Paper deep dive
VLANeXt: Recipes for Building Strong VLA Models
Xiao-Ming Wu, Bin Fan, Kang Liao, Jian-Jian Jiang, Runze Yang, Yihang Luo, Zhonghua Wu, Wei-Shi Zheng, Chen Change Loy
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/20/2026, 10:45:43 PM
Summary
The paper introduces VLANeXt, a Vision-Language-Action (VLA) model derived from a systematic dissection of the VLA design space. By evaluating over 500 experiments across foundational components, perception essentials, and action modeling, the authors distill 12 key findings. VLANeXt utilizes a soft VLM-policy connection, multi-view perception with proprioception conditioning in the VLM, and flow-matching for action generation. It achieves state-of-the-art performance on LIBERO and LIBERO-plus benchmarks with a smaller parameter count (2.5B) compared to prior methods like OpenVLA-OFT (7B).
Entities (9)
Relation Signals (7)
VLANeXt â evaluatedon â LIBERO-Plus
confidence 95% · We evaluate all variants on two commonly used VLA benchmarks, including ... LIBERO-plus
VLANeXt â evaluatedon â LIBERO
confidence 95% · We evaluate all variants on two commonly used VLA benchmarks, including LIBERO
VLANeXt â outperforms â OpenVLA-OFT
confidence 95% · Despite its smaller model size, VLANeXt achieves higher success rates than prior methods on both standard task performance (LIBERO) and robustness/generalization (LIBERO-plus)
VLANeXt â conditions â Proprioception
confidence 90% · conditioning proprioception in the VLM yields the best performance
VLANeXt â uses â Flow Matching
confidence 90% · We therefore adopt the flow-matching objective.
VLANeXt â uses â Soft Connection
confidence 90% · We adopt the soft connection in subsequent models.
Qwen3-VL-2B â backbonefor â VLANeXt
confidence 85% · We use Qwen3-VL-2B in subsequent experiments as a strong yet efficient choice.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Following the rise of large foundation models, Vision-Language-Action models (VLAs) emerged, leveraging strong visual and language understanding from Vision-Language Models for general-purpose policy learning. Yet, the current VLA landscape remains fragmented and exploratory. Although many groups have proposed their own VLA models, inconsistencies in training protocols and evaluation settings make it difficult to identify which design choices truly matter. To bring structure to this evolving space, we reexamine the VLA design space under a unified framework and evaluation setup. Starting from a simple VLA baseline similar to RT-2, which is the origin of VLA, we systematically dissect design choices along three dimensions: foundational components, perception essentials, and action modelling perspectives. From this study, we distill 12 key findings that together form a practical recipe for building strong VLA models. The outcome of this exploration is a simple yet effective model, VLANeXt. It outperforms the state-of-the-art methods on the LIBERO and LIBERO-plus benchmarks and demonstrates strong performance in real-world experiments. We release a unified and easy-to-use codebase to reproduce our findings, explore the design space, and develop new VLA variants on top of a shared foundation. The codebase is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2602.18532v2
- Canonical: https://arxiv.org/abs/2602.18532v2
Trouble viewing inline? Open PDF directly â
Full Text
72,413 characters extracted from source content.
Expand or collapse full text
VLANeXt: Recipes for Building Strong VLA Models Xiao-Ming Wu 1 Bin Fan 2 Kang Liao 1 Jian-jian Jiang 2 Runze Yang 3 Yihang Luo 1 Zhonghua Wu 4 Wei-Shi Zheng 2 Chen Change Loy 1 5 https://dravenalg.github.io/VLANeXt/ Abstract Following the rise of large foundation models, Vi- sionâLanguageâAction models (VLAs) emerged, leveraging strong visual and language understand- ing from Vision-Language Models for general- purpose policy learning. Yet, the current VLA landscape remains fragmented and exploratory. Although many groups have proposed their own VLA models, inconsistencies in training protocols and evaluation settings make it difficult to identify which design choices truly matter. To bring struc- ture to this evolving space, we reexamine the VLA design space under a unified framework and evalu- ation setup. Starting from a simple VLA baseline similar to RT-2, which is the origin of VLA, we systematically dissect design choices along three dimensions: foundational components, percep- tion essentials, and action modelling perspectives. From this study, we distill 12 key findings that to- gether form a practical recipe for building strong VLA models. The outcome of this exploration is a simple yet effective model, VLANeXt. It outper- forms the state-of-the-art methods on the LIBERO and LIBERO-plus benchmarks and demonstrates strong performance in real-world experiments. We release a unified and easy-to-use codebase to reproduce our findings, explore the design space, and develop new VLA variants on top of a shared foundation. The codebase is available athttps: //github.com/DravenALG/VLANeXt. 1. Introduction Recent advances in foundation models have reshaped how we think about general-purpose robot control. Instead of 1 S-Lab, Nanyang Technological University 2 Sun Yat-sen University 3 Shanghai Jiao Tong University 4 SenseTime Re- search 5 ACE Robotics. Correspondence to: Chen Change Loy <ccloy@ntu.edu.sg>. Proceedings of the43 rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026. Copyright 2026 by the author(s). LIBERO-plus Acc. 85 80 75 70 65 60 55 LIBEROLIBERO-plus 50 LIBERO Acc. 100 95 90 85 80 75 70 65 í 0 í 0 OpenVLA-OFT VLANeXt VLANeXt OpenVLA-OFT Diameter 2B4B 7B Figure 1.Performance comparison on the LIBERO and LIBERO-plus benchmarks. We compare VLANeXt with repre- sentative VLA baselines across model scales. Despite its smaller model size, VLANeXt achieves higher success rates than prior methods on both standard task performance (LIBERO) and robust- ness/generalization (LIBERO-plus), demonstrating the effective- ness of the design recipe distilled in this work. training task-specific policies, a growing line of work builds VisionâLanguageâAction (VLA) models that leverage large Vision-Language Models to map visual observations and language instructions directly to robot actions. By inheriting rich visual understanding and language grounding from foundation models, VLAs 1 offer a scalable route toward general-purpose, language-conditioned robot policies (Ma et al., 2024; Ravichandar et al., 2020; Xiao et al., 2025c). Since the emergence of VLAs (Zitkovich et al., 2023), both academia and industry have proposed a wide range of models demonstrating strong performance and encouraging generalization across diverse tasks (Zitkovich et al., 2023; OâNeill et al., 2024; Li et al., 2023b; 2024; Kim et al., 2024a; Black et al., 2024; Team et al., 2025; Hung et al., 2025; Kim et al., 2025; Shukor et al., 2025; Intelligence et al., 2025b;a; Liu et al., 2026). Most VLA approaches build on a sim- ilar paradigm: they build on pre-trained LLMs or VLMs, processing visual observations together with language in- structions to derive action-relevant representations for policy learning. This pipeline introduces numerous design choices, including how to interface the VLM with the policy module, how to train the policy, how to select essential perceptual 1 An overview of VLAs is provided in the Appendix. 1 arXiv:2602.18532v2 [cs.CV] 20 May 2026 VLANeXt: Recipes for Building Strong VLA Models 090100 Baseline OpenVLA-OFT (7B) VLANeXt (2.5B) Foundation Component Perception Essentials Action Modeling Perspective 11.4 action chunking flow matching loss stronger VLM soft VLM-policy connection temporal history 30.3 75.4 80 90 add policy head & class token 92 LIBERO LIBERO-plus (ge ner alization ability) 56.2 50 multi-view80.5 proprioception to VLM 87.7 world modelling time series forecasting perspective 69.7 93.1 90.3 (not use due to 3x training time) 64.4 metaquery (more tokens, larger policy) transformer projector 88.8 (not use for simplicity) Figure 2. Ablation trajectory across the VLA design space (spatial suite). We progressively evolve a baseline VLA through changes in foundational components, perception, and action mod- eling. Results are reported on LIBERO initially, and on LIBERO- plus once LIBERO performance saturates, providing a more sensi- tive test of robustness and generalization. The trajectory culminates in the final VLANeXt model (2.5B) vs. OpenVLA-OFT (7B). inputs, and how actions should be represented and modeled. Despite rapid progress, early exploration of VLAs remains something of a âprimordial soupâ, rich in ideas but lacking clear structure. While prior work has explored VLA design from certain perspectives (Zhen et al., 2024; Qu et al., 2025; Zhang et al., 2025c; Cen et al., 2025b; Zhang et al., 2025b;d; Lu et al., 2025), differences in training protocols and evalua- tion setups make it difficult to identify which design choices in the shared VLA design space truly matter. This work aims to provide a more systematic understanding of this fragmented design space by comprehensively reex- amining VLA design spaces under a unified framework and evaluation protocol. While several prior works (Kim et al., 2025; Liu et al., 2026) have