Fast-in-Slow: A Dual-System VLA Model Unifying Fast Manipulation within Slow Reasoning

Dear Reviewer x4Fe,

First of all, thanks for highlighting the clear visualization, fine-grained experiments, and performance gains of our work. We also sincerely appreciate your thoughtful comments and questions, which we address in detail below.

---

[W1 & W3.4 & Q3]. Clarification on Theoretical Alignment and System 2's Reasoning Capability.

A. FiS-VLA draws inspiration from Kahneman’s dual-system theory not to replicate its full cognitive mechanisms, but to adopt its functional abstraction. In this theory, System 1 is fast and automatic, while System 2 is slow, logical, and deliberative. This framework has been widely adopted as a functional abstraction in recent dual-system VLA models [1, 2, 3], where System 2 is implemented as a VLM. Specifically, GR00T N1 [1] treats VLM as a System 2 providing latent features for the diffusion head, while Hi Robot [2] generates low-level language commands. In both our approach and GR00T N1, System 2 goes beyond feature extraction, it performs multimodal fusion, semantic reasoning, and task understanding, supplying latent features that directly guide action generation, consistent with System 2’s deliberative role.

Moreover, unlike prior training paradigms such as GR00T N1, we introduce an dual-aware co-training strategy to preserve System 2’s reasoning. In FiS-VLA, System 2 reasons about the obervations and task instructions, further generates discretized actions explicitly.

B. Evaluation of System 2's reasoning capability. As shown in Lines 239–241, our System 2’s autoregressive supervision can leverage language-based task plans. To further assess whether preserving System 2’s reasoning improves FiS-VLA’s action generation, we replaced discrete action supervision with task plans automatically generated by Gemini 2.5 Pro and manually verified. For example, in the "phone_on_base" task from RLBench, the plan is “Step 1: Move the gripper close to the phone. Step 2: Grasp the phone. Step 3: Move it above the phone base. Step 4: Place it on the phone base.” We also use Gemini 2.5 Pro to augment plan diversity.

As shown in the table below, FiS-VLA achieves a 73% average success rate with plan-based co-training, outperforming the 69% obtained using discrete actions. This suggests that explicit reasoning supervision leads to more accurate conditioning of System 1 and improved performance.

We would like to discuss the insights behind this phenomenon. We find that preserving System 2’s reasoning supports deeper instruction comprehension and learning of richer contextual knowledge. This enables higher-quality conditioning for System 1. With shared parameters, System 1 inherits pretrained knowledge and better interprets latent features from System 2, thus improving robustness. We will integrate these findings in the revised version.

C. For the writing of "brain", it is intended as a metaphor. It does not suggest that we have constructed a complete brain using VLM. The writing of Line 43-46 will be revised as follows: "Considering these limitations, and motivated by the functional abstraction of Kahneman's dual-system framework, we raise the following question: “If a VLM model serves as the central decision-making module of the robot, can it integrate System 1 and System 2 processes to enable coordinated reasoning and execution?"

---

[W2]. Architectural Innovation.

We would like to clarify that the primary contribution of this work is proposing a new dual-system VLA paradigm. Unlike prior works [1, 3] that rely on a separate execution model as System 1, which lacks internet-scale pretraining and introduces information loss at the boundary between the two systems (as validated in W3.1 and Q1), FiS-VLA is the first to repurpose the final layers of an LLM to serve as a fast execution module (System 1) embedded within a pretrained VLM (System 2). This tightly integrated design directly addresses the bottleneck of previous works, which, in our perspective, constitutes a nontrivial and novel dual-system paradiam contribution. Building on this core design, we conduct an in-depth investigation into how to fully exploit FiS-VLA's potential for coordinated dual-system operation, focusing specifically on three key aspects:

1. The coordination of operating frequencies between the two systems, exploring how to significantly improve the model's inference speed while maintaining manipulation accuracy and stability.

2. The design of heterogeneous modality inputs for each system, which efficiently embeds point cloud as observation conditions for the execution component (System 1), enhancing the model's spatial understanding in a simple yet effective manner.

3. The co-optimization of System 1 and System 2, strengthening their cooperation and improving overall task performance.

---

[W3.1 & Q1]. Parameter Reuse Validation.

