Vision-Language-Action (VLA) models map visual observations and language instructions directly to robotic actions. While effective for simple tasks, standard VLA models often struggle with complex, multi-step tasks requiring logical planning, as well as precise manipulations demanding fine-grained spatial perception. Recent efforts have incorporated Chain-of-Thought (CoT) reasoning to endow VLA models with a ``thinking before acting'' capability. However, current CoT-based VLA models face two critical limitations: 1) an inability to simultaneously capture low-level visual details and high-level logical planning due to their reliance on isolated, single-modal CoT; 2) high inference latency with compounding errors caused by step-by-step autoregressive decoding. To address these limitations, we propose DualCoT-VLA, a visual-linguistic CoT method for VLA models with a parallel reasoning mechanism. To achieve comprehensive multi-modal reasoning, our method integrates a visual CoT for low-level spatial understanding and a linguistic CoT for high-level task planning. Furthermore, to overcome the latency bottleneck, we introduce a parallel CoT mechanism that incorporates two sets of learnable query tokens, shifting autoregressive reasoning to single-step forward reasoning. Extensive experiments demonstrate that our DualCoT-VLA achieves state-of-the-art performance on the LIBERO and RoboCasa GR1 benchmarks, as well as in real-world platforms.
Figures & tables
Figure 1: Overview of our DualCoT-VLA. The VLM backbone processes a unified sequence comprising visual observations, language instructions, and two sets of learnable query tokens. our DualCoT-VLA employs a visual-linguistic CoT for VLA models: (1) Visual CoT: visual query hidden states are aligned with dense spatial features from a frozen Depth Anything 3 ( Lin et al., 2025 ) ; (2) Linguistic CoT: linguistic query hidden states act as conditioning prefixes for a frozen auxiliary LLM to predict explicit CoT text. Finally, the reasoning-enriched hidden states guide a Flow-Matching DiT (Action Expert) to predict continuous action. During inference, the auxiliary modules are discarded for highly efficient, single-forward-pass action generation.
Figure 2: Detailed illustrations of the dual-stream implicit CoT mechanism. (a) Visual CoT via geometric distillation from a frozen Depth Anything 3 encoder. (b) Linguistic CoT via step-level supervision from a frozen auxiliary LLM.
Figure 3: Qualitative result of our Visual-Linguistic CoT on LIBERO benchmark ( Liu et al., 2023a ) . During a robotic manipulation task, given the Language Instruction: “placing the bowl on the plate” and Input Image , we visualize the Ouput Visual CoT and Output Linguisitic CoT . To visualize the implicit visual reasoning, we employ a lightweight visual probe to project the compressed hidden states of the visual query tokens back into depth maps. This visualization demonstrates that the continuous latent tokens successfully capture dense 3D geometric structures and spatial priors. The explicitly decoded Linguistic CoT text, demonstrating the model’s capacity for scene comprehension, dynamic state tracking, and step-level logical planning.
Method
Spatial
Object
Goal
Long
Average
Diffusion Policy ( Chi et al., 2025 )
78.5
87.5
73.5
64.8
76.1
OpenVLA ( Kim et al., 2024 )
84.7
88.4
79.2
53.7
76.5
PD-VLA ( Song et al., 2025a )
95.5
96.7
94.9
91.7
94.7
π0 ( Black et al., 2025a )
98.0
96.8
94.4
88.4
94.4
π0 -Fast ( Pertsch et al., 2025 )
96.4
96.8
88.6
60.2
85.5
π0.5 ( Black et al., 2025a )
98.8
98.2
98.0
92.4
96.9
Table 1: Evaluation results on the LIBERO benchmark ( Liu et al., 2023a ) . We train our DualCoT-VLA jointly on four suites. We report the success rate (%) evaluated over 500 episodes for each task. Bold indicates the best result, and underline indicates the second best.
Task
GR00T-N1.5
GR00T-N1.6
Qwen3GR00T
Qwen3PI
Qwen3OFT
Qwen3FAST
DualCoT-VLA (Our)
BottleToCabinetClose
54.0
51.5
46.0
26.0
30.0
38.0
66.0
CanToDrawerClose
50.0
13.0
80.0
62.0
76.0
44.0
64.0
CupToDrawerClose
38.0
8.5
54.0
42.0
44.0
56.0
46.0
MilkToMicrowaveClose
60.0
14.0
48.0
50.0
44.0
44.0
58.0
PotatoToMicrowaveClose
32.0
41.5
28.0
42.0
32.0
14.0
30.0
WineToCabinetClose
38.0
16.5
46.0
32.0
36.0
14.0
38.0
Table 2: Evaluation Results on RoboCasa GR1 Tabletop Tasks ( Bjorck et al., 2025 ) . We train one model jointly on all 24 tasks, and report mean results over 50 rollouts per task. Bold indicates the best result, and underline indicates the second best.
Figure 4: Real-world Experiments. (a) Our robotic arm platform setup. (b) Success rates (SR%) of our proposed DualCoT-VLA compared to baseline methods (OpenVLA-OFT ( Kim et al., 2025 ) and GR00T-N1.6 ( Bjorck et al., 2025 ) ) across three real-world manipulation tasks of varying complexity: placing bread (easy), placing blocks (medium), and placing fruits (hard). (c) Qualitative execution sequences using our DualCoT-VLA across the three manipulation tasks with increasing levels of difficulty.