made preliminary attempts to explore VLA designs, their investigations remain limited in scope. This study aims to provide a more comprehensive and in-depth analysis of this domain. In detail, we begin with a simple baseline VLA, similar to RT-2 (Zitkovich et al., 2023), which is the origin of VLA and serves as a strong reference point for analyzing the effectiveness of different design choices. We evaluate all variants on two commonly used VLA benchmarks, including LIBERO (Liu et al., 2023a) and LIBERO-plus (Fei et al., 2025b), where LIBERO-plus extends the original benchmark with con- trolled and unseen perturbations to better assess robustness and generalization. Within this setup, we systematically explore the design space along three dimensions: 1) founda- tional components, covering core VLM-policy architectures and action learning objectives; 2) perception essentials: ex- amining the role of visual, language, and proprioceptive inputs; and 3) action modeling perspective: investigating designs and auxiliary objectives that facilitate action gener- ation. We conduct more than 500 distinct experiments over the above three dimensions, and distill 12 key findings that together form a practical recipe for building strong VLA models, summarized in Fig. 2. We highlight several findings that we believe are novel and noteworthy for the field: 1) a soft connection between the VLM and the policy module performs slightly better than both loose and tight coupling strategies; 2) video inputs, even though the VLM is already pretrained on video under- standing, still fails to distill useful information for action learning; 3) conditioning proprioceptive input in the VLM yields better performance than either omitting propriocep- tion or injecting it directly into the policy module; and 4) framing action generation as a time-series forecasting prob- lem and incorporating frequency-domain modeling provides an effective and efficient way to improve action prediction. The outcome of this study is a simple yet effective VLA model, VLANeXt, derived directly from the design princi- ples uncovered in our systematic exploration. Rather than relying on aggressive model scaling or task-specific engi- neering, VLANeXt achieves state-of-the-art performance on both LIBERO (Liu et al., 2023a) and LIBERO-plus (Fei et al., 2025b) (Fig. 1), and adapts effectively to real-world manipulation tasks. These results show that strong VLA per- formance can emerge from principled design choices within a unified framework. To support further progress in this direction, we release a unified and easy-to-use codebase that standardizes training and evaluation while exposing the key components of the VLA design space. The framework is intentionally lightweight and minimally encapsulated, enabling researchers to reproduce our findings, probe alter- native design choices, and build new VLA variants on top of a shared, transparent foundation, which is available at https://github.com/DravenALG/VLANeXt. 2. Recipes for Building Strong VLA Models In this section, we detail the step-by-step evolution from a simple baseline to the final VLANeXt model. We organize our exploration along three aspects: foundational compo- 2 VLANeXt: Recipes for Building Strong VLA Models ... text, image class token action MLLM head text, image (b) separate policy head(c) MetaQuery MLLM action (a) reuse text tokens action query action MLLM policy text, image meta query text tokens Figure 3. Design choices for the policy module. nents (Sec. 2.1), perception essentials (Sec. 2.2), and action modeling perspectives (Sec. 2.3). An overview is shown in Fig. 2, with full results in Table 1. Evaluation Setup. We perform the roadmap exploration on LIBERO and LIBERO-plus (Liu et al., 2023a; Fei et al., 2025b). Main experiments are conducted on the spatial suite as our primary testbed, while the resulting insights generalize across the other suites (Object, Goal, and Long), which can also be seen in Table 5. Baseline. Our baseline follows the VLA pipeline introduced in RT-2 (Zitkovich et al., 2023), the origin of VLAs, and later adopted by OpenVLA (Kim et al., 2024a). We use LLaMA as the language backbone (Grattafiori et al., 2024). Since LLaMA does not natively support visual inputs, we also paired our backbone with the SigLIP2 (Tschannen et al., 2025) as the vision encoder, A subset of rarely used text tokens is repurposed as action tokens, enabling action pre- diction in the same autoregressive framework. Continuous actions are discretized using a simple binning strategy and modeled as classification over bin indices. We intentionally start from this minimal, classical RT2-style setup to provide a clean reference point for analyzing the effects of different design choices. Our implementation adopts a more recent LLaMA version (LLaMA 3.2) but at a smaller scale (3B parameters), compared to OpenVLA (Kim et al., 2024a). 2.1. The Foundational Components In this section, we investigate some core design choices of VLAs, including architectures and training losses. Policy Module Design.Our baseline follows RT- 2 (Zitkovich et al., 2023) and OpenVLA (Kim et al., 2024a), reusing text tokens for action classification. We first exam- ine whether an explicit policy head is necessary. To this end, we append a class token to the text and visual embeddings and feed its LLM output into a two-layer policy head (trans- former architecture) for action classification (Fig. 3(a)(b)). Results show that introducing a separate policy head per- forms slightly better than directly reusing text tokens (Ta- ble 1), suggesting that decoupling action prediction from the linguistic token space is beneficial. We further investigate whether a more expressive policy module brings additional gains. Specifically, we replace text, image MLLM (a) loose connection meta query noise flows of actions MLLM diffusion actions (b) tight connection flows of actions diffusion text, image (c) soft connection flows of actions diffusion meta query MLLM text, image noise noise text, image MLLM (a) loose connection meta query noise flows of actions MLLM diffusion actions (b) tight connection flows of actions diffusion text, image (c) soft connection flows of actions diffusion meta query MLLM text, image noise noise text, image MLLM (a) loose connection meta query noise flows of actions MLLM diffusion actions (b) tight connection flows of actions diffusion text, image (c) soft connection flows of actions diffusion meta query MLLM text, image noise noise text, image MLLM (a) loose connection meta query noise flows of actions MLLM diffusion actions (b) tight connection flows of actions diffusion text, image (c) soft connection flows of actions diffusion meta query MLLM text, image noise noise text, image MLLM (a) loose connection meta query noise flows of actions MLLM diffusion actions (b) tight connection flows of actions diffusion text, image (c) soft connection flows of actions diffusion meta query MLLM text, image noise noise Figure 4. Design choices for the VLM-Policy connection. the single class token with multiple tokens (16) and ex- pand the policy network from 2 to 12 layers, making the design conceptually similar to MetaQuery (Pan et al., 2025) (Fig. 3(c)). This enlarged policy module yields a significant performance improvement (Table 1). Our final model adopts this design. Action Chunking. Our baseline predicts actions one step at a time. Here, we evaluate action chunking, which pre- dicts multiple future actions jointly and is known to improve inference efficiency (Kim et al., 2025). Results show that longer chunk horizons consistently improve action gener- ation performance (Table 1), suggesting that modeling a longer temporal window of action provides a more coher- ent view of the action sequence. We therefore adopt action chunking with a chunk size of 8. Action Learning Objective. An action chunk is a continu- ous vector of shape(t,dim). Our baseline discretizes this vector using binning (first normalizing toâ1and1, then dividing into 256 bins) and treats action prediction as clas- sification, following OpenVLA (Kim et al., 2024a). We compare this with alternative objectives, including direct regression (Kim et al., 2025), diffusion-based losses such as DDIM (Song et al., 2021; Zhang et al., 2025c), flow matching (Lipman et al., 2021; Lv et al., 2025), and VQ- VAEâbased codebook (codebook size 1024 and each action assigns 3 codes) classification (Van Den Oord et al., 2017; Esser et al., 2021). Results show that regression achieves the strongest perfor- mance, with diffusion-based objectives close behind, while classification-based approaches perform worst (Table 1). In addition, we also notice that when the performance gets higher, the flow-matching objective will eventually outperform the regression loss, since it can represent pre- cise control signals. We therefore adopt the flow-matching objective. We also observe that classification using the VQâVAEâbased codebook underperforms relative to the binning strategy. We attribute this to the fact that the action spaces are low-rank, meaning a simple binning approach provides sufficient resolution. VLM Backbone Capacity. Our baseline uses LLaMA as the backbone (Grattafiori et al., 2024). We evaluate