To evaluate the effectiveness of our parameter-sharing design, we conducted the following ablation studies:

• Experiment 1: The VLM serves as System 2, and the last two LLM layers are duplicated to form an independent System 1. This configuration supports both autoregressive generation from System 2's VLM and diffusion-based action prediction through the duplicated LLM layers. The entire system is trained using identical settings with FiS-VLA.
• Experiment 2: This experiment mirrors the setup of Experiment 1, with the difference being that four LLM layers are duplicated instead of two.

As shown in the table below, both variants underperform compared to FiS-VLA. We attribute this to feature misalignment. The LLM was pretrained across all 32 layers, enabling hierarchical (layer-by-layer) processing. Duplicating the last few layers breaks this chain (e.g., in Experiment 1, the new 31st layer in System 1 must process outputs from System 2’s 32nd layer), despite never being pretrained to do so. This disrupts feature compatibility and limits performance.

In contrast, FiS-VLA’s parameter-sharing keeps the LLM’s original structure and information flow intact. System 1 remains embedded within System 2, preserving pretrained knowledge while enabling action execution. This design proves more efficient and effective than separated architectures.

---

[W3.2 & Q2]. Asynchronous Training Compatibility.

Our goal is not to simply use the asynchronous training method, but rather to explore an asynchronous training strategy that is optimally aligned with the new FiS-VLA dual-system paradigm. The proposed asynchronous operation relies on the effective integration between System 2 and System 1, as System 1 must maintain high-performance execution with only low-frequency guidance from System 2.

To validate the superiority of our proposed asynchronous training approach for FiS-VLA, we applied the same asynchronous strategy to CogACT, using frequency ratios of 1:2 and 1:4. As shown in the table below, CogACT's performance showed a significant decline. We find that naively applying an asynchronous frequency strategy to a VLA model can significantly reduce the informational context for the execution module during action generation. These results highlight the importance of FiS-VLA’s unified dual-system design and our dual-aware co-training strategy, both built with asynchronous frequency mechanism in mind.

---

[W3.3]. Modal Input Necessity.

Heterogeneous modality input is a unique design of FiS-VLA. Following your suggestion, we conducted additional experiments on CogACT under matched input and frequency settings. Specifically, we adopted the 1:4 asynchronous frequency to align CogACT’s System 1 updates with System 2. We also integrated our point cloud encoding into CogACT’s System 1. As shown in the table, FiS-VLA still outperforms CogACT, highlighting the advantage of our holistic dual-system design.

---

[Q4]. RLBench Experimental Parity.

Thank you for your comments. We will include the RLBench experimental details in the revised version.

• π₀ used both open-source and over 10,000 hours of self-collected data (~10M frames) for pretraining. CogACT relied solely on open-source data (~22.5M frames), comparable to our 860K videos (~19M frames). As noted in Appendix Line 17, we followed CogACT's data processing.

• FiS-VLA, π₀, and CogACT all used full fine-tuning for downstream tasks.

• While FiS-VLA uniquely uses heterogeneous modality input, other baselines can also incorporate point cloud data. For example, CogACT's results are reported in W3.3.

---

[Q5]. FiS-VLA and π₀.₅

We consider π₀.₅ a valuable work that emphasizes the importance of data, leveraging large-scale self-collected robotic and internet-sourced data to boost generalization. However, our approach differs in both implementation objectives and strategies. π₀.₅ extends π₀’s architecture by adding an action expert (System 1) after the VLM (System 2) and using additional training data to equip the VLM with task planning capabilities. In contrast, FiS-VLA focuses on a new and unified dual-system paradigm designed to achieve more effective and lossless coordination between System 2 and System 1.

---

Paper Formatting Concerns

Thank you for your suggestion. We will separate the figures and tables in the revised version.

---

[1] GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

[2] Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models

[3] Towards Synergistic, Generalized, and Efficient Dual-System for Robotic Manipulation

Thank you for your detailed reply.

After reviewing it, I still feel that there is insufficient evidence to strongly support the extensive association drawn between FiS-VLA and the Dual-process theory, particularly regarding the role of System 2 reasoning and its specific function within the FiS-VLA model.​ Accordingly, I would suggest that in addition to including the promised supplementary experiments, the writing should adopt a more conservative approach when integrating FiS-VLA with the Dual-process theory. This would enhance the rigor of the paper, help avoid potential controversies.

