acceptodds
Withdrawn from ICLR 2027

LUNAR: Language-Grounded Latent Reasoning for Text-to-Motion Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We propose LUNAR, a structured textual–latent reasoning framework built on a core principle: use language to scope latent reasoning, and motion to ground it. LUNAR organizes complex motion into atomic motions and represents each as a paired textual–latent reasoning unit. The textual block specifies the action intent and how it is realized across the body, defining the semantic scope of the latent block. The latent block is grounded in the corresponding motion segment through a pretrained text–motion space. After completing the reasoning trajectory, LUNAR predicts motion tokens that can be decoded into 3D motion by a frozen motion codec. We train LUNAR with progressive supervised fine-tuning followed by trajectory-level group relative policy optimization. Experiments show that LUNAR establishes state-of-the-art in-domain (ID) text–motion alignment on HumanML3D, achieving a Top-3 R-Precision of 0.838 and an MM-Dist of 2.714. In out-of-domain (OOD) evaluation across compositional, paraphrastic, and cross-dataset shifts, LUNAR consistently outperforms all compared baselines, improving overall R@3 from 0.517 to 0.593 in the T2M space and from 0.606 to 0.696 in the independent LaMP space. Our code will be publicly available upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
human motion generation, reasoning, animation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GBF1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The modeling was conducted through part-level segmentation and the analysis of atomic motions. While this approach holds value by addressing a practical problem frequently explored in the field, the lack of a thorough review of existing literature leaves the research’s positioning unclear, making the problem formulation appear unoriginal. Furthermore, the models used for comparison are outdated, and the evaluation is incomplete.

Critical strengths
  1. Addresses a practical aspect within the field of Text-to-Motion.
  • Although the field of Text-to-Motion has become quite stable and has seen a proliferation of research papers, it remains underutilized in the industry. I believe this is because, while current models can generate simple motions effectively, they struggle with motions derived from complex prompts. Research that tackles this challenge represents a valuable contribution to the community.
Critical weaknesses
  1. Lack of Prior Research and Novelty
  • Although the proposed approach utilizes a part-level methodology, it fails to provide a comparison or analysis against existing part-level methods such as ParCo (ECCV 2024), LGTM (SIGGRAPH 2024), BiPO (WACV 2026), and ParTY (CVPR 2026).
  • Furthermore, given that the problem formulation and methodology appear closely related to "events," the paper lacks a comparative analysis with existing methods in this domain, such as AToM (CVPR 2025) and Event-T2M (ICLR 2026). Consequently, the positioning of this paper remains unclear, and it appears to lack sufficient novelty.
  1. Evaluation Fails to Support Claims
  • The FID performance is significantly worse than that of MoMask, which does not utilize an LLM.
  • Multimodality performance is not reported.
  • For non-LLM-based approaches, comparisons are limited to models from 2024, which are now outdated.
  1. Absence of User Study
  • Evaluating human motion solely through metrics is often insufficient to capture the full picture; therefore, a user study is essential. However, no such evaluation is provided.
  1. Insufficient Demonstration of Model Efficiency
  • The paper fails to demonstrate efficiency; it should provide a comparison of parameters and latency against both non-LLM and LLM-based methodologies.
Questions for authors
  1. The current comparative evaluation and analysis of existing literature are insufficient. Please provide a comparison and analysis against the papers mentioned above.
  2. To my understanding, R-Precision is a more robust metric than GT (Ground Truth); please clarify whether this metric can be used to substantiate superior performance. Additionally, given the significantly lower FID score, please provide an analysis of this result and demonstrate why the model still outperforms others despite this.
  3. Please demonstrate the model's diversity by showcasing its multimodality.
  4. Please include a user study.
  5. Please demonstrate the model's efficiency by comparing its FLOPs, parameter count, and latency against those of state-of-the-art models.
Reviewer_8K7kOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a new method named LUNAR, a structured textual–latent reasoning framework for LLM-based text-to-motion generation. The core idea is to organize complex motion into atomic motions and represent each as a paired textual–latent reasoning unit. Specifically, the proposed method first decomposes the prompt into atomic motions and describes each of them with body-group text, followed by a block of latent states. These latent states are aligned with the embedding of the corresponding ground-truth motion segment in a pretrained text–motion space. The authors train the method with progressive SFT followed by GRPO. The authors conduct experiments on HumanML3D and three self-constructed OOD splits to validate the effectiveness of the proposed method.

