF) for integrating spatial and temporal geometric priors. Experiments across nine benchmarks demonstrate strong overall performance, supported by ablation studies and parameter sensitivity analyses.

Critical strengths
  1. The checkpoint interpolation and parameter restoration experiments reveal non-uniform sensitivity across parameter groups, particularly later MLP layers. This provides useful empirical motivation for capability-preserving post-training.

  2. DynThinker achieves competitive results across diverse static and dynamic benchmarks, with substantial improvements over GeoThinker on several dynamic reasoning tasks. The extensive evaluations demonstrate the practical potential of unified spatial reasoning.

Critical weaknesses
  1. Insufficient isolation of D-EMA's contribution. Although D-EMA is motivated by group-specific parameter sensitivity, the main Qwen3-VL experiments use shared EMA decay factors for pretrained parameters. Consequently, these results do not directly establish the benefit of sensitivity-aware group-wise averaging. Moreover, the comparison with conventional EMA uses a single decay setting without sufficiently matching the averaging budget or treatment of new parameters. Stronger controlled comparisons are needed to establish the proposed mechanism's distinct contribution.

  2. Concerns regarding baseline comparability. Table 7 shows a substantial discrepancy between the officially reported Qwen3-VL-8B VLM4D score (52.2) and the authors' evaluation (60.07). Rescoring released predictions does not fully resolve potential differences in inference settings. Additionally, 4DThinker is categorized as a 3D-specialized method, and its reported DSR score (9.3) differs substantially from its original results. These issues weaken the reliability of some cross method comparisons.

Questions for authors
  1. For Qwen3-VL, what is the benefit of group-specific D-EMA over carefully tuned uniform EMA applied only to pretrained parameters?

  2. Can the authors explain the large VLM4D baseline discrepancy and provide comparisons under matched inference configurations?

  3. Why is 4DThinker classified as 3D-specialized, and what explains its unusually low DSR score?

  4. Can the authors provide controlled comparisons between SDF and simpler fusion strategies, such as direct feature addition or fixed gating?

4 comments
AuthorsOpenReview ↗

Reply to Reviewer 83fQ (Part 1)

We sincerely thank you for your detailed and insightful review. We appreciate your recognition of our parameter sensitivity analysis and strong performance across static and dynamic benchmarks, which support our core contribution of mitigating spatial–dynamic interference while preserving complementary capabilities. To address your concerns about D-EMA's contribution and baseline comparability, we provide matched-budget EMA comparisons, VLM4D evaluation, corrected 4DThinker results, and additional SDF ablations.


Q1 & W1 (D-EMA vs. uniform EMA on Qwen3-VL)
“For Qwen3-VL, what is the benefit of group-specific D-EMA over carefully tuned uniform EMA applied only to pretrained parameters?”

Thank you for the suggestion. D-EMA differs from conventional EMA in two aspects: it excludes newly introduced SDF parameters from averaging and supports group-specific decay rates for pretrained parameters. The following comparisons examine the benefits of decoupling and group-specific budget allocation.

1. Excluding SDF from EMA.

Pretrained weights contain learned knowledge worth preserving, whereas SDF modules are newly initialized and benefit from unconstrained learning. We compare two runs under identical training settings and a matched 12.5% averaging budget for pretrained weights, differing only in whether the newly introduced SDF modules are averaged.

Excluding SDF from averaging improves VSI-Bench by 1.63 points and VLM4D by 6.49 points compared with averaging all weights. This comparison is also reported in Appendix F.2.

Table A. EMA scope on Qwen3-VL-8B with 298K training samples and 32-frame evaluation.

Averaged parametersVSI-BenchVLM4D
D-EMA: Pretrained weights only58.1661.78
EMA: All weights, including SDF56.5355.29

2. Group-specific budgets for pretrained weights.

Group-specific budgets allow D-EMA to balance different capabilities flexibly according to the backbone and target benchmarks.

Table B. Qwen3-VL-8B budget allocations. DSR (another 4D benchmark) results are added here.

