acceptodds
Under review as a conference paper at ICLR 2027

Möbius: A Unified Motion-Language Model Bridging Understanding and Generation

Abstract

We aim to build a truly unified motion model that can understand, generate, and edit human motion within a single framework. Existing motion UMMs remain only partially unified: motion understanding is handled by a language model, while generation is often delegated to a separate diffusion module operating over learned motion latents. We introduce **Möbius**, which brings the entire diffusion process into the same multimodal representation space used for motion understanding. Clean motion, noisy motion, and text jointly participate in shared attention, while role-specific parameters process understanding and generation streams. Importantly, **Möbius** operates directly on continuous frame-level raw motion, requiring neither discrete tokenization nor a pretrained motion VAE. This simple formulation naturally supports motion captioning, text-to-motion generation, temporal completion, and instruction-guided editing within a single model. On HumanML3D, **Möbius** achieves state-of-the-art performance among unified motion models. Scaling to 1.7B parameters and 1,700 hours of motion data further demonstrates that our framework is both effective and scalable.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Unified Motion-Language Model, Continuous Motion Modeling, Motion Editing, Multimodal Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CLtQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1(−1)
Read review
Paper summary

This paper introduces Möbius, a new motion-language model that unifies four tasks: motion understanding, generation, completion, and editing. Instead of existing partial motion-language integration, this model unifies motion understanding and denoising process in two seperate channels within the same latent space without VAE. It introduces two-stage training (first understanding, then generation) to effectively retain performance, with scaling to 1.7B parameters on 1,700 hours of motion data.

Critical strengths

The proposed simple yet effective architecture unifies understanding and generation within the same representation space, avoiding the burdens of training VAEs, while achieving strong results both in quantitative and qualitative experiments.

Critical weaknesses
  1. This work shows some examples in the qualitative study part, but is in lack of systematic human evaluation that can better support subjective visual quality
  2. The quantitative results in the tables do not report the multi-round 95% confidence intervals, which is the common practice for this field. Results might vary largely in different rounds.
Questions for authors
  1. Could you add a human preference evaluation to better show subjective visual quality?
  2. Are the current results from single-round evaluation runs? Could you report the results for multi-round 95% confidence intervals?
  3. Most existing text-to-motion evaluations use the established 263-dimensional evaluators from HumanML3D. For fairer comparison, could you also show the text-to-motion evaluation results by using the 263-style evaluators?
5 comments
AuthorsOpenReview ↗

Response to Reviewer CLtQ (Part1)

Thank you for your valuable comments. We have addressed your concerns through a double-blind user study, 20-round confidence interval evaluations, and evaluation under the 263-dimensional HumanML3D protocol.

W1 & Q1: Adding a systematic human preference evaluation

Thank you for this suggestion. Following your feedback, we conducted a new randomized, double-blind pairwise user study with 30 independent participants rating 90 motion pairs across two fixed prompt sets (30 randomly sampled HumanML3D prompts and 30 practical user prompts written by professional animation artists, with attention checks), yielding 360 pairwise judgments (Section 4.2, Table 4; Appendix B.2, Table S7). The results show that human raters consistently prefer Möbius-1.7B over both Möbius-0.6B and the large-scale baseline HY-Motion (1B), confirming the visual quality and instruction adherence of our generated motions:

Table 4: User study on text-to-motion generation (wins / ties / losses for Möbius-1.7B out of 120 ratings per row).

