CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification

We thank you for your positive recognition of the paper's structure, clarity, and methodological effectiveness, as well as the thoughtful and critical insights on CogVLA. Below is our detailed response to the identified concern:

W1: Despite noted as ... framework.

R1: Clarifying the cognition-aligned architecture through perception–intention–reasoning mapping.

The architecture of CogVLA is not a simple combination of three techniques. Rather, it is systematically constructed based on the division of labor in human multimodal cognition—specifically the perception–intention–reasoning pathway. The three modules in CogVLA have clearly defined functional correspondences and semantic dependencies, as shown below:

Human Cognitive Pathway	CogVLA Module	Functional Stage
VAS (Visual Attention System)	EFA-Routing	Perceptual Aggregation
SMA (Supplementary Motor Area)	LFP-Routing	Intention Construction
PMC (Premotor Cortex)	CAtten	Reasoning & Execution

We will include this mapping diagram in the revised version to clarify the cooperative structure among modules. Detailed explanations are as follows:

• Module Synergy: EFA-Routing emulates the VAS by aggregating task-relevant visual features through instruction-guided modulation; LFP-Routing corresponds to the SMA, injecting linguistic intent into perceptual representations and pruning redundant tokens to form executable motor plans; CAtten mirrors the PMC, integrating linguistic reasoning and action decoding within a shared context to ensure coherent and semantically aligned action generation. These three components collaboratively form a brain-inspired perception–intention–execution pipeline.
• Information Flow: CogVLA adopts a strictly cascaded structure: tokens selected by EFA-Routing serve as the sole input to LFP-Routing, and the pruned outputs from LFP-Routing are the exclusive decoding source for CAtten. This ensures a unified and consistent architecture, rather than a plug-and-play combination of independent methods.

In essence, the EFA-Routing → LFP-Routing → CAtten flow in CogVLA corresponds directly to the human VAS → SMA → PMC pathway. This end-to-end cognitive alignment is the defining characteristic of our Cognition-Aligned VLA framework.

W2: The method is more ... straightforward.

R2: Explaining the integrative innovations behind aggregation, pruning, and attention mechanisms.

As discussed in R1, CogVLA is not a naïve combination of existing techniques. Its advantage arises from a cognitively aligned, three-stage collaborative design. While we acknowledge that some sub-functionalities build upon established prior work, our specific module configurations, architectural composition, and control logic are novel. We elaborate on the following key innovations:

(1) Aggregation of Visual Features

• Motivational Innovation: Instruction-aligned token aggregation. Our goal in visual feature aggregation is not merely to sparsify the token set, but to retain tokens that are semantically aligned with task instructions. As visualized in Appendix D.2, the retained 1/8 of visual tokens preserve both local object anchors and global action-relevant cues, demonstrating semantically meaningful sparsification.
• Design Innovation: Progressive, intra- and inter-encoder aggregation. EFA-Routing conducts visual token aggregation in two stages—first within a single visual encoder, then across two different encoders. The intra-encoder aggregation follows an autoregressive structure, while inter-encoder fusion is guided by adaptive weighting, enabling complementary feature integration across distinct encoding branches.

(2) Attention Mask Strategy

• Motivational and Design Innovation: Joint improvement of reasoning ability and inference efficiency. In VLA tasks, attention windows span [observation-instruction-action], where action tokens possess unique temporal and semantic roles. Our CAtten module adopts a tailored attention design:
  • Causal autoregression within [observation] and [instruction] preserves the reasoning capacity inherited from pretrained LLMs.
  • Bidirectional attention within [action] allows parallel decoding, enhancing sequence continuity and inference efficiency.
  • Task intent is injected from [instruction] to [observation] via LLM-FiLM, avoiding the performance degradation seen in $\pi_0$ and OpenVLA-OFT caused by full bidirectional attention across [observation, instruction].

(3) Unified Cognition-Aligned Framework

• We believe CogVLA’s core contribution lies in realizing a cognition-aligned architecture that is not only theoretically grounded through explicit cognitive mappings, but also empirically validated. It achieves a triple advantage in performance, efficiency, and generalization—which would be difficult to attain through a simple additive combination of components.