alter- native VLM backbones to study how backbone strength affects VLA performance, including PaliGemma-3B (Beyer 3 VLANeXt: Recipes for Building Strong VLA Models et al., 2024) (used in theÏseries (Black et al., 2024; In- telligence et al., 2025b)) and the Qwen-VL family (Bai et al., 2025a), which represent some of the most capable open-source VLMs currently available. Results show a consistent trend: stronger VLM backbones yield better VLA performance (Table 1), with Qwen3-VL- 4B outperforming Qwen3-VL-2B, which in turn outper- forms LLaMA-3.2-3B and PaliGemma-3B. We use Qwen3- VL-2B in subsequent experiments as a strong yet efficient choice. This finding differs from (Zhang et al., 2026). A possible reason is that our larger policy module can bet- ter exploit the representational capacity of stronger VLMs, whereas the lightweight policy head in (Zhang et al., 2026) may limit such gains. We leave a deeper investigation to future work. VLM-Policy Connection. We next study how different con- nection strategies between the VLM and the policy module affect performance. Our baseline adopts a MetaQuery-style design (Pan et al., 2025), as discussed in âPolicy Module Designâ. We refer to this design as the loose strategy, where the VLM and policy module are fully decoupled. We com- pare this with a tight strategy that connects the two modules layer by layer, as in theÏseries (Black et al., 2024; Intel- ligence et al., 2025b). Inspired by these two designs, we further introduce a soft strategy that also connects them layer by layer but inserts learnable queries as a latent buffer between the modules (Fig. 4). In detail, for the above three connection strategies, we all use the cross attention as the condition technique, and the timestep is conditioned by adaLN, like (Peebles & Xie, 2023). Results show that the soft strategy slightly outperforms both loose and tight connections (Table 1), suggesting that the learnable query buffer helps better transfer useful represen- tations from the VLMâs textual space to the policy moduleâs action space. This may be viewed as introducing a latent buffer between the two components, analogous to reason- ing in a latent space (Hao et al., 2024). We adopt the soft connection in subsequent models. 2.2. The Perception Essentials In this section, we shift our focus from foundational compo- nents to perception, investigating whether and how different modalities (e.g., visual observations and actions) should be provided as inputs to VLAs. Temporal Observation History. We examine whether in- corporating temporal observation history improves perfor- mance. Our baseline follows OpenVLA (Kim et al., 2024a) and uses only the current frame as input. We extend this to include multiple past frames, leveraging the video capabil- ity of the Qwen3-VL-2B backbone (Bai et al., 2025b) for a controlled comparison. Results show that adding tempo- ral history does not improve action generation and slightly degrades performance (Table 1), indicating that redundant temporal inputs may introduce noise or distract the model, even though the backbone is already pretrained in video understanding. Camera View Horizon. We study the effect of camera viewpoints on VLA performance. Our baseline uses a single third-person view, following OpenVLA (Kim et al., 2024a). Many robotics datasets (OâNeill et al., 2024; Khazatsky et al., 2024) additionally provide an in-hand wrist camera, allowing choices between single-view and multi-view in- puts. Results show that combining third-person and wrist views significantly improves performance (Table 1), suggest- ing that multi-view observations provide complementary geometric cues that help resolve spatial ambiguities. Proprioception Conditioning. We examine the role of pro- prioception, which provides information about the robotâs internal state and motion history. Our baseline, following OpenVLA (Kim et al., 2024a), does not use propriocep- tive inputs. We compare three variants: conditioning the VLM, conditioning the policy module, and conditioning both (Fig. 5). In detail, for the VLM part, we will use the proprioception as input, and for the policy part, we will use the action as input to align with the generated action. Results show that conditioning proprioception in the VLM yields the best performance (Table 1). We hypothesize that integrating proprioception at the VLM level allows better fusion with visual and language inputs, whereas injecting it directly into the policy module may reduce reliance of action prediction on visual observations and instructions. Although this appears to differ from the conclusion reported in Zhao et al. (Zhao et al., 2025a), where they claim that proprioception is not needed, their study evaluates architec- tures where proprioception is injected only into the policy module. In that setting, removing proprioception improves performance, which is consistent with our findings. We further compare three different integration mechanisms, including a linear projector, a transformer-based projector, and a transformer projector with masked reconstruction pre- training (He et al., 2022). The transformer-based projector performs slightly better (Table 1); for simplicity, we use the linear projector in the final design. 2.3. Action Modelling Perspectives Here, we examine auxiliary design and training objectives to facilitate action generation. World Modelling. We examine augmenting action predic- tion with an auxiliary world modeling objective (Lv et al., 2025; Cen et al., 2025b). To maintain relatively fair com- parison, we donât use a pretrained visual generator. Instead, we tokenize images using the Emu3.5 image tokenizer (Cui 4 VLANeXt: Recipes for Building Strong VLA Models Table 1. Ablation across the VLA design space on LIBERO, LIBERO-plus (spatial suite). Each block varies in one design aspect. LIBERO- plus evaluates robustness under diverse perturbations. Performance improves steadily as effective design choices are incorporated. LIBEROLIBERO-plus ModelOriginalCamera Robot Language Light Background Noise Layout Total Foundational Components RT-2 like Baseline Baseline19.8------- < 5.0 Policy Module Design Baseline19.8------- < 5.0 Seperate Head30.20.810.031.015.424.84.030.116.6 Large Policy Module64.40.512.679.734.632.98.563.134.0 Action Chunking Horizon Action Chunk 164.40.512.679.734.632.98.563.134.0 Action Chunk 475.45.328.067.942.550.414.870.440.0 Action Chunk 8 74.65.626.085.656.855.811.763.943.4 Action Learning Objective bin Classification 74.65.626.085.656.855.811.763.943.4 VQ-VAE Classification 58.83.244.367.242.543.87.148.336.5 Regression85.45.132.390.562.768.67.775.648.4 DDIM80.04.852.680.370.555.49.468.348.3 Flow Matching80.07.234.679.246.957.811.177.445.0 VLM Backbone Capacity Paligemma69.81.117.132.122.932.62.824.918.9 LLaMA3.2 + SigLip80.07.234.679.246.957.811.177.445.0 Qwen3VL-2B90.09.642.074.675.068.627.983.653.7 Qwen3VL-4B95.812.266.093.389.081.029.988.664.8 VLM-Policy Connection Loose Connection90.09.642.074.675.068.627.983.653.7 Tight Connection90.014.451.781.068.567.825.182.155.4 Soft Connection91.811.858.389.272.974.419.972.556.2 Perception Enssentials Temporal Observation History Current Frame Image91.811.858.389.272.974.419.972.556.2 Temporal Observation History85.07.268.651.565.862.020.880.850.2 Camera View Horizon Third-person camera view91.811.858.389.272.974.419.972.556.2 Multiview (third-person + wrist)97.664.954.091.897.793.085.590.180.5 Proprioception Conditioning No Proprioception Input 97.664.954.091.897.793.085.590.180.5 Proprioception to VLM 98.087.262.286.298.393.892.096.687.7 Proprioception to Policy96.262.869.192.392.596.587.288.383.4 Proprioception to VLM & Policy97.677.973.482.390.194.690.688.884.8 Linear Projector98.087.262.286.298.393.892.096.687.7 Transformer Projector96.496.858.084.695.298.896.993.388.8 Transformer Projector & MAE97.091.151.189.972.886.978.785.678.9 Action Modelling Perspectives World Modelling Perspective Normal 98.087.262.286.298.393.892.096.687.7 World Modelling98.094.480.376.999.798.893.293.290.3 Time Series Forecasting Perspective Normal98.087.262.286.298.393.892.096.687.7 Frequency Domain Loss99.095.778.686.999.798.898.096.693.1 5 VLANeXt: Recipes for Building Strong VLA Models (a) no proprioception input flows of actions diffusion meta query MLLM text, image noise (b) proprioception to VLM flows of actions diffusion meta query MLLM text, image noise prop. (c) action to policy flows of actions diffusion meta query MLLM text, image noise (d) proprioception to VLM & action to policy meta query MLLM text, image action flows of actions diffusion noiseaction linear projector transformer projector or prop. Figure 5. Design choices for proprioception conditioning. (b)worldmodelling image future image generator meta query MLLM text, image noise flows of actions diffusion (a) normal flows of actions diffusion meta query MLLM text, image noise prop.prop. Figure 6. Augmenting action prediction with an auxiliary world modeling objective. et al., 2025) and predict future image tokens with a next- token objective. The target is the future frame at a fixed horizon (8 steps, aligned with the action chunk length). The visual generation module is inserted between the VLM and the policy module with layer-wise connections (Fig. 6). Adding world modeling improves action generation perfor- mance (Table 1), indicating that predicting future observa- tions is beneficial. However, it nearly triples training time, substantially increasing computational cost. We therefore exclude world modeling from the final recipe. Time Series Forecasting. We also explore facilitating ac- tion generation from a time-series forecasting perspective. Inspired by frequency-domain modeling in time-series pre- diction (Zhou et al., 2022; Yi et al., 2023; Yang et al., 2024; Wang et al., 2025b), we introduce a simple auxiliary loss that minimizes the MSE between predicted and ground- truth actions in the frequency domain. We use the discrete cosine transform (Ahmed et al., 1974) to convert the ac- tion to the frequency domain, and assign higher weights to low-frequency components and lower weights to high- frequency components, as high-frequency components are often noisier. This strategy improves action generation performance, slightly surpassing the world modeling objective while adding negligible training overhead (Table 1). The gain likely arises because it serves as a regularization term to avoid the model overfitting to the jitter in trajectory, which mainly improves the modelâs generalization ability. 