\hline \text{Compared model} & \text{Prompts} & \text{Semantics} & \text{Quality} & \text{Expressiveness} & \text{Overall} & \textbf{Score} \\\\ \hline \text{Möbius-0.6B} & \text{HumanML3D} & 12\\,/\\,98\\,/\\,10 & 40\\,/\\,61\\,/\\,19 & \text{--} & \mathbf{36\\,/\\,69\\,/\\,15} & \mathbf{58.8} \\\\ \text{Möbius-0.6B} & \text{User} & 109\\,/\\,11\\,/\\,0 & 42\\,/\\,67\\,/\\,11 & 111\\,/\\,7\\,/\\,2 & \mathbf{112\\,/\\,8\\,/\\,0} & \mathbf{96.7} \\\\ \hline \text{HY-Motion (1B)} & \text{User} & 49\\,/\\,39\\,/\\,32 & 33\\,/\\,47\\,/\\,40 & 49\\,/\\,41\\,/\\,30 & \mathbf{46\\,/\\,40\\,/\\,34} & \mathbf{55.0} \\\\ \hline \end{array}$$ * **Möbius-1.7B vs. Möbius-0.6B:** Raters prefer the 1.7B model on both HumanML3D prompts and practical artist prompts (**Table 4**). * **Möbius-1.7B vs. HY-Motion (1B):** On practical artist prompts (**Table 4**), raters prefer Möbius-1.7B overall (score 55.0), driven by clear advantages in semantic alignment and expressiveness, while HY-Motion holds a slight edge in raw motion quality. Please see **Section 4.2** (**Table 4**), **Appendix B.2** (**Table S7**), the **Ethics statement**, and the **supplementary material** (for rendered user-study cases) for the full results and protocol. ### W2 & Q2: Single-round vs. multi-round evaluation and 95% confidence intervals Thank you for highlighting this standard practice. While our initial **Table 1** followed MotionStreamer's single-run MS-272 protocol, we followed your suggestion and ran 20 independent evaluation rounds with different random seeds for our models and the strongest baselines to report 95% confidence intervals (**Appendix C.1**, **Table S9**). Across all 20 runs, the resulting 95% confidence intervals are very narrow on every metric and the means closely match our single-run numbers, confirming that our improvements over prior methods are statistically stable: **Table S9: Text-to-motion results on HumanML3D with 95% confidence intervals (MS-272 evaluator, 20 runs).** (Part of experiments.) $$\begin{array}{lcccc} \hline \text{Method} & \text{R@1}\uparrow & \text{R@3}\uparrow & \text{FID}\downarrow & \text{MMDist}\downarrow \\\\ \hline \text{GT} & 0.7703 & 0.9442 & 0.0000 & 14.8785 \\\\ \hline \text{MotionStreamer} & 0.6304^{\pm .004} & 0.8497^{\pm .003} & 12.2108^{\pm .078} & 16.5812^{\pm .017} \\\\ \text{MotionGPT3} & 0.6711^{\pm .003} & 0.8816^{\pm .001} & 20.9911^{\pm .067} & 17.5660^{\pm .015} \\\\ \text{HY-Motion (1B)} & 0.7898^{\pm .004} & 0.9517^{\pm .002} & 13.5197^{\pm .081} & 14.7750^{\pm .019} \\\\ \hline \textbf{Ours (0.6B)} & \mathbf{0.8208}^{\pm .003} & \mathbf{0.9671}^{\pm .003} & \mathbf{6.1289}^{\pm .051} & \mathbf{13.9977}^{\pm .017} \\\\ \textbf{Ours (1.7B)} & 0.7969^{\pm .004} & 0.9525^{\pm .003} & 10.6640^{\pm .070} & 14.7439^{\pm .019} \\\\ \hline \end{array}$$ Please see **Table 1** (caption) and **Appendix C.1** (**Table S9**) for the full 20-round confidence interval results.
AuthorsOpenReview ↗

Response to Reviewer CLtQ (Part2)

Q3: Text-to-motion evaluation using the established 263-dimensional HumanML3D evaluator

Thank you for this constructive suggestion. Following your advice, we converted our native 30-FPS SMPL-22 outputs to the traditional 263-dimensional format and ran an additional 20-round evaluation (with 95% confidence intervals) under the original 263-dimensional HumanML3D evaluator, alongside ground-truth test motions passed through the exact same conversion pipeline (GT 272→263, Appendix C.2, Table S10). The results show that even after cross-representation conversion, both Möbius-0.6B and Möbius-1.7B surpass real motion in R-Precision and MMDist and outperform most established baselines under the 263-dimensional evaluator:

Table S10: Text-to-motion results on HumanML3D with the original 263-dimensional evaluator. (Part of experiments.)

\hline \text{Method} & \text{R@1}\uparrow & \text{R@3}\uparrow & \text{FID}\downarrow & \text{MMDist}\downarrow & \text{MModality}\uparrow \\\\ \hline \text{Real motion} & 0.523 & 0.798 & 0.001 & 2.977 & \text{--} \\\\ \text{GT } 272\to263 & 0.486 & 0.780 & 0.381 & 3.126 & \text{--} \\\\ \hline \text{T2M-GPT} & 0.492^{\pm .003} & 0.775^{\pm .002} & 0.141^{\pm .005} & 3.121^{\pm .009} & 1.831^{\pm .048} \\\\ \text{MoMask} & 0.521^{\pm .002} & 0.807^{\pm .002} & 0.045^{\pm .002} & 2.958^{\pm .008} & 1.241^{\pm .040} \\\\ \text{MoGenTS} & 0.529^{\pm .003} & 0.812^{\pm .002} & 0.033^{\pm .001} & 2.867^{\pm .006} & 0.808^{\pm .036} \\\\ \text{MoMask++} & 0.517^{\pm .002} & 0.803^{\pm .002} & 0.069^{\pm .003} & 2.948^{\pm .007} & 1.192^{\pm .053} \\\\ \text{MoScale} & 0.540^{\pm .002} & 0.817^{\pm .002} & 0.046^{\pm .002} & 2.830^{\pm .005} & 0.873^{\pm .044} \\\\ \hline \textbf{Ours (0.6B)} & \mathbf{0.554}^{\pm .003} & \mathbf{0.843}^{\pm .002} & 0.342^{\pm .012} & \mathbf{2.740}^{\pm .008} & 1.146^{\pm .051} \\\\ \textbf{Ours (1.7B)} & 0.549^{\pm .004} & 0.837^{\pm .003} & 0.399^{\pm .019} & 2.800^{\pm .010} & \mathbf{1.715}^{\pm .060} \\\\ \hline \end{array}$$ * **Retrieval and multimodal distance:** Both Möbius-0.6B and Möbius-1.7B exceed real motion and outperform most prior methods on every R-Precision and MMDist metric (**Table S10**). * **Effect of representation conversion on FID:** Passing ground-truth test motions through the 30-FPS SMPL-22 to 263-dim conversion (*GT 272→263* in **Table S10**) already yields an FID of 0.381, while Möbius-0.6B achieves an even lower FID of 0.342 ± 0.012, showing that the cross-representation conversion itself sets a floor on 263-dim FID. Please see **Appendix C.2** (**Table S10**, referenced in **Table 1**, with MModality defined in **Appendix B.3**, **Table S8**) for the complete comparison. Thank you once again for your constructive feedback, which has substantially strengthened our empirical evaluation!

