Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models
Shuanghao Bai*, Jing Lyu*, Wanqi Zhou, Zhe Li, Dakai Wang, Lei Xing, Xiaoguang Zhao, Pengwei Wang, Zhongyuan Wang, Cheng Chi, Badong Chen, Shanghang Zhang
LaRA-VLA performs iterative latent reasoning by feeding hidden states back into reasoning slots before action prediction, rather than relying on long explicit chain-of-thought generation.
- 🎉 LaRA-VLA has been accepted to ICML 2026.
- ✅ Training code is released.
- ✅ Evaluation code is released.
- ✅ Pretrained model weights are released.
- ✅ Training datasets are released.
git clone https://github.com/LoveJu1y/LaRA-VLA
cd LaRA-VLA
conda create -n lara-vla python=3.10 -y
conda activate lara-vla
pip install -r requirements.txt
pip install -e .python -c "from laravla.training.train import main; print('OK')"Before launching training, set the dataset roots and model cache path:
export BRIDGE_LEROBOT_ROOT=/path/to/bridge_datasets_parent
export LIBERO_LEROBOT_ROOT=/path/to/libero_lerobot
export HF_HOME=/path/to/qwen_cacheDataset repos:
- Bridge: https://huggingface.co/datasets/lovejuly/bridge_orig_lerobot
- LIBERO: https://huggingface.co/datasets/lovejuly/libero_lerobot_all
Model repos:
- Bridge: https://huggingface.co/lovejuly/LaRA-VLA-bridge
- LIBERO: https://huggingface.co/lovejuly/LaRA-VLA-libero
Bridge training expects:
${BRIDGE_LEROBOT_ROOT}/bridge_orig_lerobot/
annotations/
meta/
data/
videos/
The current public Bridge dataset release contains the core annotations and metadata, but does not include the raw videos/ directory. Bridge training will not run unless videos/ is available locally under the structure above.
LIBERO training expects:
${LIBERO_LEROBOT_ROOT}/
libero_goal_no_noops_1.0.0_lerobot/
libero_object_no_noops_1.0.0_lerobot/
libero_spatial_no_noops_1.0.0_lerobot/
libero_10_no_noops_1.0.0_lerobot/
Bridge:
bash scripts/run_bridge_multistage.shLIBERO:
bash scripts/run_libero_multistage.shBridge:
bash scripts/run_laravla_bridge.shLIBERO:
bash scripts/run_laravla_libero.shThe LIBERO results above correspond to the evaluation workflow documented in examples/LIBERO/README.md.
| CoT Type | Method | Spatial | Goal | Object | Long | Avg |
|---|---|---|---|---|---|---|
| No CoT | OpenVLA (Kim et al., 2025b) | 84.7 | 88.4 | 79.2 | 53.7 | 76.5 |
| π₀ (Black et al., 2024) | 96.8 | 98.8 | 95.8 | 85.2 | 94.2 | |
| OpenVLA-OFT (Kim et al., 2025a) | 97.6 | 98.4 | 97.9 | 94.5 | 97.1 | |
| Textual CoT | ThinkAct (Huang et al., 2025) | 88.3 | 91.4 | 87.1 | 70.9 | 84.4 |
| MolmoAct (Lee et al., 2025) | 87.0 | 95.4 | 87.6 | 77.2 | 86.6 | |
| π₀.₅ (Intelligence et al., 2025) | 98.8 | 98.2 | 98.0 | 92.4 | 96.8 | |
| DeepThinkVLA (Yin et al., 2025) | 99.0 | 96.6 | 96.4 | 96.2 | 97.0 | |
| Visual CoT | CoT-VLA (Zhao et al., 2025) | 87.5 | 91.6 | 87.6 | 69.0 | 81.1 |
| DreamVLA (Zhang et al., 2025b) | 97.5 | 94.0 | 89.5 | 89.5 | 92.6 | |
| F1 (Lv et al., 2025) | 98.2 | 97.8 | 95.4 | 91.3 | 95.7 | |
| UD-VLA (Chen et al., 2025b) | 94.1 | 95.7 | 91.2 | 89.6 | 92.7 | |
| Latent CoT | Fast-ThinkAct (Huang et al., 2026) | 92.0 | 97.2 | 90.2 | 79.4 | 89.7 |
| LaRA-VLA (Ours) | 96.4 | 98.6 | 99.8 | 96.6 | 97.9 |
The Bridge real-world results above are evaluated through the SimplerEnv-based pipeline documented in examples/SimplerEnv/README.md.
| CoT Type | Method | Put Spoon | Put Carrot | Stack Block | Put Eggplant | Avg |
|---|---|---|---|---|---|---|
| No CoT | OpenVLA (Kim et al., 2025b) | 0.0 | 0.0 | 0.0 | 4.1 | 1.0 |
| Octo (Ghosh et al., 2024) | 47.2 | 9.7 | 4.2 | 56.9 | 29.5 | |
| OpenVLA-OFT (Kim et al., 2025a) | 12.5 | 4.2 | 8.3 | 37.5 | 39.6 | |
| π₀ (Black et al., 2024) | 29.1 | 0.0 | 16.7 | 62.5 | 40.1 | |
| CogACT (Li et al., 2024) | 71.7 | 50.8 | 15.0 | 67.5 | 51.3 | |
| Textual CoT | ThinkAct (Huang et al., 2025) | 58.3 | 37.5 | 8.7 | 70.8 | 43.8 |
| Visual CoT | F1 (Lv et al., 2025) | 50.0 | 70.8 | 50.0 | 66.7 | 59.4 |
| UD-VLA (Chen et al., 2025b) | 58.3 | 62.5 | 54.1 | 75.0 | 62.5 | |
| Latent CoT | LaRA-VLA (Ours) | 95.8 | 62.5 | 25.0 | 91.7 | 68.8 |
Our code builds on the open-source StarVLA codebase, and incorporates ideas and components from Coconut and ECOT (Embodied Chain-of-Thought).
@article{bai2026latentreasoningvla,
title={Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models},
author={Bai, Shuanghao and Lyu, Jing and Zhou, Wanqi and Li, Zhe and Wang, Dakai and Xing, Lei and Zhao, Xiaoguang and Wang, Pengwei and Wang, Zhongyuan and Chi, Cheng and Chen, Badong and Zhang, Shanghang},
journal={arXiv preprint arXiv:2602.01166},
year={2026}
}Released under the MIT License. See LICENSE.