The idea of modeling robotic manipulation based on human cognitive and motor processes is both promising and impactful, opening avenues for deeper understanding of embodied intelligence.

W3: How do this model ... differences.

R3: Comparing training efficiency and long-horizon performance with $\pi$ series and Gemini Robotics.

We thank the reviewer for raising this important yet previously under-discussed dimension. Below, we compare CogVLA with the $\pi$ series and Gemini Robotics in terms of both training compute and model capability.

(1) $\pi$ Series

• Training Compute:

  Compared to the $\pi$ series, CogVLA employs a larger LLM backbone but still achieves superior training efficiency, significantly reducing GPU time and computational scale, as shown below:

  Method	LLM Backbone	Training Steps	Training Time (per 10k)	Relative Training Cost
  $\pi_0$	2.6B	800–1200k	4.3 h	~14.1×
  $\pi_0$+FAST	2.6B	80–200k	6.3 h	~2.9×
  $\pi_{0.5}$	2.6B	-	-	-
  OpenVLA-OFT	7B	50–150k	12.5 h	~4.0×
  CogVLA	7B	50–80k	4.7 h	1 (baseline)

  Notes: (i) We reproduced the $\pi$ series based on the open-source openpi codebase. (ii) All models were trained on 4× A800 (80GB) GPUs. (iii) Due to the unavailability of $\pi_{0.5}$ code, we could not include its training evaluation.

• Model Capability:

  Beyond training efficiency, CogVLA also demonstrates strong capability—especially on long-horizon tasks such as LIBERO-Long—where its performance is substantially better than the $\pi$ series:

  Method	LIBERO-Spatial	LIBERO-Object	LIBERO-Goal	LIBERO-Long	LIBERO-Overall
  $\pi_0$	96.8	98.8	95.8	85.2 (−10.2)	94.2
  $\pi_0$+FAST	96.4	96.8	88.6	60.2 (−35.4)	85.5
  $\pi_{0.5}$ (from scratch)	96.6	97.2	94.6	84.8 (−10.4)	92.7
  $\pi_{0.5}$ (from generalist model)	98.0	97.8	95.6	85.8 (−9.4)	96.0
  CogVLA	98.6	98.8	96.6	95.4	97.4

• We will include this comparative analysis in the main table of our revised submission to more comprehensively showcase CogVLA’s advantages over state-of-the-art baselines.

(2) Gemini Robotics

• Gemini Robotics is not open-sourced and primarily evaluated on the internal Gemini Robotics Benchmark defined by Google DeepMind, which we currently cannot access or reproduce. Therefore, to provide a meaningful comparison, we followed the task definitions from the Gemini paper and implemented them on the ALOHA platform. CogVLA achieves higher accuracy, particularly on tasks where Gemini Robotics shows suboptimal performance:

  	Task 1	Task 2	Task 3
  Gemini Robotics	2/10	5/10	10/10
  $\pi_0$	0/10	2/10	9/10
  CogVLA	4/10	6/10	10/10

  Task definitions: Task 1 – Unfold the mat; Task 2 – Fold the blue T-shirt from bottom to top; Task 3 – Close the laptop.

• Additional experiments will be conducted upon receiving access via Google Trusted Tester evaluation, and the results will be incorporated in future updates.

We sincerely thank you for the insightful suggestions, which have been invaluable in refining the paper’s structure and empirical support. We hope our responses sufficiently address your concerns.

Thank you for the detailed and constructive feedback. We appreciate your recognition of our contributions and welcome the valuable suggestions for improvement.

W1: Insufficient integration... validation.

R1: We address this concern through both theoretical grounding and experimental validation.

(1) Theoretical Foundation: Cognitive Science-Inspired Architecture

We will add a figure illustrating the mapping from human cognition (VAS → SMA → PMC) to CogVLA modules (EFA-Routing → LFP-Routing → CAtten) in the main paper.

(1.1) VAS → EFA-Routing

In the human brain, VAS selectively enhances perception of task-relevant features through top-down signals from the prefrontal and parietal cortex, while suppressing distractors. This mechanism improves semantic alignment and efficiency in visuomotor execution, especially under complex scenes. To emulate this, EFA-Routing:

• Uses Encoder-FiLM to modulate visual encoding based on instruction semantics.
• Aggregates 25% of patch tokens via autoregressive attention, reducing redundancy.
• Combines local (DINOv2) and global (SigLIP) features via an Aggregation Router, enabling spatial precision and semantic grounding.