We sincerely appreciate the time and effort you have dedicated to reviewing our manuscript. We have tried our best to provide responses to all of your questions and we hope our responses would address your concerns.

As the discussion period nears its conclusion, we would be grateful for any final feedback you may have. Please don't hesitate to let us know if further clarifications would be helpful - we remain ready to provide additional details as needed.

Thank you again for your valuable insights and constructive feedback throughout this process.

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

The paper presents Mobius, a unified model for motion understanding, generation, editing, and completion. It directly models frame-level continuous motion and connects understanding with denoising through role-specific parameters and shared attention. It also adopts an understanding-first training strategy. Experiments show strong results across multiple motion-language tasks.

Critical strengths
  1. The core idea is reasonable. Role-specific parameters and shared attention offer a clean way to connect motion understanding and generation.
  2. The quantitative results and demo videos are convincing, showing good motion quality and instruction following.
  3. The same architecture supports multiple tasks, including motion captioning, generation, editing, and several types of temporal completion.
Critical weaknesses
  1. The architecture follows MoT ideas already explored in other multimodal domains. While this connection is cited, the paper could explain more clearly what new design choices are needed for motion.
  2. A generation-first, understanding-second result under the same role-routing architecture is missing, which would help assess the benefit of the proposed training order.
Questions for authors
  1. Could you report results for training generation first and then understanding, using the same role-routing architecture and a comparable training budget?
  2. Why does the 1.7B model perform worse than the 0.6B model on HumanML3D text-to-motion generation?
4 comments
AuthorsOpenReview ↗

Response to Reviewer afCa (Part1)

Thank you for your valuable comments. We have addressed your concerns through a structured summary of our motion-specific MoT designs, generation-first training ablations, and controlled scaling and user-study evaluations.

W1: Clarifying what new MoT design choices are needed specifically for motion

Thank you for pointing this out. We have added a more structured and systematic summary of the design choices of our method in our revised version (Section 3 and Appendix D.7). Specifically, we contrast our approach with image and video Mixture-of-Transformers (MoT) models (such as BAGEL and Cosmos3), which typically rely on separate pretrained encoders (e.g., a visual encoder for understanding and a VAE latent space for generation) and joint training on web-scale paired data. Because human motion lacks these conditions, our existing ablations (IDs 3, 5, 10–14 in Table 5) together with our newly added control variants (IDs 2, 8, and 9 in Table 5) show that adapting MoT to human motion requires four motion-specific design choices:

  • Routing by computational role rather than modality: In our encoder-free setup, clean observed motion and noisy denoising states share the exact same frame-level interface within a single sequence, but have very different feature statistics and training objectives. Routing them to the same motion parameters by modality (ID 5) causes motion understanding to collapse (M2T R@1 drops to 0.0328), whereas routing clean motion and text to understanding parameters and noisy motion to generation parameters (ID 14) preserves strong performance in both directions (M2T R@1 0.6001, T2M R@1 0.8207).
  • Encoder-free continuous frame-level interface: Existing motion tokenizers and VAEs introduce quantization error or temporal downsampling (IDs 2 and 3). Modeling continuous frames directly (ID 14) improves both captioning and generation FID (4.88 vs. 6.01 for VQ and 6.00 for VAE), while keeping a strict one-to-one frame alignment that allows temporal completion and editing to be formulated simply by changing which frames are observed.
  • Understanding-first parameter transfer for scarce data: Because paired motion-language data is orders of magnitude smaller than image-text data, the generation branch cannot learn rich motion-language semantics from scratch. Initializing the generation parameters from the trained understanding parameters and freezing the understanding branch (ID 14) clearly outperforms random initialization (ID 13, FID 4.88 vs. 5.88) and joint or reverse training schedules (IDs 8–11).
  • Layer-wise bidirectional shared attention: Generating accurate sub-actions and preserving observed keyframes require fine-grained temporal conditioning throughout denoising. Allowing noisy frames to attend bidirectionally to layer-wise understanding keys and values at every layer (ID 14) substantially outperforms conditioning only on final-layer context representations (ID 12, T2M R@1 0.8207 vs. 0.7621, FID 4.88 vs. 5.76).

