Vision-Language-Action (VLA) models rely on current observations, including images, language instructions, and robot states, to predict actions and complete tasks. While accurate visual perception is crucial for precise action prediction and execution, recent work has attempted to further improve performance by introducing explicit reasoning during inference. However, such approaches face significant limitations. They often depend on data-intensive resources such as Chain-of-Thought (CoT) style annotations to decompose tasks into step-by-step reasoning, and in many cases require additional visual grounding annotations (e.g., bounding boxes or masks) to highlight relevant image regions. Moreover, they involve time-consuming dataset construction, labeling, and retraining, which ultimately results in longer inference sequences and reduced efficiency. To address these challenges, we propose ATA, a novel training-free framework that introduces implicit reasoning into VLA inference through complementary attention-guided and action-guided strategies. Unlike CoT or explicit visual-grounding methods, ATA formulates reasoning implicitly by integrating attention maps with an action-based region of interest (RoI), thereby adaptively refining visual inputs without requiring extra training or annotations. ATA is a plug-and-play implicit reasoning approach for VLA models, lightweight yet effective. Extensive experiments show that it consistently improves task success and robustness while preserving, and even enhancing, inference efficiency.
Figures & tables
Fig. 1 : Overview of the proposed ATA framework. ( Left ) Given a language instruction and multimodal observations, the VLA model processes tokens through stacked layers to produce an action chunk. Our framework injects implicit reasoning cues via two complementary strategies: ( i ) Attention-guided reasoning extracts attention maps from intermediate layers to focus on task-relevant visual regions; ( ii ) Action-guided reasoning leverages the end-effector pose and camera parameters to construct a directional region of interest (RoI). ( Right ) In the original pipeline (top), an error at an early step may propagate along the horizon (up to n steps), leading to task failure (marked with × ) and risky behaviors such as knocking over a bottle (highlighted with a red circle ). The proposed pipeline (bottom) applies guidance at selected steps, resulting in a shorter effective horizon ( m<n ), successful task completion (marked with ✓ ), and more robust execution.
Fig. 2 : Illustration of the action-guided strategy. The end-effector (EEF) pose (REEFW,tEEFW) is obtained in the world coordinate system and transformed into the camera coordinate (REEFC,tEEFC) , allowing projection of the EEF direction onto the image plane. The resulting region of interest (RoI) is represented as a conic sector in the pixel coordinate system. For visualization, we overlay a semi-transparent red mask so that pixels inside the RoI appear in red, while the actual inference mask is computed from Eq. ( 9 ).
Model
Dataset
L
Freq
Time
Step
OpenVLA
Spatial
20
100
periodic
50
Goal
5
100
periodic
80
Object
20
100
periodic
60
10
20
50
periodic
170
π0 -fast
LIBERO
16/12/12/12
∞
t=0
10/7/5/2
HybridVLA
RLBench
5
∞
t=0
3
TABLE I : Best configurations of Attention-Guided and Action-Guided strategies. L denotes the attention layer index, Freq denotes the trigger frequency. For π0 -fast and HybridVLA, the attention-guided strategy is applied only once ( Freq=∞ ) at the first timestep ( t=0 ).
Models
Method
LIBERO-Spatial
LIBERO-Goal
LIBERO-Object
LIBERO-Long
Avg S.R.
Avg S.I.C ↓
Avg I.C. ↓
OpenVLA
Baseline
84.2
75.6
88.6
55.2
75.9
165
235
OpenVLA
API
86.2
78.2
88.6
52
76.3
167
232
OpenVLA
ATTN
87
82
92
57.4
79.6
165
228
OpenVLA
ATA
88.8
82.6
94.6
58.2
81.1
164
225
π0 -fast
Baseline
96.2
89
98.2
60.2
85.9
30
41
π0 -fast
API
97.6
88.8
98
60.8
86.3
31
41
TABLE II : Comparison of ATA and Baselines on LIBERO and RLBench. LIBERO includes four datasets, each with 10 tasks and 50 demonstrations per task. The RLBench environment evaluates 8 tasks, each with 100 demonstrations. Avg S.R. denotes the average success rate. Attn denotes the use of only the Attention-Guided strategy, and ATA integrates both Attention-Guided and Action-Guided strategies. Avg S.I.C. indicates the average number of inference calls required for successfully completed tasks, and Avg I.C. represents the overall average number of inference calls across all tasks.
Models
Method
LIBERO-Spatial
LIBERO-Goal
LIBERO-Object
LIBERO-Long
Avg S.R.
OpenVLA
Baseline
84.2
75.6
88.6
55.2
75.9
OpenVLA
Blur
83.4
74.6
84.2
53.2
73.9
OpenVLA
RandomBlur
84.6
77.2
87.2
53.8
75.7
OpenVLA
ATTN*
86.6
80.8
89.2
56.4
78.3
Models
Method
Close box
Close laptop lid
Toilet seat down
Sweep to dustpan
Close fridge
Umbrella out
Wine at rack
Water plants
Avg S.R.
HyBridVLA
Baseline
85
90
90
100
90
40
50
25
71.3
TABLE III : Comparison of Blur, RandomBlur, Attention-Guided, and Baselines on LIBERO and RLBench datasets. Blur represents the Gaussian Blur applied to the first frame, RandomBlur denotes the Gaussian Blur applied randomly to frames after the first frame, and ATTN* indicates the attention-guided strategy applied only on the first frame.