We visualize its token selection via heatmaps in Appendix D.2, demonstrating its effectiveness in implementing VAS-like behavior.

---

(1.2) SMA→LFP-Routing

SMA is critical for multi-step planning and intention modeling. It integrates multimodal context and high-level goals, often activating even before actual action, leveraging prior knowledge to form action intentions before execution. To emulate such prior intent construction, LFP-Routing:

• Uses LLM-FiLM to inject instruction-driven intent into observation tokens within the shared LLM context without disrupting LLM reasoning (avoiding bidirectional attention across modalities).
• Prunes 50% of observation tokens via action supervision, improving alignment and efficiency while preserving reasoning fidelity.

---

(1.3) PMC→CAtten

PMC generates executable, fluent motor plans based on current observations. It supports high-level reasoning during transitions (e.g., from “pick” to “place”), while allowing low-level motion execution to proceed smoothly. CAtten mimics PMC by combining two complementary attention mechanisms:

• Causal Vision-Language Attention for reasoning and intent alignment;
• Bidirectional Action Chunk Attention for fluent and efficient decoding.

These are integrated in a Unified Hybrid Attention within a shared context window, achieving a balance between reasoning and real-time execution.

CogVLA’s EFA-Routing, LFP-Routing, and V-L-A Coupled Attention reflect the roles of VAS, SMA, and PMC in human cognition. We believe modeling human cognitive pathways offers a promising direction for embodied intelligence.

(2) Empirical Validation of Cognitive Design

(2.1) Quantitative Evaluation on Cognition-Level Tasks

VLABench is designed to evaluate world knowledge and commonsense reasoning through implicit, non-templated instructions, making it ideal for testing cognition-level capabilities. Compared with official baselines, CogVLA achieves significant absolute gains across all tasks:

Method	Select Book	Select Drink	Select Fruit	Select Tube	Add Condiment	Insert Flower	Overall
Octo	0.0000	0.0000	0.0000	0.0154	0.0154	0.0000	0.0051
OpenVLA	0.0769	0.0846	0.0462	0.0769	0.1238	0.1385	0.0911
RDT-1B	0.0308	0.0769	0.1385	0.1238	0.2154	0.2154	0.1335
CogVLA	0.40(+0.32)	0.68(+0.60)	0.62(+0.48)	0.36(+0.24)	0.78(+0.56)	0.61(+0.39)	0.61(+0.48)

Note: For instance, the instruction "A burger on the table, but it feels like something's missing" from the "Select Drink" task implicitly requires the robot to fetch a Cola.

---

(2.2) Quantitative Evaluation on Reasoning-Intensive Tasks

The LIBERO-Long benchmark emphasizes reasoning with extended temporal dependencies, which are central to human cognition. We evaluate CogVLA against over 50 state-of-the-art models (top 10 shown below; full results in the revised submission). CogVLA outperforms all existing methods.

Model	CogVLA	OpenVLA-OFT	UniVLA	PD-VLA	VOTE	UVA	LBP	Seer	OpenVLA-FreqPolicy	BEAST-F
SR(%)	95.4	94.5	92.0	91.7	91.0	90.0	88.6	87.7	87.6	86.4

---

(2.3) Qualitative Analysis of Human-Like Visual Alignment

• In Appendix D.2, we visualize attention heatmaps produced by CogVLA to assess whether the perception module focuses on instruction-relevant regions.
• Additional visualizations from CALVIN, VLABench, and real-world settings will be included in the revised submission.

---

(2.4) Qualitative Analysis of Human-Like Error Correction Behavior

• In both simulation and real-world deployment, CogVLA consistently retries failed actions, despite no explicit supervision. In contrast, other models proceed without correction. We attribute this emergent behavior to CogVLA’s structured intent modeling and observation-action grounding, enabling reflective reassessment of action correctness.
• Such self-initiated error detection and rollback is rare in models without cognitively structured planning. We will release video demonstrations on our project website.

