mlsanityData to 7 Oct 2026

Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models

arXiv 2603.19233 ↗First posted 19 Mar 2026cs.ROcited by 14

On the map: Vision-language-action models → VLA architectures and generalization

Abstract (arXiv)

Vision-Language-Action (VLA) models combine perception, language, and motor control in a single architecture, yet how they translate multimodal inputs into actions remains poorly understood. We apply activation injection, sparse autoencoders (SAEs), and linear probes to six models spanning 80M--7B parameters across 394,000+ rollout episodes on four benchmarks. The visual pathway dominates action generation across all architectures: injecting baseline activations into null-prompt episodes recovers near-identical behavior, while cross-task injection steers robots toward source-task positions (99.8% of X-VLA episodes align with the source trajectory), exposing spatially bound motor programs tied to scene coordinates rather than abstract task representations. Language sensitivity depends on task structure, not model design: when visual context uniquely specifies the task, language is ignored; when multiple goals share a scene, language becomes essential (X-VLA textttliberogoal: 94%→10% under wrong prompts vs. textttliberoobject: 60--100% regardless). In all three multi-pathway architectures (pizhalf, SmolVLA, GR00T), expert pathways encode motor programs while VLM pathways encode goal semantics (2× greater behavioral displacement from expert injection), and subspace injection confirms these occupy separable activation subspaces. Per-token SAE processing is essential for action fidelity on most architectures, though mean-pooling improves fidelity on X-VLA. Contrastive identification recovers 82+ manipulation concepts, and causal ablation reveals sensitivity spanning 28--92% zero-effect rates independent of representation width. We release Action Atlas (https://action-atlas.com) for interactive exploration of VLA representations across all six models.

Key facts. Cited by 14 papers in the map (data through 7 Oct 2026). Placed in VLA architectures and generalization, one of 13 directions in Vision-language-action models. Builds on 59 of its 69 references that are in the map. First posted on arXiv 19 Mar 2026.

Builds on

59 of 69 references.

Closest ancestors · 2

Same direction, newest first

  1. 2026VLM4VLA: Revisiting Vision-Language-Models in Vision-Language-Action Models6 Jan
  2. 2025Emergent World Representations in OpenVLA29 Sep