Regarding the supplementary experiments on asynchronous training compatibility (ATC), could you provide more in-depth analysis? For instance, why does the introduction of ATC lead to poorer performance in CogACT, and what specific differences between CogACT and FiS-VLA might account for this discrepancy?​

Additionally, in the experiments on modal input necessity, I would appreciate seeing the results for CogACT + Point cloud. This would allow for a more direct comparison of the performance gaps between FiS-VLA and CogACT in scenarios with and without point cloud input.

I would be willing to revise my evaluation positively once the remaining issues are properly addressed.

Dear Reviewer x4Fe,

First of all, we sincerely appreciate your constructive feedback on our rebuttal. We will address each of your points in detail.

---

[Point-A]. More conservative writing.

We sincerely apologize for any confusion our original writing may have caused. We will revise the descriptions regarding the connection to Dual-process theory more carefully, rather than simply following prior works. In the revised version, we will rewrite Lines 43–46 as follows:

Considering these limitations, and in order to improve VLA execution efficiency while maintaining its inherent capabilities, we raise the following question: If a VLM model serves as the central decision-making module of the robot, can it integrate System 2 and System 1 processes to enable coordinated multimodal comprehension and fast execution?

In addition, as you suggested, we will revise the usage of the term “reasoning” by providing a more precise description of System 2’s specific role within the FiS-VLA model, such as multimodal comprehension, providing latent multimodal conditions, and generating discrete action or language planning.

For instance, we will modify Lines 4–7 and 12–14 of the main paper as follows:

To mitigate this dilemma, dual-system approaches have been proposed to leverage a VLM-based System 2 module for handling high-level multimodal comprehension, and a separate System 1 action module for ensuring real-time control.

This innovative paradigm not only enables high-frequency execution in System 1, but also facilitates coordination between multimodal comprehension and execution components within a single foundation model of System2.

Meanwhile, we will revise Lines 17–20, 47–49, and 66-68 as follows:

To enable coordination between the two systems, a dual-aware co-training strategy is proposed that equips System 1 with action generation capabilities while preserving System 2’s contextual understanding to provide stable latent conditions for System 1.

To this end, we propose Fast-in-Slow (FiS), a VLA foundation model that integrates the fast execution capabilities of System 1 into a pretrained VLM, while preserving its inherent System 2 multimodal understanding and generation capabilities.

For the multimodal comprehension component (System 2), we exploit an autoregressive next-token prediction objective to maintain its discrete action generation or high-level language planning capabilities and preserve the overall coherence and integrity of System 2.

We greatly appreciate your thoughtful suggestions. In the revised version, we will not only include the promised planning experiments, but also adopt a more conservative and detailed writing to avoid all potential controversies. Indeed, we aim to propose a new dual-system VLA paradigm to the embodied intelligence community while also providing a more conservative and rigorous writing style for future work to follow.

Dear Reviewer y6xs,

First of all, thanks for describing our FiS-VLA as an elegant approach and quite reasonable. We also sincerely appreciate your thoughtful comments and questions, which we address in detail below.

---

[W1]. Additional dynamic change experiments to further evaluate our method.

Thank you for your suggestion. We have added a new set of experimental scenarios involving dynamic changes to the manipulated objects. This experiment further demonstrates our model’s response frequency and closed-loop manipulation capability. The task is "grasp wine bottles and place them on the rack," as shown in Table 2 of the main paper and Figure 9 of the Appendix. During task execution, we manually move the bottles in four directions: up, down, left, and right. The setup of this real-world experiment is consistent with that described in Section 4.3 of the main paper.

We evaluated 20 rollouts, each using a different approach to movement. The mean success rates of FiS-VLA and π₀ on this task are reported in the table below. In this task, our model continuously and smoothly adjusts the manipulation pose in response to the moving object, achieving satisfactory manipulation results. In the revised version, we will also include more specially designed tasks that better highlight the advantages of our fast-slow system design.

---

[Q1]. Why can FiS-VLA achieve a higher success rate in static tasks?

Thank you for your insightful question. Our model not only achieves high-frequency action generation capability but also contains a series of robotic-specific designs that contribute to its strong performance across diverse tasks:

1. We innovatively integrate System 1 into the VLM (System 2) and adopt a dual-aware co-training strategy to achieve deep fusion between semantic understanding and action generation. This enables the construction of more robust action representations, allowing the model to adapt to various task types, from static operations to complex manipulation tasks. For example, our method can handle both precise pick-and-place tasks and relatively complex tasks such as "wiping a blackboard" and "folding a deformable towel and placing it into a basket".