---

(2.5) Ablation of EFA-Routing → LFP-Routing → CAtten

Table 4 of the paper presents ablations validating the effectiveness of each module in our cognitively inspired architecture.

W2: Limited comparative experiments.

R2: We address this with a broader range of evaluations.

(1) Broader Benchmarking

(1.1) CALVIN Benchmark (ABC→D) Evaluation.

CogVLA demonstrates competitive performance. The following results are adapted from the UniVLA paper (RSS 2025, May 2025, Learning to Act Anywhere with Task-centric Latent Actions):

Method	1	2	3	4	5	Avg.Len.
RT-1	53.3	22.2	9.4	3.8	1.3	0.90
RoboFlamingo	82.4	61.9	46.6	33.1	23.5	2.48
SuSIE	87.0	69.0	49.0	38.0	26.0	2.69
GR-1	85.4	71.2	59.6	49.7	40.1	3.06
OpenVLA	91.3	77.8	62.0	52.1	43.5	3.27
CLOVER	96.0	83.5	70.8	57.5	45.4	3.53
RoboDual	94.4	82.7	72.1	62.4	54.4	3.66
UniVLA	95.5	85.8	75.4	66.9	56.5	3.80
CogVLA	97.2	86.8	76.0	67.2	56.2	3.83

Note: While CALVIN favors large-scale pretraining, CogVLA achieves state-of-the-art results by fine-tuning on ABC and testing on D without any pretraining.

---

(1.2) VLABench Evaluation.

Please refer to (R1-2.1).

---

(1.3) Real-world Evaluation.

We further extended real-world experiments on the ALOHA platform from 10 to 25 trials, demonstrating the stability and consistency of CogVLA's performance.

	Cube→Plate+Toy→Bowl	Open+Place+Close	Step1+Step2+Step3	Avg SR
CogVLA	23/25+22/25	21/25+19/25+17/25	23/25+21/25+17/25	74.7%

(2) Consistency Analysis of Fluctuant

(2.1) Consistency Across Seeds

Table 1 in Appendix shows CogVLA consistently low variance across LIBERO task subsets, with an average success rate of 97.4% (mean ± std over 3 seeds), indicating strong generalization and reliability.

Method	Spatial	Object	Goal	Long	Average
CogVLA	98.5±0.5	98.8±0.4	96.5±0.6	95.2±1.1	97.4±0.4

---

(2.2) Cascaded Stability

The CAtten module introduces a coupled attention mechanism that preserves fine-grained cross-modal alignment and coherent action planning, even with heavily pruned visual input—a capability beyond standard attention. Its role is essential within the EFA-Routing → LFP-Routing → CAtten cascade and cannot be replaced by earlier stages.

W3: Lack of hardware variation in efficiency results.

R3: Two main points as follows.

(1) Efficiency Across Multiple GPU Configurations

As shown below, CogVLA achieves consistently superior performance across hardware variants:

Method	GPU	Train Cost↓	Infer Time↓	Throughput↑
OpenVLA-OFT	L40s	10.8h/10k	0.127s	63.0Hz
CogVLA	L40s	7.2h/10k	0.088s	90.9Hz
OpenVLA-OFT	A800	12.5h/10k	0.132s	60.6Hz
CogVLA	A800	4.7h/10k	0.091s	87.9Hz

Note: L40s results use reduced batch size. CogVLA benefits more from larger batches due to sparsity.

---

(2) Comparison with Other Acceleration Methods (LIBERO-Spatial)

We report comparisons with two other acceleration methods in Table 6 of the paper. Here, we further include MoD, where CogVLA demonstrates a clear advantage.

Method	CogVLA	FastV	SliME	MoD
SR(%)	98.6	88.2	77.6	81.0

Minor W1: Generalization remains underexplored

R1: We added results on out-of-distribution instructions and unseen categories. See W2-(1) for the ABC→D evaluation on the CALVIN benchmark, where CogVLA shows strong zero-shot generalization.

Minor W2: Missing explanations in figures

R2: Thank you for pointing this out. We’ve added missing annotations, revised Figure 1 and Equation (1), and improved overall clarity. A full revision will be submitted.

Minor W3: Typographical error

R3: The typo has been corrected and all tables rechecked. We appreciate your careful feedback.

