Research map / Vision-language and multimodal / Robot action and navigation
ACoT-VLA: Action Chain-of-Thought for Vision-Language-Action Models
Linqing Zhong, Yi Liu, Yifei Wei, Ziyu Xiong, Si Liu, Guanghui Ren
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 Explicit Action Reasoner (EAR) (ablation) | pi0.5 · LIBERO | avg success rate: 96.9 -> 98.3 | Improved |
| add Implicit Action Reasoner (IAR) (ablation) | pi0.5 · LIBERO | avg success rate: 96.9 -> 98.1 | Improved |
| add both EAR and IAR (ablation) | pi0.5 · LIBERO | avg success rate: 96.9 -> 98.5 | Improved |
| add EAR (ablation) | pi0.5 (frozen LLM) · LIBERO-Plus (SFT) | avg success rate: 75.7 -> 83.7 | Improved |
| add IAR (ablation) | pi0.5 (frozen LLM) · LIBERO-Plus (SFT) | avg success rate: 75.7 -> 80.4 | Improved |
| add EAR and IAR (ablation) | pi0.5 (frozen LLM) · LIBERO-Plus (SFT) | avg success rate: 75.7 -> 84.1 | Improved |
| use Downsample KV-cache interaction in IAR instead of Query (ablation) | IAR · LIBERO | avg success rate: 97.0 -> 98.1 | Improved |
| use Downsample instead of Attention Pooling in IAR (ablation) | IAR · LIBERO | avg success rate: 97.3 -> 98.1 | Improved |
| double action head size to 600M (10 denoise steps) without EAR (ablation) | pi0.5 (frozen LLM) · LIBERO / LIBERO-Plus | avg success rate: LIBERO 96.6 -> 97.6; LIBERO-Plus 75.7 -> 74.9 | Mixed |
| increase denoising steps from 10 to 20 at 600M head (ablation) | pi0.5 (frozen LLM) · LIBERO-Plus | avg success rate: 74.9 -> 75.1 | No clear effect |
| add 300M EAR to matched-size baseline (600M head, 10 steps) (ablation) | pi0.5 (frozen LLM) · LIBERO-Plus | avg success rate: 74.9 -> 83.7 | Improved |
| scale EAR from 300M to 500M (ablation) | ACoT-VLA · LIBERO-Plus (SFT) | avg success rate: 83.7 -> 80.9 | Hurt |
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.
- VLA-Adapter — VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model (AAAI 2026) · better on 2 of 2 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
- CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · similarity 0.91)
- Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.89)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.89)
- Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.89)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
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.
