Research map / Vision-language and multimodal / Robot action and navigation
M2Tok: Multi-head Multi-codebook Discrete Action Tokenization for Vision-Language-Action Models
Chunpu Xu, Zhixuan Liang, Yuhao Zhang, Chi-Min Chan, Jiashuo Wang, Yang Xiao, Mengkang Hu, Xiaokang Yang, Yao Mu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-head decomposition (vs single head) (ablation) | M2Tok · RoboTwin | average success rate: 0.22 -> 0.51 | Improved |
| add multiple codebooks (vs single codebook) (ablation) | M2Tok · RoboTwin | average success rate: 0.34 -> 0.51 | Improved |
| add conversion module (vs direct embedding layer) (ablation) | M2Tok · RoboTwin | average success rate: 0.42 -> 0.51 | Improved |
| add multi-head to Move Pillbottle Pad (ablation) | M2Tok · RoboTwin Move Pillbottle Pad | success rate: 0.00 -> 0.33 | Improved |
| increase number of heads h up to 8 (ablation) | M2Tok · RoboTwin | success rate | Improved |
| increase number of heads h beyond 8 (ablation) | M2Tok · RoboTwin | success rate | No clear effect |
| use M2Tok instead of VQ-VLA tokenizer | VLA (Qwen2.5-0.5B) · RoboTwin | average success rate: 0.45 -> 0.51 | Improved |
| use M2Tok instead of VQ-VLA tokenizer | VLA (Qwen2.5-0.5B) · Simpler-Env | average success rate: 0.21 -> 0.28 | Improved |
| use M2Tok instead of VQ-VLA tokenizer | VLA (Qwen2.5-0.5B) · Real-world zero-shot | average success rate: 0.28 -> 0.33 | Improved |
| use M2Tok instead of VQ-VLA on Stack Green on Yellow | VLA (Qwen2.5-0.5B) · Simpler-Env | success rate: 0.21 -> 0.08 | Hurt |
| use M2Tok instead of Binning tokenizer for inference speed | VLA (Qwen2.5-0.5B) · RoboTwin latency | speed (Hz): 2.0 -> 8.2 | Improved |
| use vLLM inference engine | M2Tok VLA · RoboTwin latency | speed (Hz): 8.2 -> 56.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VQ-VLA — VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers (ICCV 2025) · better on 9 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- Grounding Multimodal Large Language Models in Actions (NeurIPS 2024 · similarity 0.88)
- RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.88)
- CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification (NeurIPS 2025 · similarity 0.88)
- Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