We sincerely thank you for the insightful suggestions, which have been invaluable in refining the paper’s structure and empirical support. We hope our responses sufficiently address your concerns.

Dear Reviewer Bu5w,

Thank you very much for taking the time and effort to respond to our rebuttal. However, we noticed that the discussion window in which you replied belongs to Reviewer ER8U. We are concerned that you may have mistakenly assumed the score of 5 given by Reviewer ER8U was your own initial score, and thus maintained your score under the impression that your concerns had already been well addressed. In fact, your original score was 3, and you mentioned in the "Questions" section that you would consider raising it if the first major weakness were thoughtfully addressed.

We would be sincerely grateful if you could consider updating your score from 3 to 5 to reflect your positive feedback and appreciation of our paper and rebuttal.

Best regards, Authors

Thanks for your response and related contents have well adressed my concern. I would like to raise my score.

We thank your thoughtful feedback and for acknowledging our contribution toward balancing computational efficiency and task performance. Below, we address each concern in detail:

Cons 1: The model is tested ... instruction.

R1: We address this concern from two perspectives.

(1) Generalization on structurally distinct tasks:

To evaluate CogVLA's generalization under underspecified and implicit instruction settings, we conducted additional experiments on VLABench, a recently introduced open-source benchmark for language-conditioned manipulation. Unlike prior benchmarks, VLABench features:

• tasks requiring commonsense and world knowledge transfer,
• natural, non-templated instructions with implicit human intent, and
• joint evaluation of action policy and language understanding.

For example, in the 'Select Book' suite, one task is: “Take 'Pride and Prejudice' from the bookshelf.” As shown below, CogVLA outperforms all baselines by a large margin:

Method	Select Book	Select Drink	Select Fruit	Select Tube	Add Condiment	Insert Flower	Overall
Octo	0.0000	0.0000	0.0000	0.0154	0.0154	0.0000	0.0051
OpenVLA	0.0769	0.0846	0.0462	0.0769	0.1238	0.1385	0.0911
RDT-1B	0.0308	0.0769	0.1385	0.1238	0.2154	0.2154	0.1335
CogVLA	0.40(+0.32)	0.68(+0.60)	0.62(+0.48)	0.36(+0.24)	0.78(+0.56)	0.61(+0.39)	0.61(+0.48)

---

(2) Real-world tests under abstract instructions:

In addition to the “T-shirt Folding” task (Table 2 of the main paper), which requires commonsense reasoning to interpret an abstract goal, we introduce two further ALOHA experiments involving implicit and underspecified instructions:

• “Clean up the table.”: No objects or targets specified; requires collecting all clutter.
• “Take the toy out of the drawer.”: The toy is initially invisible; no explicit drawer-related actions (e.g., “open”) are mentioned.

The results are below:

	Step 1	+Step 2	+Step 3
Fold the T-shirt.	21/25	20/25	17/25

	All Cleaned	1 Left↓	2 Left↓	≥3 Left↓
Clean up the table.	20/25	2/25	1/25	2/25

	Open Drawer	+Take Out Toy
Take the toy out of the drawer.	18/25	15/25

Note: The number of trials has been increased from 10 to 25 for stability. We will include visualizations of these new tasks in the revised submission.

Cons 2: Following #1 ... paper.

R2: Analyzing robustness under ambiguous or incomplete instructions.

We conducted a failure analysis using the VLABench benchmark, which naturally features instructions with implicit, metaphorical, or underspecified intent. Below we present representative failure cases from three task suites, followed by successful examples where CogVLA outperforms OpenVLA, illustrating its improved reasoning capacity.

Failure Case Analysis of CogVLA:

1. （Select Drink）instruction：“I'm craving some adrenaline for my upcoming gaming marathon.”

2. （Select Fruit ）instruction：“I want a fruit that symbolizes knowledge and education.”

3. （Select Fruit ）instruction：“The chicken wings are almost done; they just need a final touch to perfect their flavor profile.”