W2 & Q1: Generation-first, understanding-second training under the same role-routing architecture and budget

Thank you for suggesting this control experiment. Under the exact same Qwen3-0.6B role-routed architecture and per-stage step budget as our understanding-first model (ID 14), we trained two new generation-first baselines—freezing understanding during Stage 1 generation (Gen. → Und. (F), ID 8) or updating it (Gen. → Und. (T), ID 9), with generation frozen in Stage 2—alongside Gen. → Align → Joint (ID 10). Comparing these variants in Table 5 (Section 4.4, Appendix D.8, Table S18) shows that reversing the training order substantially degrades performance:

Table 5: Ablation study on training schedules. (Part of experiments.)

\hline \text{ID} & \text{Training} & \text{T2M R@1}\uparrow & \text{FID}\downarrow & \text{M2T R@1}\uparrow & \text{CIDEr}\uparrow \\\\ \hline 7 & \text{Gen. only} & 0.7877 & 6.0073 & \text{--} & \text{--} \\\\ 8 & \text{Gen.}\to\text{Und.}^{\mathrm{F}} & 0.7567 & 8.1031 & 0.6001 & 0.5629 \\\\ 9 & \text{Gen.}\to\text{Und.}^{\mathrm{T}} & 0.7471 & 8.1985 & 0.5115 & 0.3916 \\\\ 10 & \text{Gen.}\to\text{Align}\to\text{Joint} & 0.8096 & 6.1139 & 0.5454 & 0.5010 \\\\ \hline \mathbf{14} & \textbf{Und.}\to\textbf{Gen.} & \mathbf{0.8207} & \mathbf{4.8772} & \mathbf{0.6001} & \mathbf{0.5629} \\\\ \hline \end{array}$$ Because generation attends to layer-wise understanding K/V at every layer, updating understanding in Stage 2 (**ID 8**) shifts the conditioning context away from what the frozen generation branch was trained on (**Eq. 9**), while updating understanding in Stage 1 (**ID 9**) fails to prevent this shift and harms understanding. Training understanding first and freezing it (**ID 14**) fixes a stable conditioning space before learning generation.
AuthorsOpenReview ↗

Response to Reviewer afCa (Part2)

Q2: Why does the 1.7B model perform worse than the 0.6B model on HumanML3D text-to-motion generation?

Thank you for this question. To thoroughly investigate why the 1.7B model obtains slightly lower automatic scores on HumanML3D T2M in our initial tables, we conducted a new controlled experiment training Möbius-1.7B solely on HumanML3D (Table S16), analyzed our 1,700-hour training corpus against HumanML3D (Table S1, Figure S1), and carried out a double-blind human preference study (Table 4), alongside our existing benchmark evaluations (Tables 1, 2, S10, and S15). Together, these existing and new results confirm that Möbius-1.7B actually delivers stronger overall quality and generalization, as consistently supported by our user study and multi-task evaluations. Training the 1.7B model on the much larger 1,700-hour corpus substantially strengthens its general motion generation and understanding capabilities; however, as also explicitly noted in recent large-scale motion studies such as LLaMo [1], Go to Zero (MotionMillion) [2], and ViMoGen [3], models scaled on broad, open-domain motion corpora do not necessarily achieve better automatic scores (especially FID) on the small HumanML3D test set, because HumanML3D's automatic metrics measure closeness to its narrow test distribution rather than true motion quality or open-domain generalization:

  • Controlling for training data: When we train Möbius-1.7B exclusively on HumanML3D under the same setup as the 0.6B model (Table S16), the two models achieve nearly identical HumanML3D T2M scores (FID 6.24 vs. 6.13, R@1 0.8161 vs. 0.8209), showing that the small HumanML3D dataset (<30 hours) already saturates at 0.6B. Adding our 1,700-hour corpus (60× larger, covering a much broader range of combat, dance, and stylized actions; Table S1, Figure S1) equips the 1.7B model with much stronger open-domain generalization, even though distribution-based HumanML3D metrics (such as FID) fail to reflect this broader capability [1, 2, 3].
  • Strong quantitative performance across benchmarks: Even on HumanML3D T2M, Möbius-1.7B still achieves higher R@1 than real ground-truth motion (0.7968 vs. 0.7703) and outperforms all other large-scale baselines in Table 1 (including HY-Motion and Go to Zero [2]) across R@1–3, FID, and MMDist. Moreover, it achieves higher per-prompt generation diversity (MModality 1.715 vs. 1.146 in Table S10), the best motion captioning results across all metrics (Table 2), and the lowest FID across all four temporal completion tasks (Table S15).
  • Superior human preference: In our double-blind user study (Table 4), human raters prefer Möbius-1.7B over the 0.6B model even on HumanML3D test prompts (overall preference score 58.8, winning 40 to 19 in motion quality), and overwhelmingly prefer the 1.7B model on practical prompts written by professional animators (112 wins, 8 ties, and 0 losses; score 96.7), while also preferring Möbius-1.7B overall over HY-Motion (1B):

Table 4: User study on text-to-motion generation (wins / ties / losses for Möbius-1.7B out of 120 ratings per row).