2.4. Summary of Recipes Starting from a classical RT-2/OpenVLA-style baseline, we find that strong VLA performance emerges from a series of principled design choices. Beneficial changes include: re- placing token reuse with a deeper, dedicated policy module; prop. tokenstext tokensimage tokens z x y roll yaw hist. proprioceptions instruction multi-view frames proprioception projector text embedding vision encoder meta query (soft) Put the black objects into the draw. MLLM pitch Frequency (Hz) z x y roll yaw future actions pitch diffusion noise flows of actions Figure 7. VLANeXt architecture. Multi-view visual inputs, lan- guage instructions, and proprioception are tokenized and processed by a multimodal LLM, with meta queries enabling soft interaction with the policy module. Action chunks are predicted using flow matching and further regularized by a frequency-domain objective. adopting action chunking to model longer temporal action horizons; using continuous objectives such as flow matching (with regression also effective under simple distributions); employing a stronger VLM backbone (Qwen3-VL-2B as an effectiveâefficient choice); and connecting the VLM and policy module through soft, layer-wise interactions with learnable query buffers. On the perception side, multi-view inputs (third-person + wrist) and VLM-side proprioception conditioning improve performance, while redundant temporal observation history is unnecessary. Moreover, adding a lightweight frequency- domain auxiliary loss further boosts action generation with negligible cost. Although world modeling also improves performance, its substantially higher training cost makes it less practical. Together, these choices form a practical recipe for building a strong and efficient VLA model, which we call VLANeXt. 3. Benchmarks Evaluations 3.1. Settings To evaluate both standard performance and generalization robustness, we employ the LIBERO ecosystem. We first evaluate our VLANeXt on the standard LIBERO bench- mark (Liu et al., 2023a), which provides four distinct cat- egories (Spatial, Object, Goal, and Long) to test the task learning ability, each providing 500 expert demonstrations 6 VLANeXt: Recipes for Building Strong VLA Models across 10 tasks to assess policy generalization to different spatial layouts, objects, goals, and long-horizon tasks. To test the generalization boundaries of our model fur- ther, we evaluate our method on LIBERO-plus (Fei et al., 2025b). Unlike the static conditions in standard LIBERO, LIBERO-plus introduces systematic variations to the evalua- tion episodes, comprising 10,030 demonstrations across the above four suites in LIBERO, with perturbations in visual (e.g., lighting, background, camera pose), physical (e.g., object layout, robot state), and semantic (e.g., language instruction rewrites) dimensions. Following the standard setting in OpenVLA (Kim et al., 2024a), we train our models on the modified LIBERO dataset for each suite (Spatial, Object, Goal, and Long), and evaluate performance on both the LIBERO and LIBERO- plus benchmarks (which include unseen perturbations) for the corresponding suite. For fair comparisons across differ- ent design choices, we directly fine-tune all models on the LIBERO dataset. All experiments in our recipes use 10,000 training steps with a batch size of 256. The learning rate is set to1Ă 10 â4 for models smaller than 3B parameters and 5Ă 10 â5 otherwise in our explorations. 3.2. LIBERO Benchmark Results On the LIBERO benchmark, we compare our method against two categories of approaches: (i) direct policy learn- ing methods that are trained solely on robotic datasets, and (i) VLA methods that leverage knowledge from pre- trained VLMs for policy learning. For direct policy learning methods, we include Diffusion Policy (Chi et al., 2025), Octo (Ghosh et al., 2024), and MDT (Reuss et al., 2024). For VLA methods, we compare against OpenVLA (Kim et al., 2024a), TraceVLA (Zheng et al., 2025a), Spa- tialVLA (Qu et al., 2025), WorldVLA (Cen et al., 2025b), CoT-VLA (Zhao et al., 2025b),Ï 0 (Black et al., 2024), Ï 0 -Fast (Pertsch et al., 2025), NORA (Hung et al., 2025), SmolVLA (Shukor et al., 2025), UniVLA (Wang et al., 2025e), FLOWER (Reuss et al., 2025), and OpenVLA- OFT (Kim et al., 2025). The comparison results are shown in Table 2. As we can ob- serve, following our recipes allows us to build a strong VLA that achieves state-of-the-art performance, demonstrating the effectiveness of the design choices. 3.3. LIBERO-plus Benchmark Results For the LIBERO-plus benchmark, we compare our model with several VLA models such as OpenVLA (Kim et al., 2024a), WorldVLA (Cen et al., 2025b), NORA (Hung et al., 2025), UniVLA (Wang et al., 2025e),Ï o (Black et al., 2024), Ï o -Fast (Pertsch et al., 2025), and OpenVLA-OFT (Kim et al., 2025). Table 2. LIBERO benchmark performance. The results are shown in success rate (%). S, O, G, L: Spatial, Object, Goal, and Long suites, respectively. We color thebestandsecond bestresults. ModelSOGLAvg Baseline Direct Policy Models Diffusion Policy 78.392.568.350.572.4 Octo78.985.784.651.575.1 MDT78.587.573.564.876.4 Baseline VLA Models TraceVLA84.685.275.154.174.8 OpenVLA84.788.479.253.776.5 SpatialVLA88.289.978.655.578.1 WorldVLA 85.689.082.659.079.1 CoT-VLA87.591.687.669.083.9 Ï 0 -Fast96.496.888.660.285.5 Ï 0 90.086.095.073.086.0 NORA92.295.489.474.687.9 SmolVLA 93.094.091.077.088.8 UniVLA96.596.895.692.095.2 FLOWER97.599.196.194.996.9 Ï 0.5 98.898.29892.496.9 OpenVLA-OFT97.698.497.994.597.1 Ours VLANeXt 99.099.296.694.897.4 As shown in Table 3, the proposed VLANeXt model demon- strates strong generalization ability across different types of unseen perturbations. Moreover, our model shows a sig- nificant improvement (13% in success rate) over the state- of-the-art method OpenVLA-OFT (Kim et al., 2025) on the LIBERO-plus benchmark compared to previous methods, suggesting the effectiveness of the explored recipes. 4. Real-World Evaluations To comprehensively assess the performance of our method, we additionally evaluate it in real-world deployments. 4.1. Settings We design four tasks, including two single-arm tasks and two bimanual tasks, to evaluate our method. The single- arm tasks include table cleaning, which involves picking up objects from a table and placing them into a container, and drawer manipulation, where the robot opens a drawer, places objects inside, and closes it. The bimanual tasks include basket lifting, which requires lifting a basket using both hands, and bimanual table cleaning, where two arms coordinate to collect objects from a table and place them into a container. The single-arm experiments use Franka Emika, while the bimanual experiments are conducted on the Aloha system (Zhao et al., 2023). A visualization of the experimental setup for each task is depicted in Figure 8. For training, we collect 50 episodes per task and evaluate each model over 20 trials, reporting the success rate. We 7 VLANeXt: Recipes for Building Strong VLA Models Table 3. LIBERO-plus benchmark performance. The results are shown in success rate (%). We color thebestandsecond bestresults (in average). The complete per-suite results of the listed methods can also be found in (Fei et al., 2025b). ModelSuiteCameraRobotLanguageLightBackgroundNoiseLayoutTotal Baseline VLA Models OpenVLAAverage0.83.523.08.134.815.228.515.6 WorldVLAAverage0.127.941.643.717.110.938.025.0 NORAAverage2.237.065.145.758.612.862.139.0 UniVLA Average1.846.269.669.081.021.231.942.9 Ï 0 Average13.86.058.885.081.479.068.953.6 Ï 0 -FastAverage65.121.661.073.273.274.468.861.6 OpenVLA-OFT Spatial88.340.080.598.397.396.393.984.0 Object38.925.499.073.797.672.371.866.5 Goal62.025.253.293.992.575.259.163.0 Long38.738.287.089.486.863.576.966.4 Average56.431.979.588.793.375.874.269.6 Ours VLANeXt Spatial95.778.686.999.798.898.096.693.1 Object99.548.598.699.384.799.878.286.5 Goal96.663.651.597.570.896.863.976.2 Long69.772.090.386.975.881.784.379.7 Average90.465.781.895.982.594.180.883.9 (a) Clean Table(b) Open Drawer and Place Object (d) Bimanual Clean Table(c) Lifting Figure 8. Our single-arm and bimanual arm tasks for the real-world experiments. first pretrain the model on the DROID dataset (Khazatsky et al., 2024) for 100k steps, then fine-tune it on each task for 20k steps with a learning rate of1Ă 10 â4 . Because DROID contains only single-arm data, adapting the model to bimanual tasks requires reinitializing the proprioception projector and the final layer of the action generation module, while keeping all other pretrained weights. 