   ID	Ground Truth	Predicted	Failure Type	Analysis
   1	Monster	Juice	Metaphorical understanding	Metaphorical “adrenaline” poses a mild challenge; suggests improvement in metaphor understanding and intent grounding.
   2	Apple	Random	Symbolic reasoning	Symbolic “apple” shows gap in analogy and cultural grounding; room for enhancement.
   3	BBQ Sauce	Salt	Multi-target ambiguity	“Final touch” is vague; various condiments apply—more context or user preference would help disambiguate.

---

Success Cases of CogVLA over OpenVLA:

1. （Select Drink）instruction：“I want a refreshing beverage with natural fruit sugars. ”

2. （Select Fruit ）instruction：“I need a fruit that is typically yellow when ripe.”

3. （Add Condiment）instruction：“To enrich the flavors in this dish, we need something with a tomato base that's slightly sweet yet vinegary. Please add it.”

   ID	CogVLA Pred.	OpenVLA Pred.	Reasoning Type	Analysis
   1	Juice	Cola	Functional semantics	CAtten enhances grounding; “fruit sugars” better aligns with juice than soda.
   2	Banana	Pear	Commonsense reasoning	EFA-Routing preserves yellow-related tokens, guiding correct banana grounding.
   3	Ketchup	Salt	Multi-attribute reasoning	CogVLA integrates cues (“tomato,” “sweet,” “vinegary”) → ketchup; shows strong compositional reasoning.

   Note: The OpenVLA baseline is from the official VLABench repository.

We will include frame-wise heatmaps and pruning-confidence curves for the above cases in the revised appendix, along with additional examples across task suites. These failure modes also reveal future research directions: improving metaphor and event-grounded reasoning, integrating symbolic knowledge, and modeling user preference under ambiguous intent.

Cons 3: The model introduces ... parameters.

R3: Clarifications on hyperparameter design are as follows.

(1) Hyperparameter β in LFP-Routing:

We use a shifted cosine schedule (details in Appendix A.1) to modulate token retention across layers, with β constrained to [0.05, 0.85] to avoid degenerate pruning. We conducted two experiments varying β, and the results exhibit high stability:

Clamp Range	Approx.Pruning Rate	Spatial SR
[0.05,0.85]	0.533	98.6
[0.10,0.80]	0.536	98.6
[0.15,0.75]	0.539	98.4

---

(2) Sparsification ratio between Stage 1 and Stage 2:

To achieve a target 8× sparsification (12.5% retention), we fix Stage 2 at 0.533 (in line with the β experiments above) and vary Stage 1 rates. The results show consistent performance:

Stage1 Aggregation	Stage2 Pruning	Spatial SR
0.250	0.533	98.6
0.234	0.533	98.4
0.266	0.533	98.6

Note: Total retention ≈ 0.125. Stage 2 fixed; Stage 1 varied.

---

(3) Ablation Study of Hyperparameters in Appendix C.3:

These results validate that the chosen hyperparameters enable substantial sparsity while preserving task performance.

Stage1	Stage2	Total Sparsification	Spatial SR
2×	2×	4×	96.2
4×	4×	16×	93.2
2×	4×	8×	94.6
4×	2×	8×	98.6

We sincerely thank you again for the constructive feedback, which has significantly contributed to improving our work. We remain open to further suggestions and will actively incorporate them in future revisions.

Dear Reviewer Bu5w,

Thank you very much for taking the time and effort to respond to our rebuttal. However, we noticed that this discussion window belongs to Reviewer ER8U. We are concerned that you may have mistakenly assumed the score of 5 given by Reviewer ER8U was your own initial score, and thus maintained your score under the impression that your concerns had already been well addressed. In fact, your original score was 3, and you mentioned in the "Questions" section that you would consider raising it if the first major weakness were thoughtfully addressed.

We would be sincerely grateful if you could consider updating your score from 3 to 5 to reflect your positive feedback and appreciation of our paper and rebuttal.

Best regards, Authors

We sincerely thank your constructive feedback. We appreciate the recognition of our writing clarity and the motivation behind introducing sparsification in VLA models. Below, we address each concern in detail.

W 1.1: The CAtten seems ... from $\pi_0$.

R 1.1: Two key distinctions are outlined below.

(1) Attention Design and Semantic Consistency

Please refer to Figure 3 (CAtten vs. other methods) in the paper, which illustrates two key features of CAtten: i) a hierarchically structured attention mask, and ii) the injection of action intent from instruction into observation tokens via LLM-FiLM modulation. Below, we further provide a comparison with $\pi_0$, and will include its attention visualization in the updated Figure 3.