\hline \text{Compared model} & \text{Prompts} & \text{Semantics} & \text{Quality} & \text{Expressiveness} & \text{Overall} & \textbf{Score} \\\\ \hline \text{Möbius-0.6B} & \text{HumanML3D} & 12\\,/\\,98\\,/\\,10 & 40\\,/\\,61\\,/\\,19 & \text{--} & \mathbf{36\\,/\\,69\\,/\\,15} & \mathbf{58.8} \\\\ \text{Möbius-0.6B} & \text{User} & 109\\,/\\,11\\,/\\,0 & 42\\,/\\,67\\,/\\,11 & 111\\,/\\,7\\,/\\,2 & \mathbf{112\\,/\\,8\\,/\\,0} & \mathbf{96.7} \\\\ \hline \text{HY-Motion (1B)} & \text{User} & 49\\,/\\,39\\,/\\,32 & 33\\,/\\,47\\,/\\,40 & 49\\,/\\,41\\,/\\,30 & \mathbf{46\\,/\\,40\\,/\\,34} & \mathbf{55.0} \\\\ \hline \end{array}$$ Please see **Appendix D.5** (**Table S16**), **Appendix A.1** (**Table S1**, **Figure S1**), **Appendix C.2** (**Table S10**), and **Section 4.2** (**Table 4**) for the complete analysis. > [1] Li et al., *LLaMo: Scaling Pretrained Language Models for Unified Motion Understanding and Generation with Continuous Autoregressive Tokens*, arXiv:2602.12370, 2026. > [2] Fan et al., *Go to Zero: Towards Zero-shot Motion Generation with Million-scale Data*, ICCV 2025. > [3] Lin et al., *The Quest for Generalizable Motion Generation: Data, Model, and Evaluation*, arXiv:2510.26794, 2025. Thank you again for your insightful and constructive feedback!

We sincerely appreciate the time and effort you have dedicated to reviewing our manuscript. We have tried our best to provide responses to all of your questions and we hope our responses would address your concerns.

As the discussion period nears its conclusion, we would be grateful for any final feedback you may have. Please don't hesitate to let us know if further clarifications would be helpful - we remain ready to provide additional details as needed.

Thank you again for your valuable insights and constructive feedback throughout this process.

Thanks for the reply! The generation-first experiments answered my question about the training order. The explanation of why 1.7B is worse than 0.6B on HumanML3D also makes sense to me. I'll keep my score at 3.

Reviewer_UhBuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose Möbius, a unified model for motion understanding, generation, completion, and editing. The framework directly processes continuous motion frames without discrete tokenization or a pretrained motion VAE. Text, clean motion, and noisy motion share attention, with separate parameters for understanding and generation. The authors first train the model for motion understanding, then train the generation branch with the understanding parameters frozen. Experiments show improved motion generation and captioning on HumanML3D, and better instruction-guided motion editing on MotionFix.

Critical strengths
  1. The framework unifies multiple motion tasks within a single architecture that directly processes continuous motion frames.
  2. Experiments on two datasets show improved performance in motion generation, understanding, and editing.
  3. The demo video presents convincing qualitative results across multiple motion tasks.
Critical weaknesses
  1. The scaling experiments change both model size and training data, making their effects difficult to separate. This is particularly relevant given the limited scale of motion data compared with the language pretraining data used by the 1.7B backbone.
  2. The additional training corpus is described only as internal data, without details on its sources or composition.
  3. In Table 4, the VQ variant uses dense computation, while the continuous variants use role-specific parameters. Changing both representation and routing makes it difficult to isolate the benefit of raw motion representations.
  4. Kimodo is included in Table S8 for temporal completion but omitted from the main text-to-motion evaluation, despite being a strong baseline for that task.
Questions for authors
  1. How does the proposed method compare with Kimodo on public datasets such as BONES-SEED?
  2. Is there evidence that the pretrained model retains its reasoning abilities after training on motion tasks?
  3. Have the authors tested whether the model can retain context and follow earlier instructions across multiple generation–understanding turns?
3 comments
AuthorsOpenReview ↗

Response to Reviewer UhBu (Part1)

Thank you for your valuable comments. We have addressed your concerns through controlled scaling and representation ablations, dataset analysis, comparisons with Kimodo, and evaluations of language retention and multi-turn interaction.

W1: Separating the effects of model size and training data

Thank you for raising this question. Following your suggestion, to separate model size from data scale, we trained a new Möbius-1.7B baseline exclusively on HumanML3D under the exact same setup as Möbius-0.6B (Table S16). Comparing this baseline against our existing models shows that 0.6B already saturates the small HumanML3D dataset (<30 hours), so scaling to 1.7B brings clear gains only when paired with our 60× larger 1,700-hour corpus:

Table S16: Disentangling model size and training data. (Part of experiments.)