Critical strengths
  1. The authors introduce a structured textual–latent reasoning method that organizes complex motion into atomic motions and represents each as a paired textual–latent reasoning unit.
  2. The OOD evaluation covers composition, paraphrasing and cross-dataset transfer, and the results are cross-checked with an independent evaluator.
Critical weaknesses
  1. The in-domain comparison in Table 1 is incomplete and not convincing.
  • The non-LLM group in Table 1 misses recent methods such as SALAD [1] and MotionHiFlow [2], and LaMP-T2M is cited but not compared. All of them report better R-Precision and FID than LUNAR under the same evaluator, so the state-of-the-art claim does not hold.
  • The rewards use the same encoder as the evaluator (lines 972–973). As shown in Table 3, adding GRPO improves R-Precision but worsens FID, which suggests overfitting to the evaluator.
  1. The technical contribution is limited. As discussed in the appendix (lines 1101–1104), Motion-R1 already uses decomposed CoT, cold-start SFT and GRPO with similar rewards. The main new component is the latent block, but its benefit is not well supported.
  • Compared with the textual reasoning variant, the latent block brings only a small gain, and the confidence intervals in Appendix C.4 include zero.
  • The intervention study shows that changing the latent states does not affect text–motion alignment (lines 1232–1235), so it is unclear whether the latents carry motion semantics.
  1. A large part of the OOD gain comes from GRPO rather than from the reasoning.
  • As shown in Table 3, Baseline-RL, which predicts motion directly without reasoning, already outperforms all external baselines in Table 2.
  • Baseline-RL and Baseline-RL-Reasoning should be included in Table 2.
  • The paraphrase analysis in Table 10 shows no robustness advantage over the textual reasoning variant.
  1. The paper emphasizes how motion is realized through the body, but does not report physical metrics such as foot skating, nor does it include a user study.

[1]: Seokhyeon Hong, Chaelin Kim, Serin Yoon, et al. SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing. In CVPR, 2025.

[2]: Heng Li, Xiaotong Lin, Ling-An Zeng, et al. MotionHiFlow: Text-to-Motion via Hierarchical Flow Matching. In CVPR, 2026.

Questions for authors

See Weaknesses.

Reviewer_WYu9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LUNAR, which generates motion from text through textual-latent reasoning. During inference, the model first generates a high-level language reasoning trace, then temporally ordered textual-latent units which contain a phrase, body-part level descriptions, and latents, finally the model generates motion tokens. The training procedures contain a progressive cold-start SFT and trajectory-level GRPO. This framework improves text-motion alignment in both in-distribution and out-of-distribution settings.

Critical strengths

The two-stage training strategy together with the textual-latent reasoning trace is effective enough, which gain substantial text-motion alignment improvements in both ID and OOD settings. In the OOD settings, the score improvements are consistent across two different evaluators, enhancing reliability of the results.

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. In the videos of the supplementary materials, there are a lot of twisted motions. Accoding to my experience, this is usually due to failure of inverse kinematics from HumanML3D style 263-dimensional representations. If the models are correctly trained and used for inference, these twisted motions should not appear, this brings doubt to the implementation of model comparisons.
Questions for authors
  1. Could you add a human preference evaluation to better show subjective visual quality?
  2. Could you explain why there are twisted motions in the provided video examples?

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

Related papers

Discussion (0)

Sign in to comment.