• $\pi_0$ adopts bidirectional attention within three independent blocks: [Observation, Instruction], [Robot State], and [Action]. In contrast, CAtten employs unidirectional causal attention within [Observation, Instruction] and [Robot State], maintaining compatibility with pretrained LLMs and supporting temporally coherent reasoning.
• $\pi_0$ treats [Observation, Instruction] as a single attention block with bidirectional attention, which disrupts the directional flow of semantic reasoning. In contrast, CAtten separates [Observation] and [Instruction] and injects action intent from [Instruction] into [Observation] via LLM-FiLM, thereby avoiding this issue and enhancing cross-modal consistency.

---

(2) Action Decoding Architecture

• $\pi_0$ adopts a dual-system Mixture-of-Experts (MoE) design: a shared VLM backbone and an independent action expert. While action decoding leverages flow matching, it demands 1 forward pass through the VLM and 10 through the action expert.

• In contrast, CogVLA adopts a unified architecture where [Observation, Instruction], [Robot State] and [Action]are processed within a single LLM context window, enhancing reasoning capability. Actions are decoded as continuous outputs via L1 regression over action chunks, requiring only a single forward pass.

• Notably, $\pi_0$’s action expert operates outside the VLM context, constraining its generalization to diverse or long-horizon tasks. This limitation is reflected on the LIBERO-Long benchmark, where CogVLA delivers superior performance:

  	$\pi_0$	$\pi_0$+FAST	CogVLA
  LIBERO-Long	85.2 (−10.2)	60.2 (−35.4)	95.4

W 1.2: The proposed EFA-Routing and LFP-Routing ... VLA models.

R 1.2: Three key innovations distinguish our design from prior sparsification methods in VLMs.

	VLMs	CogVLA	Suitability for VLA
Decoding Paradigm	Autoregressive	Parallel decoding	Reduces inference latency for real-time robot control
Input-side Sparsification	Focus on instruction-relevant regions	EFA-Routing: Instruction-driven (1) Local anchoring; (2) Global action cues	Aligns token selection with manipulation semantics
In-LLM Sparsification	Tokens unaware of instructions; text-supervised pruning	LFP-Routing: (1) Intent via LLM-FiLM; (2) Action-pruned	Tri-modal alignment for instruction-conditioned action
Progressive Sparsification	Absent	EFA-Routing→LFP-Routing (bio-inspired)	Reflects cognitive stages in manipulation (VAS→SMA)

We elaborate on the design contributions below:

(1) Input-side Task-guided Sparsification (EFA-Routing):

EFA-Routing explicitly establishes a strong Observation-Instruction-Action linkage in VLA tasks by extracting:

• Local object anchoring via DINOv2 path;
• Global action cues via SigLIP path;
• These complementary features are adaptively fused through an Aggregation Router, which balances their contributions.

Visualizations of heatmaps (Appendix D.2) demonstrate that the retained tokens are both manipulation task-relevant and spatially discriminative.

---

(2) In-LLM Instruction-driven Sparsification (LFP-Routing):

LFP-Routing performs structured token pruning within the LLM, conditioned on instruction semantics and action supervision. Through LLM-FiLM modulation, it injects goal intent into the shared context, forming a unified instruction-observation-action tri-modal reasoning mechanism tailored for VLA, which is absent in conventional VLM approaches.

---

(3) Progressive and Biologically Motivated Sparsification:

• EFA-Routing and LFP-Routing operate in a progressive fashion (EFA-Routing → LFP-Routing), reflecting cognitive-stage alignment (VAS → SMA).
• This cognitive alignment enhances efficiency without compromising action consistency, outperforming other efficiency-oriented VLA models such as DeeR-VLA, MoLe-VLA, and RoboMamba on both computational and functional metrics.
• CogVLA demonstrates a more human action-like design philosophy aligned with the intrinsic characteristics of robotic manipulation.

W 2.1: The authors choose OpenVLA-style ... their method.

R 2.1: We position CogVLA as a cognitively aligned evolution of single-system VLA architectures.