\hline \text{Model} & \text{Training data} & \text{T2M R@1}\uparrow & \text{FID}\downarrow & \text{M2T R@1}\uparrow & \text{CIDEr}\uparrow \\\\ \hline 0.6\text{B} & \text{HumanML3D} & \mathbf{0.8209} & \mathbf{6.1291} & 0.6056 & 0.5754 \\\\ 1.7\text{B} & \text{HumanML3D} & 0.8161 & 6.2357 & 0.5937 & 0.5699 \\\\ \mathbf{1.7\text{B}} & \textbf{HumanML3D + Internal} & 0.7968 & 10.6642 & \mathbf{0.6152} & \mathbf{0.5836} \\\\ \hline \end{array}$$ * **Varying model size on HumanML3D:** Trained only on HumanML3D, 1.7B performs almost identically to 0.6B in both T2M and M2T, as 0.6B already captures this small dataset (both surpass real motion in R@1 in **Table 1**). * **Varying training data at 1.7B:** Adding the 1,700-hour corpus utilizes the 1.7B capacity, improving all M2T metrics (best in **Table 2**), achieving the lowest FID across all four temporal completion tasks (**Table S15**), and winning 112–0 over 0.6B in human preference on practical animator prompts (**Table 4**). See **Section 4.1** and **Appendix D.5** (**Table S16**). ### W2: Sources and composition of the additional training corpus Thank you for pointing this out. Following your suggestion, we added a detailed analysis of our internal corpus in **Appendix A.1** (**Table S1**, **Figure S1**): it contains about 770K text-annotated clips (about 1,700 hours) of professionally captured motion retargeted to SMPL-22 at 30 FPS (201-dim), used strictly for training while all evaluations remain on public benchmarks. Whereas roughly 34% of HumanML3D is basic walking/running, our corpus covers far more dynamic and stylized actions scarce in HumanML3D—including combat/martial arts (about 27% vs. 4%), hit reactions/falls (about 6% vs. 1%), and dance (about 4% vs. 2%, **Table S1**)—and spans a much broader semantic and pose space in t-SNE projections (**Figure S1**). ### W3: Isolating raw motion representation from routing in the ablation study Thanks for the comment. Following your suggestion, we added a *Role + VQ* variant (**ID 2** in **Table 5**) to separate motion representation from routing. Holding role routing and the training schedule fixed across *Role + VQ* (**ID 2**), *Role + VAE* (**ID 3**), and *Role + Frame* (**ID 14**), our continuous frame-level representation clearly outperforms VQ on both tasks and achieves lower FID and stronger captioning than VAE (**Section 4.4**, **Table 5**; **Appendix D.8**, **Table S18**): **Table 5: Ablation study on motion representation and routing.** (Part of experiments.) $$\begin{array}{cllcc|cc} \hline \text{ID} & \text{Routing} & \text{Encoding} & \text{T2M R@1}\uparrow & \text{FID}\downarrow & \text{M2T R@1}\uparrow & \text{CIDEr}\uparrow \\\\ \hline 1 & \text{Dense} & \text{VQ} & 0.5627 & 6.0912 & 0.5695 & 0.4568 \\\\ 2 & \text{Role} & \text{VQ} & 0.6713 & 6.0121 & 0.5643 & 0.4471 \\\\ 3 & \text{Role} & \text{VAE} & \mathbf{0.8251} & 5.9970 & 0.5846 & 0.5522 \\\\ \hline \mathbf{14} & \textbf{Role} & \textbf{Frame} & 0.8207 & \mathbf{4.8772} & \mathbf{0.6001} & \mathbf{0.5629} \\\\ \hline \end{array}$$ * **Raw frames vs. VQ (ID 2 vs. ID 14):** While role routing improves VQ T2M R@1 from 0.5627 (**ID 1**) to 0.6713 (**ID 2**), replacing VQ with continuous frames (**ID 14**) under identical routing further boosts both M2T and T2M, showing the gains come largely from avoiding quantization loss. * **Raw frames vs. VAE (ID 3 vs. ID 14):** Under identical role routing, raw frames (**ID 14**) achieve lower FID (4.88 vs. 6.00) and better M2T across all metrics while preserving 1-to-1 frame alignment without a pretrained VAE.
AuthorsOpenReview ↗

Response to Reviewer UhBu (Part2)

W4: Including Kimodo in the main text-to-motion evaluation

Thank you for pointing this out. Following your suggestion, we evaluated Kimodo (Kimodo-SOMA-RP-ArdyRep-v1.2) on HumanML3D T2M via its official inference and SOMA-to-SMPL retargeting pipeline (Table 1, Appendix D.6, Figure S4): Möbius-1.7B outperforms Kimodo across all automatic metrics (R@1 0.7968 vs. 0.5161, FID 10.66 vs. 78.58) and executes multi-part prompts more completely (Figure S4). That said, Kimodo is a very strong baseline (with competitive completion results in Table S15), and we explicitly note that SOMA-to-SMPL retargeting and training-distribution differences disadvantage its HumanML3D FID, making our native BONES-SEED comparison below (Q1, Table S17) a fairer test.