2. In addition, our proposed heterogeneous modality input helps System 1 dynamically and efficiently acquire 2D images, 3D point clouds, and robot states at each timestep. Even for static tasks, this provides richer perceptual information and improves action precision.

3. Finally, our asynchronous frequency design enables high-frequency robot control, generating smoother and more stable control trajectories, and avoiding issues such as object slippage or inaccurate positioning caused by jerky motion.

---

[Q2]. The experiment of feeding point clouds into the beginning of the VLM.

This is an important ablation study explored in our work, and we have included the experimental results in the Appendix. As shown in Section B.2 and Figure 2 of the Appendix, we investigate various combinations of heterogeneous modality inputs for the fast and slow systems. Specifically, Variant 1 feeds the point cloud directly into the beginning of the VLM model, together with the 2D image and language instruction as input. The corresponding results are shown in the table below.

The results demonstrate that using the point cloud solely as input to the slow system does not improve manipulation accuracy, while requiring additional training and inference time. In order to maintain both model accuracy and speed, we only introduce the 3D point cloud as a condition for the execution component (System 1).

Dear Reviewer y6xs,

First and foremost, we sincerely appreciate the time and effort you dedicated to reviewing our paper, as well as your valuable suggestions and recognition of our work. To further demonstrate the superiority of FiS-VLA's high-frequency policy, we have incorporated additional task featuring dynamic changes. Additionally, we provide insights into why FiS-VLA performs well in both static and dynamic tasks, along with further clarification regarding the point cloud input concerns.

Therefore, we hope our rebuttal has addressed your concerns, and would be truly grateful for your consideration in updating the score if you find our responses satisfactory. If any questions remain, we would be more than happy to discuss and provide a prompt response. Thank you again for your valuable feedback.

Paper 4688 authors

Dear Reviewer qC7A,

First of all, thanks for recognizing that our method is sound and reasonable. We also sincerely appreciate your thoughtful comments and questions, which we address in detail below.

---

[W1]. The comparison of execution efficiency in real-world scenarios.

The model inference speed is similar in both real-world and simulation environments, as we deploy the model on an NVIDIA 4090 GPU for both. However, real-world deployment introduces additional latency due to hardware and communication overhead. Under the same hardware setup (Agilex dual-arm robot, as shown in Figure 1 of the appendix) and with an action chunk size of one, the control frequency comparison between our method and the baselines is summarized in the table below.

For the detailed latency in real-world control, the baseline models take approximately 34 ms to obtain 2D observations, whereas our method requires 40 ms to acquire 2D + 3D data. Meanwhile, there are some unavoidable sources of latency, including communication between the edge device and the robot, the robot’s actuation time, and the inverse kinematics computation used to convert the end-effector pose into joint positions (depending on the control paradigm), which together contribute around 1–2 ms. The majority of the latency comes from model inference: our model takes approximately 45 ms per forward pass, whereas π₀ requires around 72 ms per pass.

To improve control frequency in real-world scenarios, we incorporate an action chunking strategy into the FiS-VLA framework, as described in Section B.1 and shown in Figure 2 of the Appendix. For example, with a chunk size of 8, the model achieves a satisfactory manipulation success rate, while the control frequency on the real robot can exceed 20 Hz. This enables real-time control even when accounting for additional latency caused by hardware and communication overhead.

Finally, due to rebuttal constraints that prohibit the inclusion of visualizations and videos, we will provide execution video demos comparing other baselines with FiS-VLA in the revised supplementary material.

---

[Q1]. Explanation of the experiments in Table 2 of main paper.

Thank you for your constructive question. First, we would like to clarify that the comparison between our 7B version of FiS-VLA and π₀ is fair, for the following reasons:

1. Both models were trained on the same amount of downstream robotic data and employed a full-parameter fine-tuning strategy, ensuring a fair training setup.
2. FiS-VLA outperforms π₀ in both performance and speed, and our comparison is conducted along these two dimensions, which demonstrates the fairness of our evaluation.
3. Due to our unique dual-system asynchronous frequency design, the actual calling frequency of the entire LLM in our framework is lower than that of π₀.