Metric
Non-CoT
AR CoT
Parallel CoT (our)
VLM Forward
53.7
3156.0
58.1
Action Head
22.5
27.5
25.1
Total Time
76.2
3178.5
83.2
Table 3: Detailed inference latency breakdown. The execution time of the VLM forward pass, the action head, and the total inference latency in milliseconds (ms) are reported.
CoT
LIBERO Benchmark
Visual
Linguistic
Spatial
Object
Goal
Long
Avg.
×
×
97.8
98.8
97.4
92.0
96.5
✓
×
99.4
99.6
97.4
95.0
97.9
×
✓
98.4
98.4
96.6
96.0
97.4
✓
✓
99.4
99.8
97.8
98.2
98.8
Table 4: Ablation study of the impact of the Visual CoT and Linguistic CoT across four task suites on the LIBERO benchmark ( Liu et al., 2023a ) . Success rates (%) are reported.
Vision-Language-Action (VLA) models have become the dominant recipe for generalist manipulation, yet they are almost universally trained by behavior cloning: a policy imitates expert action chunks conditioned on a static image and a fixed instruction. A natural remedy is to inject explicit reasoning through textual chain-of-thought (CoT). We show, both empirically and analytically, that free-form textual CoT degrades low-level control: the reasoning it produces is ungrounded, its latency breaks closed-loop timing, and, crucially, the reasoning and action tokens are optimized against conflicting objectives so that the policy learns to narrate rather than to act. We argue that what a VLA needs is not the ability to generate language, but the ability to consume grounded language. To this end we introduce \ourmethod{}, a framework that endows a VLA with language competence through (i) in-context post-training, in which perceptual evidence is injected as structured context and the model is supervised only on actions, and (ii) an agentic tool-use interface, in which the policy queries open-vocabulary detectors, monocular depth, and a vision--language model to actively acquire task-relevant information. Rather than emitting a single templated caption, our data engine produces diverse, paraphrased, and evidence-conditioned spatial descriptions, so that the policy learns to interpret language it has never seen verbatim. Across the RoboCasa-GR1, SimplerEnv, and LIBERO simulation benchmarks, together with 8 real-world robot manipulation tasks, our method consistently achieves SOTA results in both performance and efficiency when compared with CoT-based approaches under matched configurations.
Jiarui Yang, Wen Huang, Jiale Zhang +2
Nankai University · Tsinghua University · Swiss Federal Institute of Technology in Lausanne (EPFL)
Vision-Language-Action (VLA) models have shown promising capabilities for embodied intelligence, but most existing approaches rely on text-based chain-of-thought reasoning where visual inputs are treated as static context. This limits the ability of the model to actively revisit the environment and resolve ambiguities during long-horizon tasks. We propose VLA-Thinker, a thinking-with-image reasoning framework that models perception as a dynamically invocable reasoning action. To train such a system, we introduce a two-stage training pipeline consisting of (1) an SFT cold-start phase with curated visual Chain-of-Thought data to activate structured reasoning and tool-use behaviors, and (2) GRPO-based reinforcement learning to align complete reasoning-action trajectories with task-level success. Extensive experiments on LIBERO and RoboTwin 2.0 benchmarks demonstrate that VLA-Thinker significantly improves manipulation performance, achieving 97.5% success rate on LIBERO and strong gains across long-horizon robotic tasks. Project and Codes: https://cywang735.github.io/VLA-Thinker/ .
Chaoyang Wang, Wenrui Bao, Sicheng Gao +5
University of Central Florida · University of Würzburg · University of Southern California +1
Most Vision-Language-Action (VLA) models map observations directly to actions without explicit reasoning, limiting their capacity for reasoning-intensive long-horizon tasks. To address this, existing approaches adopt Chain-of-Thought (CoT) reasoning to enable subgoal decomposition and spatial anticipation. However, those methods lack a unified architecture for effective cross-modal reasoning and fail to explicitly include inverse reasoning ability based on the target state. We argue that manipulation planning naturally decomposes into prediction, anticipating the next visual state, and inverse dynamics, inferring the actions to reach it. Bridging both requires a unified autoregressive architecture that interleaves textual and visual reasoning in a single generation process. We propose ThinkingVLA, a generative model that realizes this decomposition within a unified Mixture-of-Transformers architecture. ThinkingVLA consists of a forward CoT that identifies the immediate subgoal and guides the visual forecasting; the predicted image then serves as the target state, grounding an inverse CoT that reasons about spatial relationships and action intent based on the predicted image; and the final action is generated conditioned on this full reasoning context. Extensive experiments on simulation and real-world benchmarks demonstrate that ThinkingVLA consistently outperforms state-of-the-art baselines, with particularly large gains on long-horizon manipulation tasks.
Tianyi Lu, Hui Zhang, Zijie Diao +8
Fudan University · Shanghai Innovation Institue · China Unicom +1