Q1: Comparing with Kimodo on public datasets such as BONES-SEED

Thank you for this suggestion. For a fair, retargeting-free comparison, we trained a new Möbius-1.7B model exclusively on the public BONES-SEED training split using Kimodo's native 369-dim SOMA representation at 30 FPS and evaluated it with the official Kimodo pipeline against Kimodo-SOMA-SEED-v1.1 (Appendix D.6, Table S17). Under identical training data, skeleton, and evaluation, Möbius-1.7B outperforms Kimodo in both R@3 and FID on Overview and Timeline single across both splits, with lower foot skating on all six subsets:

Table S17: Text-following comparison with Kimodo on BONES-SEED. (Part of experiments.)

\hline \text{Split / Subset} & \text{Method} & \text{R@3}\uparrow & \text{FID}\downarrow & \text{Skate}\downarrow \\\\ \hline \text{Content / Overview} & \text{Kimodo / }\textbf{Ours (1.7B)} & 81.13\\,/\\,\mathbf{81.35} & 0.035\\,/\\,\mathbf{0.032} & 3.703\\,/\\,\mathbf{3.631} \\\\ \text{Content / Timeline single} & \text{Kimodo / }\textbf{Ours (1.7B)} & 73.50\\,/\\,\mathbf{73.69} & 0.029\\,/\\,\mathbf{0.024} & 3.535\\,/\\,\mathbf{3.320} \\\\ \hline \text{Repetition / Overview} & \text{Kimodo / }\textbf{Ours (1.7B)} & 91.05\\,/\\,\mathbf{93.61} & 0.004\\,/\\,\mathbf{0.003} & 4.238\\,/\\,\mathbf{4.027} \\\\ \text{Repetition / Timeline single} & \text{Kimodo / }\textbf{Ours (1.7B)} & 80.08\\,/\\,\mathbf{85.71} & 0.007\\,/\\,\mathbf{0.006} & 4.134\\,/\\,\mathbf{4.115} \\\\ \hline \end{array}$$ ### Q2: Retaining pretrained reasoning abilities after motion training Thank you for raising this point. We added evaluations on the full 5-shot MMLU benchmark (57 subjects, 14,042 questions; **Table S11**) and text-only biomechanical QA probes (**Table S12**) in **Appendix D.3**. Because our architecture freezes understanding parameters during Phase II generation and interleaves text replay during Phase I, our models retain nearly all backbone reasoning and give accurate biomechanical explanations: **Table S11: General language ability after motion training (5-shot MMLU accuracy, %).** $$\begin{array}{lccc} \hline \text{Model} & \text{Qwen3 backbone} & \textbf{Ours} & \textbf{Retained} \\\\ \hline 0.6\text{B} & 47.34 & \mathbf{46.97} & \mathbf{99.2\\%} \\\\ 1.7\text{B} & 60.35 & \mathbf{58.51} & \mathbf{96.9\\%} \\\\ \hline \end{array}$$ ### Q3: Retaining context and following instructions across multiple turns Thank you for the review. Maintaining context across multiple generation–understanding turns is a challenging setting, and we want to note candidly that our model was not explicitly trained on multi-turn motion dialogues. To test whether our model can still support this capability, we conducted an additional four-turn interactive experiment with Möbius-1.7B interleaving instruction grounding, motion generation, motion captioning, temporal continuation, and multi-turn recall (**Table S13**). The results show that our model can follow instructions across turns to generate, describe, and continue motions coherently while accurately recalling earlier dialogue context and requests, indicating that our unified architecture and preserved LLM backbone still retain a solid degree of multi-turn capability even without dedicated multi-turn training. Developing native multi-turn motion–language training is a very interesting and promising direction for future work, and we thank you for this inspiring suggestion. Please see **Appendix D.3** (**Table S13**) for the complete transcript and rendered motion frames. Thank you again for your thorough and constructive review!

We sincerely appreciate the time and effort you have dedicated to reviewing our manuscript. We have tried our best to provide responses to all of your questions and we hope our responses would address your concerns.

As the discussion period nears its conclusion, we would be grateful for any final feedback you may have. Please don't hesitate to let us know if further clarifications would be helpful - we remain ready to provide additional details as needed.

Thank you again for your valuable insights and constructive feedback throughout this process.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response

We thank all three reviewers (UhBu, afCa, and CLtQ) for their time and constructive feedback. We are glad to see your positive reception of our unified framework, the role-routed continuous motion architecture, and our results across understanding, generation, completion, and editing tasks.