To further address your question, we conducted an additional comparison using the 2.7B Phi-2 model as the LLM backbone. To ensure fairness, FiS-VLA (2.7B) was also pretrained on the same assembled robotic datasets under identical settings. Due to the time constraints of the rebuttal period, we only trained for one epoch. As shown in the table below, the compact-sized FiS-VLA also outperforms π₀ in performance, which demonstrates the effectiveness of our approach and its generalizability to different VLM backbones.

Dear Reviewer Vt2R,

First of all, thanks for recognizing that the fast-in-slow architecture is novel and promising. We also sincerely appreciate your thoughtful comments and questions, which we address in detail below.

---

[W1]. Why is sharing VLM layers necessary?

We sincerely appreciate your suggestion, which prompted us to provide further discussion on the necessity of sharing VLM layers. In our FiS-VLA framework, we unify System 2 and System 1 within a single VLM architecture by sharing the last few Transformer blocks. To systematically validate the advantages of this parameter-sharing approach, we conducted the following ablation studies:

• Experiment 1: The VLM serves as System 2, while the last two layers of the LLM are copied to form a diffusion head after the VLM, which serves as an independent System 1. This configuration enables: (1) autoregressive generation through the VLM's outputs, and (2) diffusion-based action generation via the duplicated Transformer blocks. The entire model is jointly optimized using our proposed dual-aware co-training strategy.
• Experiment 2: The experimental setup remains largely the same as in Experiment 1, except that the two copied Transformer blocks in System 1 are replaced with four blocks.

As shown in the table below, the results demonstrate that simply replicating additional Transformer blocks yields inferior performance compared to FiS-VLA's shared-parameter design. We attribute this to the fact that the VLM is pretrained on internet-scale data using the full 32-layer Transformer architecture for forward feature propagation. When System 1 is implemented as an independent head composed of copied 31st and 32nd layers, it inherits pretrained weights but lacks the context in which those layers were originally trained. Specifically, the duplicated 31st layer in System 1 must process latent features from System 2’s 32nd layer, an input distribution it was never exposed to during pretraining. In contrast, the FiS-VLA framework shares the last few Transformer blocks as System 1, allowing for better interpretation of earlier-layer features without compromising the integrity of the pretrained VLM. This newly proposed paradigm, which embeds System 1 within System 2, enables lossless utilization of the VLM’s pretrained knowledge and facilitates more efficient model learning.

---

[W2]. An additional comparision with π₀ using 2.7B LLM.

First, following your suggestion, we adopt the 2.7B Phi-2 model as our LLM backbone to not only further validate the effectiveness of our approach, but also demonstrate its generalizability to different LLM architectures. Given the critical role of large-scale robotic pretraining in achieving robust downstream performance for VLA tasks, we conduct the same pretraining on our assembled robotic datasets. Due to the time constraints of the rebuttal period, we only trained for one epoch. As shown in the table below, the results indicate that our lightweight FiS-VLA implementation still outperforms π₀.

Second, we would like to clarify that the comparison between our 7B version of FiS-VLA and π₀ is fair. Both models underwent downstream training on the same dataset, ensuring fair training conditions. Additionally, FiS-VLA outperforms π₀ in both performance and speed, and our comparison across these two metrics aims to support the fairness of our evaluation.

---

[Q1]. The control frequency comparison in real-world deployment.

First of all, thank you for raising this highly practical and important question. In the main paper, we primarily report speed-related quantitative results in the simulator, where all comparisons are conducted fairly. During real-world deployment with an action chunk size of one, FiS-VLA achieves a control frequency of 11.6 Hz, compared to 9.2 Hz for π₀ and 7.1 Hz for CogACT.

Next, we outline the main sources of latency in real-world deployment. The baseline model takes approximately 34 ms to obtain 2D observations, whereas our method requires 40 ms to acquire 2D + 3D data. The additional overhead primarily arises from communication between the edge device (NVIDIA 4090 GPU) and the robot, the robot’s actuation time, and the inverse kinematics (IK) computation to convert the end-effector pose into joint position, together contributing around 1 to 2 ms. The dominant latency source is model inference, which takes approximately 45 ms per step.

Meanwhile, to further improve control frequency in the real world, we adopt an action chunking scheme, as shown in Section B.1 and Figure 2 of Appendix. For example, when the chunk size is set to 8, the model accuracy remains nearly unchanged, but the real-world control frequency can exceed 20 Hz. This enables real-time control even when accounting for additional latency.