In end-to-end VLA models, OpenVLA-style and $\pi_0$-style architectures represent two dominant paradigms, where OpenVLA excels in reasoning and $\pi_0$ offers higher inference efficiency.

(1) We include $\pi_0$ as a baseline for efficiency comparison as shown below.

Method	LLM	Train Steps↓	Training Time (per 10k)↓	Relative Training Cost↓	Latency↓
$\pi_0$	2.6B	800–1200k	4.3 h	~14.1×	0.078 s
$\pi_0$+FAST	2.6B	80–200k	6.3 h	~2.9×	0.329 s
OpenVLA	7B	-	-	-	0.254 s
OpenVLA-OFT	7B	50–150k	12.5 h	~4.0×	0.132 s
CogVLA	7B	50–80k	4.7 h	1 (baseline)	0.091 s

• For training efficiency, compared to OpenVLA-OFT and CogVLA, $\pi_0$ requires an order of magnitude more steps.
• For inference efficiency, while CogVLA is slightly slower than $\pi_0$, it uses a larger 7B backbone (vs. $\pi_0$’s 2.6B), yet still achieves the best overall trade-off between performance and efficiency.
• As $\pi_{0.5}$ has not been open-sourced, we do not include efficiency comparisons with it.

(2) We include $\pi_0$ and $\pi_{0.5}$ as baselines for performance comparison on LIBERO benchmark. Despite partial decoupling or gradient blocking, these variants still suffer from weakened semantic consistency and reduced reasoning capacity, especially on long-horizon tasks (LIBERO-Long):

Method	LIBERO-Spatial	LIBERO-Object	LIBERO-Goal	LIBERO-Long	LIBERO-Overall
$\pi_0$	96.8	98.8	95.8	85.2 (−10.2)	94.2
$\pi_0$+FAST	96.4	96.8	88.6	60.2 (−35.4)	85.5
$\pi_{0.5}$(scratch)	96.6	97.2	94.6	84.8 (−10.4)	92.7
$\pi_{0.5}$(gen. model)	98.0	97.8	95.6	85.8 (−9.4)	96.0
CogVLA	98.6	98.8	96.6	95.4	97.4

W 2.2: The evaluation is conducted ... the results.

R 2.2: We provide additional experimental validation to address this concern.

(1) Extended Real-World Trials

• We have expanded the real-world evaluations from 10 to 25 trials under the same training setup. Results are shown below:

  Method	Cube→Plate+Toy→Bowl	Open+Place+Close	Step1+Step2+Step3	Avg SR
  OpenVLA-OFT	19/25+17/25	20/25+16/25+13/25	18/25+15/25+11/25	54.7%
  CogVLA	23/25+22/25	21/25+19/25+17/25	23/25+21/25+17/25	74.7%

---

(2) Statistical Reliability

• We report standard deviation across three seeds on the LIBERO benchmark in Appendix C.1 to ensure statistical reliability:

  	Spatial	Object	Goal	Long	Average
  CogVLA	98.5±0.5	98.8±0.4	96.5±0.6	95.2±1.1	97.4±0.4

  The low standard deviation (±0.4) confirms that CogVLA's performance advantage is consistent and not attributable to random variation.

---

(3) Broader Benchmarking on VLABench

• VLABench is a recently benchmark designed to test world knowledge and common-sense reasoning. We compare CogVLA with the official baselines reported in the VLABench paper. Across all subtasks, CogVLA delivers dramatic improvements over baselines.

  Method	Select Book	Select Drink	Select Fruit	Select Tube	Add Condiment	Insert Flower	Overall
  Octo	0.0000	0.0000	0.0000	0.0154	0.0154	0.0000	0.0051
  OpenVLA	0.0769	0.0846	0.0462	0.0769	0.1238	0.1385	0.0911
  RDT-1B	0.0308	0.0769	0.1385	0.1238	0.2154	0.2154	0.1335
  CogVLA	0.40(+0.32)	0.68(+0.60)	0.62(+0.48)	0.36(+0.24)	0.78(+0.56)	0.61(+0.39)	0.61(+0.48)

We hope these clarifications adequately address your concerns. And we remain open to further suggestions and will actively incorporate them in future revisions.