Models
Freq
Spatial
Goal
Object
Long
OpenVLA
0 (Baseline)
84.2
75.6
88.6
55.2
OpenVLA
∞ (ATTN*)
86.6
80.8
89.2
56.4
OpenVLA
20
83.9
78.8
87.8
55.6
OpenVLA
50
86
79.6
89.2
57.4
OpenVLA
100
87
82
92
56.4
OpenVLA
200
86.2
81.2
89.6
55.8
TABLE IV : Comparison of Different Frequencies of Attention-Guided Strategy on LIBERO Dataset Performance. 0 denotes the baseline, ∞ denotes applying Attention-Guided ( ATTN* ) only on the first timestep ( t=0 ), and 20 , 50 , 100 , 200 indicate applying Attention-Guided every 20, 50, 100, and 200 inference steps, respectively.
TABLE V : Real-world stacking results with and without ATA. Performance is measured on stacking one , two , and three -layer towers of blocks. For the three-tower case, we also tested robustness by adding irrelevant and unseen objects ( Three* ). The bottom-left shows the real-world robotic setup, while the right part illustrates task execution sequences in both simple ( top ) and complex ( bottom ) scenes.
Existing Vision-Language-Action (VLA) models predominantly rely on explicit Chain-of-Thought (CoT) reasoning to bridge perception and action. While effective, this paradigm suffers from high computational costs and error propagation in multi-step tasks. In this paper, we propose Adaptive Variable Alignment VLA (AVA-VLA), a novel Latent Reasoning VLA framework that models reasoning as a sequence of unobservable latent variables, bypassing the need for explicit text generation. However, latent trajectories are inherently susceptible to noise interference and misalignment with downstream objectives. To address this, we introduce a Reinforcement Learning-based Denoising mechanism that treats latent state generation as a sequential decision process, optimizing reasoning trajectories via task-level rewards. Furthermore, we incorporate an Early-Exit Strategy that adaptively terminates reasoning based on state confidence, enabling a dynamic trade-off between depth and efficiency. Extensive experiments on embodied decision benchmarks demonstrate that AVA-VLA achieves a 6x inference speedup over explicit CoT methods while attaining a 98.3% average success rate on LIBERO, improving both efficiency and long-horizon stability over full-reasoning baselines.
Current research on Vision-Language-Action (VLA) models predominantly focuses on enhancing generalization through reasoning techniques. While effective, these improvements increase computational complexity and inference latency. Furthermore, these mechanisms are typically applied indiscriminately, wasting resources on trivial tasks while failing to provide the uncertainty estimation necessary to prevent catastrophic failure on out-of-distribution scenarios. Inspired by human cognition, we propose an adaptive framework that dynamically routes VLA execution based on the complexity of the perceived state. Our approach transforms the VLA's vision-language backbone into an active detection tool by projecting latent embeddings into a set of parametric and non-parametric estimators. This allows the system to execute known tasks immediately (Act), reason about ambiguous scenarios (Think), and preemptively halt execution when encountering physical or semantic anomalies (Abstain). We find that a Gaussian Mixture Model fitted to fused vision-language embeddings provides the most reliable task-complexity signal, combining visual novelty with instruction context and cross-modal compatibility. Evaluated on the LIBERO and LIBERO-PRO benchmarks as well as on a real robot, our fused configuration achieves up to 87.5% F1-score across two VLA backbones (SmolVLA and π0), retains 83% with as little as 5% of training data, and surpasses state-of-the-art failure detectors.
Riccardo Andrea Izzo, Gianluca Bardaro, Matteo Matteucci
All authors are with the Department of Electronics, Information, and Bioengineering, Politecnico di Milano, Milan, Italy.
Vision-Language-Action (VLA) models aim for general robot learning by aligning action as a modality within powerful Vision-Language Models (VLMs). Existing VLAs rely on end-to-end supervision to implicitly enable the action decoding process to learn task-relevant features. However, without explicit guidance, these models often overfit to spurious correlations, such as visual shortcuts or environmental noise, limiting their generalization. In this paper, we introduce GuidedVLA, a framework designed to manually guide the action generation to focus on task-relevant factors. Our core insight is to treat the action decoder not as a monolithic learner, but as an assembly of functional components. Individual attention heads are supervised by manually defined auxiliary signals to capture distinct factors. As an initial study, we instantiate this paradigm with three specialized heads: object grounding, spatial geometry, and temporal skill logic. Across simulation and real-robot experiments, GuidedVLA improves success rates in both in-domain and out-of-domain settings compared to strong VLA baselines. Finally, we show that the quality of these specialized factors correlates positively with task performance and that our mechanism yields decoupled, high-quality features. Our results suggest that explicitly guiding action-decoder learning is a promising direction for building more robust and general VLA models.
Xiaosong Jia, Bowen Yang, Zuhao Ge +17
Institute of Trustworthy Embodied AI (TEAI), Fudan University · 2Shanghai Key Laboratory of Multimodal Embodied AI · Shanghai Jiao Tong University +1