4.2. Results We compare against two representative VLA baselines, OpenVLA-OFT (Kim et al., 2025) andÏ 0 (Black et al., 2024). We load their pretrained checkpoints and fine-tune Table 4. Real-world evaluation results. Results are shown with (success count/total count). We color thebestandsecond best. Single ArmBimanual Arm Model CleanDrawerCleanLifting Baseline VLA Models OpenVLA-OFT 7/207/205/209/20 Ï 0 10/208/2010/2013/20 Ours VLANeXt 14/2011/2011/2015/20 them on each task in the same manner as our method to ensure a fair comparison. The results are shown in Table 4. As can be seen, our model performs well in real-world ex- periments, demonstrating that the recipes we propose lead to a strong VLA model that can be effectively deployed in real- world settings. In addition, even without bimanual training, our method can adapt to bimanual robotics tasks with decent performance, demonstrating the cross-embodiment adapt- ability of the method. Additional video demonstrations of our experimental results are provided in the supplementary materials. 5. Conclusion This work moves toward a more systematic understanding of VLA models. Rather than introducing another standalone architecture, we revisit the VLA pipeline and show that many gains arise from principled design choices within a unified framework. In particular, how the VLM interacts with the policy module, how multimodal signals such as proprioception are fused, and how temporal structure in ac- tions is modeled all play central roles. Several observations carry broader implications. Modest architectural refine- 8 VLANeXt: Recipes for Building Strong VLA Models ments, such as soft VLMâpolicy coupling or VLM-side proprioception conditioning, can meaningfully influence performance, indicating that where information is injected matters as much as what information is used. Viewing action generation as structured sequence modeling, for example, through frequency-domain objectives, also shows that ideas from time-series learning transfer effectively to robotics. Meanwhile, richer objectives like world modeling improve performance but introduce notable computational overhead, highlighting the importance of efficiency-aware design. We hope this work encourages a shift from ad-hoc model variations toward more controlled exploration of the VLA design space. By releasing a unified, lightweight framework, we aim to support systematic studies and shared progress. Extending this perspective to more diverse embodiments, longer-horizon reasoning, extensive mid-training, and richer world-interaction objectives remains an important direction for future research. Acknowledgements This research is supported by cash and in-kind funding from NTU S-Lab and industry partner(s). It is also sup- ported by Singapore MOE AcRF Tier 2 (MOE-T2EP20224- 0003), and by Guangdong Key Research and Development Program (No.2024B0101040004, No. 2025B0909020002). Also, thanks to victkk (Zicheng Zhang) for giving us valu- able suggestions to improve the paper. Impact Statement This paper presents work aimed at advancing the field of ma- chine learning, specifically Vision-Language-Action models for robotic control. While our work contributes to the devel- opment of more capable embodied agents, we believe that its potential societal implications fall within well-established discussions in the field and therefore do not require special emphasis here. References Ahmed, N., Natarajan, T., and Rao, K. R. Discrete cosine transform. IEEE Transactions on Computers, 1974. Bai, S., Cai, Y., Chen, R., Chen, K., et al. Qwen3-vl techni- cal report. arXiv preprint arXiv:2511.21631, 2025a. Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025b. Bai, Z., Gao, C., and Shou, M. Z. Evolve-vla: Test-time training from environment feedback for vision-language- action models. arXiv preprint arXiv:2512.14666, 2025c. Beyer, L., Steiner, A., Pinto, A. S., Kolesnikov, A., Wang, X., Salz, D., Neumann, M., Alabdulmohsin, I., Tschan- nen, M., Bugliarello, E., et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. Bhat, V., Lan, Y.-H., Krishnamurthy, P., Karri, R., and Khor- rami, F. 3d cavla: Leveraging depth and 3d context to generalize vision language action models for unseen tasks. arXiv preprint arXiv:2505.05800, 2025. Bi, H., Tan, H., Xie, S., Wang, Z., Huang, S., Liu, H., Zhao, R., Feng, Y., Xiang, C., Rong, Y., et al. Mo- tus: A unified latent action world model. arXiv preprint arXiv:2512.13030, 2025. Bjorck, J., Casta Ì neda, F., Cherniadev, N., Da, X., Ding, R., Fan, L., Fang, Y., Fox, D., Hu, F., Huang, S., et al. Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734, 2025. Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., et al. pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al. Rt-1: Robotics transformer for real-world control at scale. In RSS, 2023. Cai, J.-F., Chen, Z., Wu, X.-M., Jiang, J.-J., Wei, Y.-L., and Zheng, W.-S. Real-to-sim grasp: Rethinking the gap between simulation and real world in grasp detection. In CoRL, 2024. Cen, J., Huang, S., Yuan, Y., Li, K., Yuan, H., Yu, C., Jiang, Y., Guo, J., Li, X., Luo, H., et al. Rynnvla-002: A unified vision-language-action and world model. arXiv preprint arXiv:2511.17502, 2025a. Cen, J., Yu, C., Yuan, H., Jiang, Y., Huang, S., Guo, J., Li, X., Song, Y., Luo, H., Wang, F., et al. Worldvla: To- wards autoregressive action world model. arXiv preprint arXiv:2506.21539, 2025b. Cheang, C.-L., Chen, G., Jing, Y., Kong, T., Li, H., Li, Y., Liu, Y., Wu, H., Xu, J., Yang, Y., et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. Chen, N., Wu, X.-M., Xu, G., Jiang, J.-J., Chen, Z., and Zheng, W.-S. Motiongrasp: Long-term grasp motion tracking for dynamic grasping. RA-L, 2024. Chen, P., Bu, P., Wang, Y., Wang, X., Wang, Z., Guo, J., Zhao, Y., Zhu, Q., Song, J., Yang, S., et al. Combatvla: An efficient vision-language-action model for combat 9 VLANeXt: Recipes for Building Strong VLA Models tasks in 3d action role-playing games. arXiv preprint arXiv:2503.09527, 2025a. Chen, Z., Niu, R., Kong, H., Wang, Q., Xing, Q., and Fan, Z. Tgrpo: Fine-tuning vision-language-action model via trajectory-wise group relative policy optimization. arXiv preprint arXiv:2506.08440, 2025b. Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., and Song, S. Diffusion policy: Visuomo- tor policy learning via action diffusion. IJRR, 2025. Cui, Y., Chen, H., Deng, H., Huang, X., Li, X., Liu, J., Liu, Y., Luo, Z., Wang, J., Wang, W., et al. Emu3. 5: Native multimodal models are world learners. arXiv preprint arXiv:2510.26583, 2025. Ding, P., Ma, J., Tong, X., Zou, B., Luo, X., Fan, Y., Wang, T., Lu, H., Mo, P., Liu, J., et al. Humanoid-vla: Towards universal humanoid control with visual integration. arXiv preprint arXiv:2502.14795, 2025. Du, Y., Yang, S., Dai, B., Dai, H., Nachum, O., Tenenbaum, J., Schuurmans, D., and Abbeel, P. Learning universal policies via text-guided video generation. In NeurIPS, 2023. Esser, P., Rombach, R., and Ommer, B. Taming transformers for high-resolution image synthesis. In CVPR, 2021. Fang, H.-S., Wang, C., Gou, M., and Lu, C. Graspnet- 1billion: A large-scale benchmark for general object grasping. In CVPR, 2020. Fei, S., Wang, S., Ji, L., Li, A., Zhang, S., Liu, L., Hou, J., Gong, J., Zhao, X., and Qiu, X. Srpo: Self-referential policy optimization for vision-language-action models. arXiv preprint arXiv:2511.15605, 2025a. Fei, S., Wang, S., Shi, J., Dai, Z., Cai, J., Qian, P., Ji, L., He, X., Zhang, S., Fei, Z., et al. Libero-plus: In-depth robust- ness analysis of vision-language-action models. arXiv preprint arXiv:2510.13626, 2025b. Fu, Y., Zhang, Z., Zhang, Y., Wang, Z., Huang, Z., and Luo, Y. Mergevla: Cross-skill model merging toward a generalist vision-language-action agent. arXiv preprint arXiv:2511.18810, 2025. Ghosh, D., Walke, H. R., Pertsch, K., Black, K., Mees, O., Dasari, S., Hejna, J., Kreiman, T., Xu, C., Luo, J., et al. Octo: An open-source generalist robot policy. In RSS, 2024. Goyal, A., Hadfield, H., Yang, X., Blukis, V., and Ramos, F. Vla-0: Building state-of-the-art vlas with zero modifica- tion. arXiv preprint arXiv:2510.13054, 2025. Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. Guo, W., Lu, G., Deng, H., Wu, Z., Tang, Y., and Wang, Z. Vla-reasoner: Empowering vision-language-action models with reasoning via online monte carlo tree search. arXiv preprint arXiv:2509.22643, 2025. Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J., and Tian, Y. Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769, 2024. He, K., Chen, X., Xie, S., Li, Y., Doll Ì ar, P., and Girshick, R. Masked autoencoders are scalable vision learners. In CVPR, 2022. Huang, C.-P., Wu, Y.-H., Chen, M.-H., Wang, Y.