DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action Models

TL;DR

This paper investigates when and why Chain-of-Thought reasoning improves vision-language-action models, identifying key conditions for effectiveness and proposing a new model, DeepThinkVLA, that surpasses existing baselines.

Contribution

It systematically diagnoses CoT's effectiveness in VLA models, introduces conditions for success, and develops DeepThinkVLA with novel attention mechanisms and training pipelines.

Findings

DeepThinkVLA achieves 97.0% success on LIBERO.

It outperforms baselines by up to 21.7 points.

The model demonstrates practical effectiveness in real-world robot experiments.

Abstract

Does Chain-of-Thought (CoT) reasoning genuinely improve Vision-Language-Action (VLA) models, or does it merely add overhead? Existing CoT-VLA systems report limited and inconsistent gains, yet no prior work has rigorously diagnosed when and why CoT helps robots act. Through systematic experiments, we identify two necessary conditions that must be jointly satisfied for CoT to be effective in VLA: (1) Decoding Alignment -- CoT and actions must be generated with modality-appropriate mechanisms; forcing both through a single autoregressive decoder is not merely suboptimal but actively harmful, degrading performance by 4.2 percentage points; (2) Causal Alignment -- CoT must be causally linked to task success via outcome-based optimization; without it, supervised CoT is indistinguishable from no reasoning at all under distribution shift, exhibiting a 32.0\,pp performance drop nearly identical to the 31.6\,pp drop of a reasoning-free baseline. Guided by these findings, we build DeepThinkVLA: a hybrid-attention decoder satisfies Condition~1 by pairing causal attention for language with bidirectional attention for parallel action decoding, while a two-stage SFT-then-RL pipeline satisfies Condition~2 by aligning the full reasoning--action chain with sparse task-success rewards. DeepThinkVLA achieves 97.0\% success on LIBERO, 79.0\% robustness on LIBERO-Plus (vs.\ 61.6\% for $\pi_0$-FAST), and 59.3\% success on RoboTwin~2.0, exceeding the strongest baseline by 21.7 points. Furthermore, we validate the practical effectiveness of our approach through real-world robot experiments. Code available at https://github.com/OpenBMB/DeepThinkVLA