Late MLPsOthersDSRVSI-BenchVLM4D
12.5%12.5%29.1658.1661.78
12.5%50%31.4058.0656.05

As shown in Table B, on Qwen3-VL-8B, increasing the budget for remaining pretrained weights from 12.5% to 50%, while keeping late MLPs at 12.5%, improves DSR by 2.24 points but reduces VLM4D by 5.73 points, with comparable VSI-Bench performance. Different evaluation priorities may favor different allocations. When prioritizing DSR, the group-specific allocation (12.5%/50%) performs better, highlighting D-EMA's flexibility in balancing different capabilities.

On Qwen2.5-VL-7B (Appendix F.2, Table 11), non-uniform allocation (late MLPs: 12.5%, others: 100%) achieves a better VSI-Bench/VLM4D balance than shared budgets. Compared with the shared 100% budget, it improves both VSI-Bench (49.65 → 50.37) and VLM4D (46.64 → 48.56), while the shared 12.5% budget limits VSI-Bench to 34.44.

Together, these results highlight D-EMA's flexibility in navigating capability trade-offs across backbones and benchmarks.

3. Shared decay in the main model.

Given the results on VSI-Bench and VLM4D and the high cost of large-scale training (2.68M samples, ~3,072 GPU-hours per run), we carried the shared-decay strategy into large-scale training. A finer search over the decay rates for late MLPs and the remaining weights could further improve performance across a wider range of target benchmarks.


AuthorsOpenReview ↗

Reply to Reviewer 83fQ (Part 2)

Q2 & W2 (VLM4D evaluation across protocols)
"Can the authors explain the large VLM4D baseline discrepancy and provide comparisons under matched inference configurations?"

Thank you for raising this point. To our knowledge, the 52.2 score for Qwen3-VL-8B on VLM4D is reported in MLLM-4D[1]. Differences in the prompt template explain most of the gap in our controlled comparison.

With the checkpoint, visual input, decoding, and answer extraction held fixed, switching from MLLM-4D's prompt template to ours raises Qwen3-VL-8B accuracy from 53.19 to 60.13, close to the 60.07 reported in our manuscript. Using MLLM-4D's released inference settings and extractor, we obtain 52.31, close to their reported 52.2 [1].

Our evaluation follows the official Qwen3-VL template [2, Appendix B.6], consistent with VSI-Bench in LMMs-Eval. We reevaluate key baselines and API-based models under a unified protocol to ensure fair comparisons. DynTrace (ACM MM 2026) [3] also reports a comparable Qwen3-VL-8B VLM4D score of 59.0.

To check whether the comparison depends on the choice of prompt or protocol, we evaluate MLLM-4D and DynThinker under both protocols, using frame inputs with matched prompts, decoding, and answer extraction within each protocol (Table C).

Table C. VLM4D accuracy under matched evaluation protocols (%).

ProtocolModelOverall
MLLM-4DQwen3-VL-8B52.31
MLLM-4DMLLM-4D-RFT57.76
MLLM-4DDynThinker (main)55.84
MLLM-4DDynThinker‡63.05
OursQwen3-VL-8B60.07
OursMLLM-4D-RFT59.36
OursDynThinker (main)60.07
OursDynThinker‡66.74

‡ DynThinker checkpoint trained with the same training data as MLLM-4D, distinct from the main model for balanced static and dynamic reasoning.

As shown in Table C, DynThinker‡ outperforms MLLM-4D-RFT by 5.29 points under the MLLM-4D protocol and 7.38 points under ours. The main DynThinker model achieves 60.07 under our protocol, reflecting its focus on balanced static and dynamic reasoning. We have added these results to Appendix E.1 (highlighted in red).


Q3 & W2 (4DThinker category and DSR score)
"Why is 4DThinker classified as 3D-specialized, and what explains its unusually low DSR score?"

Thank you for pointing out the classification and flagging the DSR score. We have moved 4DThinker to the 4D-specialized category in Table 1 and rechecked its evaluation.

In rechecking the evaluation, we found that our direct-answer parser could miss the final option in 4DThinker's reasoning outputs. Re-evaluating the released checkpoint with its native prompt and final-answer extraction yields 31.85 on DSR, compared with 34.2 in its original report.

DynThinker still substantially outperforms it on DSR (67.7) and in the updated AVG (60.30 vs. 29.63). We have corrected Table 1 accordingly (highlighted in red).