-C. F., and Yang, F.-E. Thinkact: Vision-language-action reason- ing via reinforced visual latent planning. arXiv preprint arXiv:2507.16815, 2025a. Huang, J., Wang, S., Lin, F., Hu, Y., Wen, C., and Gao, Y. Tactile-vla: unlocking vision-language-action modelâs physical knowledge for tactile generalization. arXiv preprint arXiv:2507.09160, 2025b. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., and Zhu, S.-C. Diffusion-based generation, optimiza- tion, and planning in 3d scenes. In CVPR, 2023a. Huang, W., Abbeel, P., Pathak, D., and Mordatch, I. Lan- guage models as zero-shot planners: Extracting action- able knowledge for embodied agents. In ICML, 2022. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., and Fei-Fei, L. Voxposer: Composable 3d value maps for robotic manipulation with language models. In CoRL, 2023b. Huang, W., Wang, C., Li, Y., Zhang, R., and Fei-Fei, L. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. In CoRL, 2025c. Hung, C.-Y., Sun, Q., Hong, P., Zadeh, A., Li, C., Tan, U., Majumder, N., Poria, S., et al. Nora: A small open- sourced generalist vision language action model for em- bodied tasks. arXiv preprint arXiv:2504.19854, 2025. Intelligence, P., Amin, A., Aniceto, R., Balakrishna, A., Black, K., Conley, K., Connors, G., Darpinian, J., Dha- balia, K., DiCarlo, J., et al. pi0.6: a vla that learns from experience. arXiv preprint arXiv:2511.14759, 2025a. Intelligence, P., Black, K., Brown, N., Darpinian, J., Dha- balia, K., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., et al. pi0.5: a vision-language-action model with open- world generalization. arXiv preprint arXiv:2504.16054, 2025b. 10 VLANeXt: Recipes for Building Strong VLA Models Janner, M., Du, Y., Tenenbaum, J., and Levine, S. Planning with diffusion for flexible behavior synthesis. In ICML, 2022. Jiang, J.-J., Wu, X.-M., He, Y.-X., Zeng, L.-A., Wei, Y.- L., Zhang, D., and Zheng, W.-S. Rethinking bimanual robotic manipulation: Learning with decoupled interac- tion framework. In ICCV, 2025. Kareer, S., Pertsch, K., Darpinian, J., Hoffman, J., Xu, D., Levine, S., Finn, C., and Nair, S. Emergence of human to robot transfer in vision-language-action models. arXiv preprint arXiv:2512.22414, 2025. Khazatsky, A., Pertsch, K., Nair, S., Balakrishna, A., Dasari, S., Karamcheti, S., Nasiriany, S., Srirama, M. K., Chen, L. Y., Ellis, K., et al. Droid: A large-scale in-the-wild robot manipulation dataset. In RSS, 2024. Kim, M. J., Pertsch, K., Karamcheti, S., Xiao, T., Balakr- ishna, A., Nair, S., Rafailov, R., Foster, E., Lam, G., San- keti, P., et al. Openvla: An open-source vision-language- action model. In CoRL, 2024a. Kim, M. J., Finn, C., and Liang, P. Fine-tuning vision- language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025. Kim, Y., Oh, H., Lee, J., Choi, J., Ji, G., Jung, M., Youm, D., and Hwangbo, J. Not only rewards but also constraints: Applications on legged robot locomotion. TRO, 2024b. Kuang, F., You, J., Hu, Y., Zhang, T., Wen, C., and Gao, Y. Adapt your body: Mitigating proprioception shifts in imitation learning. arXiv preprint arXiv:2506.23944, 2025. Kumar, A., Fu, Z., Pathak, D., and Malik, J. Rma: Rapid motor adaptation for legged robots. In RSS, 2021. Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., and Hutter, M. Learning quadrupedal locomotion over challenging terrain. Science Robotics, 2020. Lee, J., Duan, J., Fang, H., Deng, Y., Liu, S., Li, B., Fang, B., Zhang, J., Wang, Y. R., Lee, S., et al. Molmoact: Action reasoning models that can reason in space. arXiv preprint arXiv:2508.07917, 2025. Li, H., Lin, X., Zhou, Y., Li, X., Huo, Y., Chen, J., and Ye, Q. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. In IJCAI, 2023a. Li, H., Yang, S., Chen, Y., Chen, X., Yang, X., Tian, Y., Wang, H., Wang, T., Lin, D., Zhao, F., et al. Cronusvla: Towards efficient and robust manipulation via multi- frame vision-language-action modeling. arXiv preprint arXiv:2506.19816, 2025a. Li, H., Zuo, Y., Yu, J., Zhang, Y., Yang, Z., Zhang, K., Zhu, X., Zhang, Y., Chen, T., Cui, G., et al. Simplevla-rl: Scaling vla training via reinforcement learning. arXiv preprint arXiv:2509.09674, 2025b. Li, Q., Liang, Y., Wang, Z., Luo, L., Chen, X., Liao, M., Wei, F., Deng, Y., Xu, S., Zhang, Y., et al. Cogact: A foundational vision-language-action model for synergiz- ing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650, 2024. Li, S., Gao, Y., Sadigh, D., and Song, S. Unified video action model. In RSS, 2025c. Li, X., Liu, M., Zhang, H., Yu, C., Xu, J., Wu, H., Cheang, C., Jing, Y., Zhang, W., Liu, H., et al. Vision-language foundation models as effective robot imitators. In ICLR, 2023b. Liang, H., Chen, X., Wang, B., Chen, M., Liu, Y., Zhang, Y., Chen, Z., Yang, T., Chen, Y., Pang, J., et al. Mm-act: Learn from multimodal parallel generation to act. arXiv preprint arXiv:2512.00975, 2025. Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. In ICLR, 2021. Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., and Stone, P. Libero: Benchmarking knowledge transfer for lifelong robot learning. In NeurIPS, 2023a. Liu, H., Li, X., Li, P., Liu, M., Wang, D., Liu, J., Kang, B., Ma, X., Kong, T., and Zhang, H. Towards generalist robot policies: What matters in building vision-language-action models. Nature Machine Intelligence, 2026. Liu, J., Zhang, R., Fang, H.-S., Gou, M., Fang, H., Wang, C., Xu, S., Yan, H., and Lu, C. Target-referenced reactive grasping for dynamic objects. In CVPR, 2023b. Liu, M., Pathak, D., and Agarwal, A. Locoformer: Gener- alist locomotion via long-context adaptation. In CoRL, 2025. Liu, S., Zhou, Y., Yang, J., Gupta, S., and Wang, S. Contact- gen: Generative contact modeling for grasp generation. In ICCV, 2023c. Liu, S., Wu, L., Li, B., Tan, H., Chen, H., Wang, Z., Xu, K., Su, H., and Zhu, J. Rdt-1b: a diffusion foundation model for bimanual manipulation. In ICLR, 2024. Lu, G., Zhang, S., Wang, Z., Liu, C., Lu, J., and Tang, Y. Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation. In ECCV, 2024. 11 VLANeXt: Recipes for Building Strong VLA Models Lu, G., Guo, W., Zhang, C., Zhou, Y., Jiang, H., Gao, Z., Tang, Y., and Wang, Z. Vla-rl: Towards masterful and general robotic manipulation with scalable reinforcement learning. arXiv preprint arXiv:2505.18719, 2025. Lv, Q., Kong, W., Li, H., Zeng, J., Qiu, Z., Qu, D., Song, H., Chen, Q., Deng, X., and Pang, J. F1: A vision-language- action model bridging understanding and generation to actions. arXiv preprint arXiv:2509.06951, 2025. Lyu, J., Li, Z., Shi, X., Xu, C., Wang, Y., and Wang, H. Dywa: Dynamics-adaptive world action model for gener- alizable non-prehensile manipulation. In ICCV, 2025. Ma, Y., Song, Z., Zhuang, Y., Hao, J., and King, I. A survey on vision-language-action models for embodied ai. arXiv preprint arXiv:2405.14093, 2024. Margolis, G. B. and Agrawal, P. Walk these ways: Tun- ing robot control for generalization with multiplicity of behavior. In CoRL, 2023. Mason, M. T. Progress in nonprehensile manipulation. IJRR, 1999. Mordatch, I., Popovi Ì c, Z., and Todorov, E. Contact-invariant optimization for hand manipulation. In SIGGRAPH, 2012. OâNeill, A., Rehman, A., Maddukuri, A., Gupta, A., Padalkar, A., Lee, A., Pooley, A., Gupta, A., Mandlekar, A., Jain, A., et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collabora- tion 0. In ICRA, 2024. Pan, X., Shukla, S. N., Singh, A., Zhao, Z., Mishra, S. K., Wang, J., Xu, Z., Chen, J., Li, K., Juefei-Xu, F., Hou, J., and Xie, S. Transfer between modalities with metaqueries. arXiv preprint arXiv:2504.06256, 2025. Peebles, W. and Xie, S. Scalable diffusion models with transformers. In CVPR, 2023. Peng, X. B., Abbeel, P., Levine, S., and Van de Panne, M. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACMTOG, 2018. Pertsch, K., Stachowicz, K., Ichter, B., Driess, D., Nair, S., Vuong, Q., Mees, O., Finn, C., and Levine, S. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747, 2025. Qu, D., Song, H., Chen, Q., Yao, Y., Ye, X., Ding, Y., Wang, Z., Gu, J., Zhao, B., Wang, D., et al. Spatialvla: Exploring spatial representations for visual-language-action model. arXiv preprint arXiv:2501.15830, 2025. Radosavovic, I., Shi, B., Fu, L., Goldberg, K., Darrell, T., and Malik, J. Robot learning with sensorimotor pre- training. In CoRL, 2023. Ravichandar, H., Polydoros, A. S., Chernova, S., and Billard, A. Recent advances in robot learning from demonstration. ARCRAS, 2020. Reuss, M., Ya Ì gmurlu, Ì O. E., Wenzel, F., and Lioutikov, R. Multimodal diffusion transformer: Learning versatile behavior from multimodal goals. In RSS, 2024. Reuss, M., Zhou, H., R Ì uhle, M., Ya Ì gmurlu, Ì O. E., Otto, F., and Lioutikov, R. Flower: Democratizing generalist robot policies with efficient vision-language-action flow policies. arXiv preprint arXiv:2509.04996, 2025. Shi, H., Xie, B., Liu, Y., Sun, L., Liu, F., Wang, T., Zhou, E., Fan, H., Zhang, X., and Huang, G. Memo- ryvla: Perceptual-cognitive memory in vision-language- action models for robotic manipulation. arXiv preprint arXiv:2508.19236, 2025. Shridhar, M., Manuelli, L., and Fox, D. Perceiver-actor: A multi-task transformer for robotic manipulation. In CoRL, 2023. Shukor, M., Aubakirova, D., Capuano, F., Kooijmans, P., Palma, S., Zouitine, A., Aractingi, M., Pascal, C., Russi, M., Marafioti, A., et al. Smolvla: A vision-language- action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025. Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. In ICLR, 2021. Song, W., Zhou, Z., Zhao, H., Chen, J., Ding, P., Yan, H., Huang, Y., Tang, F., Wang, D., and Li, H. Reconvla: Reconstructive vision-language-action model as effective robot perceiver. In AAAI, 2025. Tan, S., Dou, K., Zhao, Y., and Kraehenbuehl, P. Interac- tive post-training for vision-language-action models. In CVPR, 2025. Team, G. R., Abeyruwan, S., Ainslie, J., Alayrac, J.-B., Arenas, M. G., Armstrong, T., Balakrishna, A., Baruch, R., Bauza, M., Blokzijl, M., et al. Gemini robotics: Bringing ai into the physical world. arXiv preprint arXiv:2503.20020, 2025. Tian, Y., Yang, S., Zeng, J., Wang, P., Lin, D., Dong, H., and Pang, J. Predictive inverse dynamics models are scalable learners for robotic manipulation. In ICLR, 2024. Tschannen, M., Gritsenko, A., Wang, X., Naeem, M. F., Alabdulmohsin, I., Parthasarathy, N., Evans, T., Beyer, L., Xia, Y., Mustafa, B., et al. Siglip 2: Multilingual 12 VLANeXt: Recipes for Building Strong VLA Models vision-language encoders with improved semantic under- standing, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. Van Den Oord, A., Vinyals, O., et al. Neural discrete repre- sentation learning. In NeurIPS, 2017. Wang, A.-L., Chen, N., Lin, K.-Y., Li, Y.-M., and Zheng, W.-S. Task-oriented 6-dof grasp pose detection in clutters. In ICRA, 2025a. Wang, C., Fang, H.-S., Gou, M., Fang, H., Gao, J., and Lu, C. Graspness discovery in clutters for fast and accurate grasp detection. In ICCV, 2021. Wang, H., Pan, L., Shen, Y., Chen, Z., Yang, D., Yang, Y., Zhang, S., Liu, X., Li, H., and Tao, D. Fredf: Learning to forecast in the frequency domain. In ICLR, 2025b. Wang, L., Chen, X., Zhao, J., and He, K.Scaling proprioceptive-visual learning with heterogeneous pre- trained transformers. In NeurIPS, 2024. Wang, S., Yu, W., Chen, X., Tian, X., Zhang, J., Lu, L., and Zhang, C. End-to-end listen, look, speak and act. arXiv preprint arXiv:2510.16756, 2025c. Wang, Y., Ding, P., Li, L., Cui, C., Ge, Z., Tong, X., Song, W., Zhao, H., Zhao, W., Hou, P., et al. Vla-adapter: An effective paradigm for tiny-scale vision-language-action model. arXiv preprint arXiv:2509.09372, 2025d. Wang, Y., Li, X., Wang, W., Zhang, J., Li, Y., Chen, Y., Wang, X., and Zhang, Z. Unified vision-language-action model. arXiv preprint arXiv:2506.19850, 2025e. Wei, Y.-L., Jiang, J.-J., Xing, C., Tan, X.-T., Wu, X.-M., Li, H., Cutkosky, M., and Zheng, W.-S. Grasp as you say: Language-guided dexterous grasp generation. In NeurIPS, 2024. Wei, Y.-L., Lin, M., Lin, Y., Jiang, J.-J., Wu, X.-M., Zeng, L.-A., and Zheng, W.-S. Afforddexgrasp: Open- set language-guided dexterous grasp with generalizable- instructive affordance. In ICCV, 2025. Wen, J., Zhu, Y., Li, J., Zhu, M., Tang, Z., Wu, K., Xu, Z., Liu, N., Cheng, R., Shen, C., et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. RAL, 2025. Wu, H., Jing, Y., Cheang, C., Chen, G., Xu, J., Li, X., Liu, M., Li, H., and Kong, T. Unleashing large-scale video generative pre-training for visual robot manipulation. In ICLR, 2024a. Wu, P., Escontrela, A., Hafner, D., Abbeel, P., and Gold- berg, K. Daydreamer: World models for physical robot learning. In CoRL, 2023. Wu, X.-M., Cai, J.-F., Jiang, J.-J., Zheng, D., Wei, Y.-L., and Zheng, W.-S. An economic framework for 6-dof grasp detection. In ECCV, 2024b. Xiao, J., Yang, Y., Chang, X., Chen, R., Xiong, F., Xu, M., Zheng, W.-S., and Zhang, Q. World-env: Leveraging world model as a virtual environment for vla post-training. arXiv preprint arXiv:2509.24948, 2025a. Xiao, L., Li, J., Gao, J., Ye, F., Jin, Y., Qian, J., Zhang, J., Wu, Y., and Yu, X. Ava-vla: Improving vision-language- action models with active visual attention. arXiv preprint arXiv:2511.18960, 2025b. Xiao, X., Liu, J., Wang, Z., Zhou, Y., Qi, Y., Jiang, S., He, B., and Cheng, Q. Robot learning in the era of foundation models: A survey. Neurocomputing, 2025c. Xu, G.-H., Wei, Y.-L., Zheng, D., Wu, X.-M., and Zheng, W.-S. Dexterous grasp transformer. In CVPR, 2024. Yang, R., Cao, L., YANG, J., et al. Rethinking fourier transform from a basis functions perspective for long- term time series forecasting. In NeurIPS, 2024. Ye, S., Jang, J., Jeon, B., Joo, S., Yang, J., Peng, B., Man- dlekar, A., Tan, R., Chao, Y.-W., Lin, B. Y., et al. Latent action pretraining from videos. In CoRL, 2024. Yi, K., Zhang, Q., Fan, W., Wang, S., Wang, P., He, H., An, N., Lian, D., Cao, L., and Niu, Z. Frequency-domain mlps are more effective learners in time series forecasting. In NeurIPS, 2023. Ze, Y., Zhang, G., Zhang, K., Hu, C., Wang, M., and Xu, H. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In ICRA, 2024. Zhang, J., Chen, Y., Xu, Y., Huang, Z., Zhou, Y., Yuan, Y.-J., Cai, X., Huang, G., Quan, X., Xu, H., et al. 4d-vla: Spatiotemporal vision-language-action pretraining with cross-scene calibration. arXiv preprint arXiv:2506.22242, 2025a. Zhang, J., Guo, Y., Hu, Y., Chen, X., Zhu, X., and Chen, J. Up-vla: A unified understanding and prediction model for embodied agent. In ICML, 2025b. Zhang, J., Chen, X., Wang, Q., Li, M., Guo, Y., Hu, Y., Zhang, J., Bai, S., Lin, J., and Chen, J. Vlm4vla: Revis- iting vision-language-models in vision-language-action models. arXiv preprint arXiv:2601.03309, 2026. Zhang, W., Liu, H., Qi, Z., Wang, Y., Yu, X., Zhang, J., Dong, R., He, J., Lu, F., Wang, H., et al. Dreamvla: a vision-language-action model dreamed with comprehen- sive world knowledge. In NeurIPS, 2025c. 13 VLANeXt: Recipes for Building Strong VLA Models Zhang, Z., Zheng, K., Chen, Z., Jang, J., Li, Y., Han, S., Wang, C., Ding, M., Fox, D., and Yao, H. Grape: Gener- alizing robot policy via preference alignment. In ICRA, 2025d. Zhao, J., Lu, W., Zhang, D., Liu, Y., Liang, Y., Zhang, T., Cao, Y., Xie, J., Hu, Y., Wang, S., et al. Do you need proprioceptive states in visuomotor policies? arXiv preprint arXiv:2509.18644, 2025a. Zhao, Q., Lu, Y., Kim, M. J., Fu, Z., Zhang, Z., Wu, Y., Li, Z., Ma, Q., Han, S., Finn, C., et al. Cot-vla: Visual chain- of-thought reasoning for vision-language-action models. In CVPR, 2025b. Zhao, T., Kumar, V., Levine, S., and Finn, C. Learning fine- grained bimanual manipulation with low-cost hardware. In RSS, 2023. Zhen, H., Qiu, X., Chen, P., Yang, J., Yan, X., Du, Y., Hong, Y., and Gan, C. 3d-vla: A 3d vision-language-action generative world model. In ICML, 2024. Zheng, R., Liang, Y., Huang, S., Gao, J., Daum Ì e I, H., Kolobov, A., Huang, F., and Yang, J. Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies. In ICLR, 2025a. Zheng, Z., Cai, J.-F., Wu, X.-M., Wei, Y.-L., Tang, Y.-M., Wu, A., and Zheng, W.-S. imanip: Skill-incremental learning for robotic manipulation. In ICCV, 2025b. Zhong, Z., Yan, H., Li, J., Liu, X., Gong, X., Zhang, T., Song, W., Chen, J., Zheng, X., Wang, H., et al. Flowvla: Visual chain of thought-based motion reason- ing for vision-language-action models. arXiv preprint arXiv:2508.18269, 2025. Zhou, J., Ma, T., Lin, K.-Y., Wang, Z., Qiu, R., and Liang, J. Mitigating the human-robot domain discrepancy in visual pre-training for robotic manipulation. In CVPR, 2025a. Zhou, J., Ye, K., Liu, J., Ma, T., Wang, Z., Qiu, R., Lin, K.-Y., Zhao, Z., and Liang, J. Exploring the limits of vision-language-action manipulation in cross-task gener- alization. In NeurIPS, 2026. Zhou, T., Ma, Z., Wen, Q., Wang, X., Sun, L., and Jin, R. Fedformer: Frequency enhanced decomposed trans- former for long-term series forecasting. In ICML, 2022. Zhou, Z., Zhu, Y., Zhu, M., Wen, J., Liu, N., Xu, Z., Meng, W., Peng, Y., Shen, C., Feng, F., et al. Chatvla: Unified multimodal understanding and robot control with vision- language-action model. In EMNLP, 2025b. Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 14 VLANeXt: Recipes for Building Strong VLA Models A. More Experimental Results A.1. Qualitative Experiments We present more demos of our model on the LIBERO and LIBERO-plus benchmarks, as well as in real-world settings (see Figures 10, 11, and 9). Video demonstrations of our experimental results are provided in the page. (a) Clean Table (b) Open Drawer and Place Object (c) Lifting (d) Bimanual Clean Table Figure 9. Qualitative experiments of our method in real-world tasks. (a) Spatial(pick up the black bowl between the plate and the ramekin and place it on the plate) (b) Object(pick up the tomato sauce and place it in the basket) (c) Goal (open the middle drawer of the cabinet) (d) Long (put both the alphabet soup and the tomato sauce in the basket) Figure 10. Qualitative experiments of our method in the four suites of the LIBERO benchmark. 