Based on your suggestions, we have run several new experiments and updated the manuscript (all changes are marked in blue). Below is a summary of the main additions:

  • Human preference study (Reviewer CLtQ): We ran a double-blind user study with 30 participants and 360 pairwise judgments across both HumanML3D prompts and practical prompts written by animators (Section 4.2, Table 4, and Appendix B.2).
  • Multi-round confidence intervals and 263-dim evaluation (Reviewer CLtQ): We repeated the HumanML3D T2M evaluation 20 times to report 95% confidence intervals under the MS-272 evaluator (Appendix C.1, Table S9), and added 20-round results using the standard 263-dimensional HumanML3D evaluator (Appendix C.2, Table S10).
  • Separating model scale from training data (Reviewers UhBu, afCa): We trained a new Möbius-1.7B model using only HumanML3D to isolate the effect of model size from the additional training data, and verified through user studies and multi-task evaluations that scaling to 1.7B with the 1,700-hour corpus substantially improves overall generalization despite HumanML3D automatic metrics already saturating at 0.6B (Section 4.1, Appendix D.5, Table S16).
  • Details on the 1,700-hour training dataset (Reviewer UhBu): We added statistics, action-category breakdowns, and t-SNE visualizations comparing our internal training corpus with HumanML3D (Appendix A.1, Table S1, Figure S1).
  • New ablations on VQ routing and generation-first training (Reviewers UhBu, afCa): We added a Role + VQ variant (ID 2) to separate motion representation from parameter routing, as well as two generation-first, understanding-second baselines (IDs 8 and 9) under the same architecture and step budget (Section 4.4, Table 5, and Appendix D.8, Table S18).
  • Comparisons with Kimodo (Reviewer UhBu): We added Kimodo to the main HumanML3D T2M table (Table 1, Figure S4) and trained a native SOMA-based Möbius-1.7B model on the public BONES-SEED dataset for a direct comparison without retargeting (Appendix D.6, Table S17).
  • Language retention and multi-turn interaction (Reviewer UhBu): We evaluated our models on the 5-shot MMLU benchmark and text-only motion QA to verify that language reasoning is preserved (Appendix D.3, Tables S11 and S12), and added a four-turn conversation example interleaving motion generation, continuation, and captioning (Table S13).
  • Motion-specific MoT design choices (Reviewer afCa): We expanded our discussion on why adapting MoT to human motion requires different design choices compared to image and video models (Section 3 and Appendix D.7).

We address each reviewer's specific questions in the individual responses below.

AuthorsOpenReview ↗

Final Remarks

Dear Area Chairs, Senior Area Chairs, and Reviewers,

As the discussion phase concludes, we would like to briefly highlight our core contribution and summarize how the paper has become much more complete through the rebuttal.

1. Our Core Idea and Conclusions

Existing unified motion–language models either discretize motion via VQ (losing motion fidelity) or rely on a separate motion VAE / external diffusion module outside the language model.

  • Core formulation: Möbius unifies motion understanding and continuous motion generation directly on raw motion frames within a single shared Transformer space—without any pretrained motion tokenizer or VAE.
  • Key motion-specific designs: By routing clean observed frames and noisy denoising frames by computational role rather than modality, maintaining 1-to-1 frame alignment, and adopting understanding-first parameter transfer, a single model naturally supports text-to-motion generation, motion captioning, instruction-guided editing, and temporal completion without task-specific heads.
  • Core conclusion: Trained solely on public benchmarks without any internal data, our model already achieves state-of-the-art performance across understanding, generation, completion, and editing, while scaling up to a larger model and corpus further improves open-domain generalization.

2. How the Paper Improved After the Rebuttal

Guided by the reviewers' constructive suggestions, we substantially expanded our experiments and analyses, making the revised paper much more complete and self-contained (with Reviewers afCa and CLtQ both acknowledging that our responses to their initial questions made sense):

  • Systematic human evaluation and statistical rigor: We added a double-blind pairwise user study, multi-round confidence intervals, and evaluations under the traditional HumanML3D evaluator, confirming that our model is consistently preferred by human raters and that our gains are statistically stable across evaluators.
  • Strict isolation of architectural gains vs. data scaling: We added a controlled public-data scaling baseline, a detailed composition and distribution analysis of our scaling corpus, and new ablations on VQ role-routing and matched-budget generation-first training. Together, these confirm that our core gains come from our architectural and training designs on public data, while the larger corpus further strengthens open-domain generalization.
  • Cross-dataset validation and preserved LLM capabilities: We trained a native SOMA-based model on the public BONES-SEED benchmark for a retargeting-free comparison with Kimodo, and added general language benchmark and multi-turn interaction evaluations—confirming that our framework generalizes across skeletons/datasets while preserving the LLM backbone's reasoning and multi-turn capabilities.
  • Clear documentation of evaluation and visualization fairness: We clarified that no post-processing, smoothing, or hand motion is used in any quantitative evaluation or user study (all user-study videos use the exact same standard SMPL mesh with rigid flat hands, and visualization presets in the teaser/demo video are applied purely for presentation and identically across all baselines).

We sincerely thank the Area Chairs, Senior Area Chairs, and Reviewers for your time, guidance, 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.