Research map / Large language models / Reasoning and chain of thought

UniMo: Unified Motion Generation and Understanding with Chain of Thought

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UniMo, an LLM-based unified model for text-to-motion generation and motion-to-text understanding that uses video-grounded chain-of-thought data, SFT, and GRPO with task-specific rewards.

Problem

Existing motion generation and understanding methods are separate or weakly unified, lack interpretability, and suffer from cumulative errors of next-token prediction over motion tokens.

Approach

Builds motion-consistent CoT annotations by rendering HumanML3D motions in Blender and captioning the videos with Qwen2.5-VL-72B, then trains a unified Qwen2.5-3B model with SFT on CoT and optimizes both T2M and M2T with GRPO using motion, semantic, format, and caption similarity rewards.

Key results

• On HumanML3D T2M, UniMo reaches R-Precision Top-1 0.539 and MM-Dist 2.768, best among reported methods, with FID 0.177 (MoMask 0.045 is best on FID).
• On M2T captioning, UniMo reports BLEU@1 63.10, BLEU@4 19.74, CIDEr 46.69, BertScore 54.26, exceeding MotionLLM (54.53 BLEU@1, 33.74 CIDEr).
• Ablation: with CoT and GRPO, T2M Top-1 R-Precision rises from 0.384 (no CoT, no reward) to 0.539.

Surprising findings

• MoMask attains a lower FID (0.045) than UniMo (0.177), so UniMo is not best on FID in the T2M table.
• Adding CoT alone increases FID (0.178 to 0.292) even though it improves R-Precision.

Limitations

• FID on T2M is worse than MoMask, indicating gaps in realism measures.
• CoT annotations depend on a 72B VLM teacher, and the paper does not quantify annotation accuracy beyond qualitative inspection.
• Evaluation is limited to HumanML3D.

Experimental setup

Datasets: HumanML3D

Benchmarks: HumanML3D

Metrics: R-Precision FID MM-Dist Diversity MModality BLEU@1 BLEU@4 ROUGE-L CIDEr BertScore

Baselines: MoMask Motion-R1 MotionLLM MotionGPT-2 MotionChain MotionGPT T2M-GPT MDM

Builds on: GRPO Qwen2.5-3B-Instruct HumanML3D MoMask

Closest papers by meaning

• AvatarGPT: All-in-One Framework for Motion Understanding Planning Generation and Beyond (CVPR 2024 · similarity 0.88)
• HSI-GPT2: A Dual-Granularity Large Motion Reasoning Model with Diffusion Refinement for Human-Scene Interaction (CVPR 2026 · similarity 0.87)
• HMVLM:Human Motion-Vision-Language Model via MoE LoRA (NeurIPS 2025 · similarity 0.87)
• The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion (CVPR 2025 · similarity 0.86)
• Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.86)
• MG-MotionLLM: A Unified Framework for Motion Comprehension and Generation across Multiple Granularities (CVPR 2025 · similarity 0.86)