15 VLANeXt: Recipes for Building Strong VLA Models (a) Background (pick up the black bowl between the plate and the ramekin and place it on the plate table) (b) Robot (pick up the black bowl between the plate and the ramekin and place it on the plate table) (c) Camera(pick up the black bowl between the plate and the ramekin and place it on the plate table) (d) Language(could you place the black bowl onto the plate for me? Itâs the one sitting between another plate and a ramekin) (e) Noise (pick up the black bowl between the plate and the ramekin and place it on the plate table) (f) Layout(pick up the black bowl between the plate and the ramekin and place it on the plate table) (g) Light (pick up the black bowl between the plate and the ramekin and place it on the plate table) Figure 11. Qualitative experiments of our method in the 7 types of perturbations in the same task in the LIBERO-plus benchmark. A.2. More Ablation Experiments We present more ablations of our model on the other suites (Object, Goal, and Long) on the LIBERO and LIBERO-plus benchmarks, as well as in the clean table task in real-world settings. The ablations in other suites on the LIBERO and LIBERO-plus benchmark follow the same experimental setup of our main experiments in Table 1, and the clean table task experiment also follows the real-world setup in Table 4 in the main paper. It can be seen from Table 5 and Table 6 that the results are consistent with the experiments in the Spatial suite on LIBERO and LIBERO-plus benchmarks (Table 1), which shows the robustness of our recipe. A.3. Detailed Experimental Settings We present the detailed training configuration of our final model. The same parameters are used across all four suites, which also shows the robustness of our recipe. 16 VLANeXt: Recipes for Building Strong VLA Models Table 5. Results on LIBERO and LIBERO-Plus Benchmarks. We color thebestandsecond bestresults. Freq. means the frequency- domain loss from the time-series forecasting perspective. Prop. means the input of proprioception to the VLM. Mv. means multi-view inputs. And soft means soft connection. Model LIBEROLIBERO-Plus SpatialObjectGoalLongAvgSpatialObjectGoalLongAvg VLANeXtâ freq.â prop.â mv.â soft90.031.293.683.474.653.738.236.750.944.9 VLANeXtâ freq.â prop.â mv.91.861.894.890.284.756.241.447.354.049.7 VLANeXtâ freq.â prop.97.698.896.093.896.680.574.469.068.173.0 VLANeXtâ freq. 98.099.495.695.097.087.777.971.371.477.1 VLANeXt99.099.296.694.897.493.186.576.279.783.9 Table 6. Ablation Study on Real-World Clean Table. We color thebest . ModelReal-world clean table VLANeXt w/o. action chunk1/10 VLANeXt w/o. multiview 6/10 VLANeXt w/o. soft connection 5/10 VLANeXt w/o. proprioception4/10 VLANeXt7/10 Table 7. Hyperparameters for VLANeXt training on LIBERO and LIBERO-plus benchmark, four suites. HyperparameterValue Optimization OptimizerAdamW Learning Rate1.0Ă 10 â4 Batch Size 256 Training Steps10,000 Warmup Steps500 Lr Scheduler Cosine Decay Weight Decay0.01 Max Grad Norm1.0 Backbone UpdateFull Finetune Data & Augmentation Observation Modality Image Camerate VideoMulti-view Proprioception to VLMtrue Proprioception to Policyfalse Transformer Proprioception Projector false History Proprioception Size8 Action Chunk Size8 Augmentation TypeRandom Crop, Color Jitter Crop Scale / Ratio[0.8, 1.0] / [0.9, 1.1] Color Jitter (B/C/S/H) 0.2/± 0.2/± 0.2/0.05 Architecture VLM BackboneQwen3-VL-2B-Instruct Condition TypeSoft Connection Num of Queries16 Action Loss Diffusion Loss ScheduleFlow Matching Diffusion Hidden Dim1024 Diffusion Depth29 Diffusion Heads16 Frequency Domain Loss Weight0.5 17 VLANeXt: Recipes for Building Strong VLA Models B. Revisiting Robot Learning and VLA Models Robot learning aims to apply machine learning techniques to robotic control, empowering robots to interact with the physical world and acquire diverse skills (Ravichandar et al., 2020). In the context of robot learning, tasks are generally categorized into locomotion and manipulation based on the active components being controlled. Locomotion is designed to maintain the stability and balance of the robot base, enabling mobility for legged systems such as quadrupeds or humanoids (Peng et al., 2018; Kumar et al., 2021; Lee et al., 2020; Margolis & Agrawal, 2023; Kim et al., 2024b; Liu et al., 2025). Since these tasks typically possess explicit objectives, they frequently use Reinforcement Learning (RL). In contrast, robotic manipulation focuses on controlling the robotic arm or whole body to execute a diverse array of interactive tasks. Previously, manipulation was usually focused on specific tasks, such as gripper grasping (Fang et al., 2020; Wang et al., 2021; Wu et al., 2024b; Cai et al., 2024; Wang et al., 2025a), dexterous hand grasping (Liu et al., 2023c; Li et al., 2023a; Huang et al., 2023a; Xu et al., 2024; Wei et al., 2024; 2025), dynamic grasping (Liu et al., 2023b; Chen et al., 2024), and non-prehensive manipulation (Mason, 1999; Mordatch et al., 2012; Lyu et al., 2025). These tasks can typically be formulated with well- defined task priors, making them relatively easier to learn and generalize. More recently, general-purpose manipulation has gradually become a central research direction (Brohan et al., 2023; Chi et al., 2025; Wang et al., 2024; Liu et al., 2024; Ghosh et al., 2024). However, due to the diversity of task goals and manipulated objects, it is often difficult to design explicit reward functions or leverage well-defined task priors for such settings. As a result, imitation learning (IL) has been widely adopted, enabling robots to acquire complex manipulation skills directly from expert demonstrations. In this paper, we primarily focus on general-purpose robotic manipulation through imitation learning. Robotic manipulation methods can be divided into standard action policies (Huang et al., 2022; Shridhar et al., 2023; Radosavovic et al., 2023; Huang et al., 2023b; Ze et al., 2024; Lu et al., 2024; Huang et al., 2025c; Jiang et al., 2025; Zheng et al., 2025b; Kuang et al., 2025; Zhou et al., 2025a; Zhao et al., 2025a) and video action policies (Janner et al., 2022; Wu et al., 2023; 2024a; Du et al., 2023; Cheang et al., 2024; Tian et al., 2024; Li et al., 2025c). The former framework naively inputs instructions and visuals, outputting the actions to complete the tasks, while the latter pipeline predicts future videos together with action generation, with the claim that these world modeling abilities can help understand the task better and generate the actions more accurately. In recent years, with the triumph of large foundation models, integrating these tremendous models into robot learning, specifically called Vision-Language-Action (VLA) Models, has become a prominent trend (Ma et al., 2024). This paradigm was pioneered by the groundbreaking RT-2 (Zitkovich et al., 2023), which formally introduced the concept of VLA. Subsequently, researchers across academia and industry have developed a diverse array of VLA models (OâNeill et al., 2024; Li et al., 2023b; 2024; Kim et al., 2024a; Black et al., 2024; Team et al., 2025; Hung et al., 2025; Kim et al., 2025; Shukor et al., 2025; Intelligence et al., 2025b;a; Liu et al., 2026). These newer iterations address specific challenges, such as leveraging 3D spatial information (Zhen et al., 2024; Bhat et al., 2025; Zhang et al., 2025a; Qu et al., 2025), exploiting intermediate data (like subtasks decomposition, future frame prediction or robot trajectory traces prediction) (Zheng et al., 2025a; Zhao et al., 2025b; Zhang et al., 2025c; Lv et al., 2025; Zhong et al., 2025; Liang et al., 2025; Lee et al., 2025; Cen et al., 2025b;a; Wang et al., 2025e; Zhang et al., 2025b; Song et al., 2025) to enhance action generation, and designing post-training optimization like planning or reinforcement learning to adapt to specific environment (Guo et al., 2025; Zhang et al., 2025d; Bai et al., 2025c; Tan et al., 2025; Li et al., 2025b; Fei et al., 2025a; Chen et al., 2025b; Huang et al., 2025a; Lu et al., 2025; Xiao et al., 2025a;b). Additionally, a subset of VLAs explores some niche but important aspects (Zhou et al., 2025b; Wang et al., 2025c; Kareer et al., 2025; Shi et al., 2025; Fu et al., 2025; Pertsch et al., 2025; Goyal et al., 2025; Zhang et al., 2026; Zhou et al., 2026), such as latent actions (Ye et al., 2024; Bi et al., 2025), lightweight VLAs (Wen et al., 2025; Li et al., 2025a; Reuss et al., 2025; Wang et al., 2025d) and VLAs in specific domains (Chen et al., 2025a; Bjorck et al., 2025; Ding et al., 2025; Huang et al., 2025b). Despite their different emphases, most VLA models follow a similar pipeline that builds on pretrained LLMs or VLMs to process visual observations and language instructions and produce action-relevant representations for policy learning, yet this pipeline admits many design choices spanning model interfacing, policy training, perception, and action modeling. As a result, early VLA research remains a âprimordial soupâ: rich in ideas but insufficiently structured, and the diversity of existing frameworks, together with inconsistent training and evaluation protocols, makes it difficult to identify truly impactful choices. Toward this end, this work aims to provide a more systematic understanding of this fragmented design space by comprehensively reexamining VLA design spaces under a unified framework and evaluation protocol. We conduct more than 500 distinct experiments over the above three dimensions, and distill 12 key findings that together form a practical recipe for building strong VLA models. The outcome of this study is a simple yet effective VLA model, VLANeXt, which achieves state-of-the-art performance on both LIBERO (Liu et al., 2023a) and LIBERO-plus (Fei et al., 2025b) (Fig. 1), and adapts effectively to real-world manipulation tasks. 18