Q4 (SDF vs. simpler fusion)
"Can the authors provide controlled comparisons between SDF and simpler fusion strategies, such as direct feature addition or fixed gating?"

Thank you for this suggestion. Under matched training settings, SDF achieves 58.16 on VSI-Bench and 61.78 on VLM4D, outperforming direct feature addition (37.00/41.40; Table D).

Table D. SDF gating ablation (Qwen3-VL-8B). All variants share the same training data and schedule.

FusionVSIVLM4D
SDF (ours)58.1661.78
Direct addition + global scaling56.8859.64
Direct addition37.0041.40

Keeping global scaling while replacing token-level importance gating with direct addition gives 56.88 on VSI-Bench and 59.64 on VLM4D. Further removing global scaling leads to a substantial performance drop. This comparison shows the contribution of both choices in the tested setup. We have added this ablation to Appendix F.2 (highlighted in red).

Thank you again for your constructive feedback. We would be happy to clarify any remaining questions.


References

[1] Yin X, Li C, Chang J, et al. MLLM-4D: Towards visual-based spatial-temporal intelligence.

[2] Qwen Team. Qwen3-VL Technical Report.

[3] Rongxin Gao et al. DynTrace: Tracking Dynamic Object Evidence for 4D Spatio-Temporal Reasoning in MLLMs.

AuthorsOpenReview ↗

Follow-up on Our Response and Revised Manuscript

Thank you once again for your time and constructive feedback on DynThinker. Your comments have helped us strengthen the paper through additional D-EMA comparisons, matched-protocol evaluations, and SDF ablations, as well as corrections to the 4DThinker results. We have incorporated these additional results and corrections into the revised manuscript.

As the discussion period draws to a close, we would greatly appreciate any further feedback you may have on our response and additional experiments. We hope these have addressed your concerns and would be happy to clarify any remaining questions. Thank you again for your time and consideration!

Area_Chair_UUzUOpenReview ↗

Feedback

Hi, the review deadline is approaching. Please take a look at the authors' rebuttal and post your feedback (and update your rating if appropriate) at your earliest convenience. Thank you! AC

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response – Thanks to the Area Chair and Reviewers

We sincerely thank the Area Chair and all reviewers for their careful reading and constructive feedback.

We appreciate the reviewers' recognition of our analysis of spatial–dynamic interference (83fQ, oVme), the simplicity and effectiveness of D-EMA and SDF in mitigating interference and facilitating spatial–dynamic fusion (H8Vv), and the resulting strong performance across static and dynamic benchmarks (83fQ, oVme, H8Vv).

To address the reviewers' concerns, we have strengthened the manuscript with additional analyses and controlled experiments:

  • D-EMA design and effectiveness (83fQ, oVme). Matched-budget comparisons clarify the benefits of D-EMA over standard EMA, while budget-allocation ablations justify its treatment of pretrained parameters and newly introduced modules, as well as the selected decay configurations (Appendix F.2).
  • Data fairness and component contributions (oVme, H8Vv). Under matched training data, our method outperforms GeoThinker (AVG: 50.6 vs. 49.1). Full-scale ablations using 2.68M samples further validate the contributions of both components: removing D-EMA or SDF reduces AVG from 60.3 to 59.4 and 57.2, respectively (Appendices D.4 and D.2).
  • Evaluation consistency and baseline corrections (83fQ). Prompt-controlled evaluations explain the VLM4D baseline discrepancy, with MLLM-4D and DynThinker evaluated under both protocols. We also correct the 4DThinker categorization and results using its native prompt and answer-extraction procedure (Appendix E.1 and Table 1).
  • Spatial–dynamic fusion paradigm (83fQ). Controlled comparisons with direct feature addition, with and without global scaling, demonstrate the effectiveness of SDF's selective integration of spatial and temporal priors (Appendix F.2).

All revisions are highlighted in red in the updated manuscript, with detailed responses provided to each reviewer.

We are encouraged that two reviewers have already provided positive follow-up feedback, indicating that our responses and additional experiments have addressed their main concerns. We appreciate their recognition and constructive suggestions, which have helped strengthen the paper. We thank the Area Chair and all reviewers again for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.