Other references · 57

  1. 2022Flow Matching for Generative Modelingcited by 5,426
  2. 2023RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Controlcited by 2,993
  3. 2018UMAP: Uniform Manifold Approximation and Projection for Dimension Reductioncited by 2,841
  4. 2024OpenVLA: An Open-Source Vision-Language-Action Modelcited by 2,777
  5. 2023Diffusion Policy: Visuomotor Policy Learning via Action Diffusioncited by 2,730
  6. 2022Locating and Editing Factual Associations in GPTcited by 2,609
  7. 2024π0: A Vision-Language-Action Flow Model for General Robot Controlcited by 2,297
  8. 2023PaLM-E: An Embodied Multimodal Language Modelcited by 2,037
  9. 2022RT-1: Robotics Transformer for Real-World Control at Scalecited by 2,007
  10. 2025π0.5: a Vision-Language-Action Model with Open-World Generalizationcited by 1,653
  11. 2023Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardwarecited by 1,490
  12. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
38 more
  1. 2023LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learningcited by 1,338
  2. 2016Understanding intermediate layers using linear classifier probescited by 1,279
  3. 2022Code as Policies: Language Model Programs for Embodied Controlcited by 1,172
  4. 2019Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learningcited by 1,168
  5. 2025GR00T N1: An Open Foundation Model for Generalist Humanoid Robotscited by 1,167
  6. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
  7. 2025Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Successcited by 844
  8. 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 831
  9. 2024DROID: A Large-Scale In-The-Wild Robot Manipulation Datasetcited by 801
  10. 2015Scheduled Sampling for Sequence Prediction with Recurrent Neural Networkscited by 779
  11. 2025FAST: Efficient Action Tokenization for Vision-Language-Action Modelscited by 643
  12. 2024Scaling and evaluating sparse autoencoderscited by 530
  13. 2025SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Roboticscited by 507
  14. 2025SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Modelcited by 502
  15. 2024Evaluating Real-World Robot Manipulation Policies in Simulationcited by 484
  16. 2025CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Modelscited by 440
  17. 2024CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulationcited by 411
  18. 2024RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robotscited by 391
  19. 2025X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Modelcited by 285
  20. 2025π*0.6: a VLA That Learns From Experiencecited by 280
  21. 2025LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Modelscited by 278
  22. 2025LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorizationcited by 166
  23. 2025SimpleVLA-RL: Scaling VLA Training via Reinforcement Learningcited by 141
  24. 2025Interactive Post-Training for Vision-Language-Action Modelscited by 129
  25. 2024AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulationcited by 118
  26. 2025GR-3 Technical Reportcited by 110
  27. 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Modelscited by 70
  28. 2024ALOHA 2: An Enhanced Low-Cost Hardware for Bimanual Teleoperationcited by 61
  29. 2025Interpreting CLIP with Hierarchical Sparse Autoencoderscited by 54
  30. 2024RACER: Rich Language-Guided Failure Recovery Policies for Imitation Learningcited by 46
  31. 2026SceneSmith: Agentic Generation of Simulation-Ready Indoor Scenescited by 39
  32. 2025Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulationcited by 36
  33. 2024Steering Large Language Models using Conceptors: Improving Addition-Based Activation Engineeringcited by 33
  34. 2025Steering CLIP's vision transformer with sparse autoencoderscited by 33
  35. 2025Interpretable Steering of Large Language Models with Feature Guided Activation Additionscited by 32
  36. 2025Pixels Versus Priors: Controlling Knowledge Priors in Vision-Language Models through Visual Counterfactscited by 30
  37. 2025Mechanistic interpretability for steering vision-language-action modelscited by 30
  38. 2025dVLA: Diffusion Vision-Language-Action Model with Multimodal Chain-of-Thoughtcited by 30

50 most cited of 57 All 57 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 14.

14 papers.

Sort:

By area:

  1. 2026Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?cited by 3
  2. 2026Steering Robustness into World Action Models via Mechanistic Interpretability and Optimal Controlcited by 3
  3. 2026Sequential Planning via Anchored Robotic Keypointscited by 2
  4. 2026Event-Grounded Sparse Autoencoders for Vision-Language-Action Policiescited by 1
  5. 2026Present but Not Remembered: Auditing How Frozen VLAs Encode, Deploy, and Steer Visual Historycited by 1
  6. 2026A Causality-aware Infer-diagnose-refine Framework for Test-time Modality Adaptation in VLA Modelscited by 0
  7. 2026Grounded Semantic Re-Binding for Robust Instruction Generalization in Vision-Language-Action Modelscited by 0
  8. 2026Multimodal Model Diffing for Feature Discovery and Controlcited by 0
  9. 2026FabriMAE I Trust Myself? Self-Evaluating VLA Action Generation with Markov Attention Entropycited by 0
  10. 2026Anatomy of a Closed-Loop Collapse: A Causal Case Study of a Compressed VLA Policycited by 0
  11. 2026SCULPT-VLA: Learning Structured Control through Staged Action Groundingcited by 0
  12. 2026Gaze Prompts: Temporally Dense Human Attention for Vision-Language-Action Fine-Tuningcited by 0
2 more
  1. 2026Natural State-Prediction Accuracy can Hide Weak Controlled Responsiveness in VLA Readoutscited by 0
  2. 2026eRLT: Efficient VLA Reinforcement Learning via Action-Relevant Token Routingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026When Does Language Matter? Multilingual Instructions Reveal Step-wise Language Sensitivity in Vision-Language-Action Models
  2. 2026Vision-Language-Action in Robotics: A Survey of Datasets, Benchmarks, and Data Engines
  3. 2026Observing and Controlling Features in Vision-Language-Action Models
  4. 2026Encoded but Not in Control: Revealing the Grounding Gap in Vision-Language Robot Policies
  5. 2025An Anatomy of Vision-Language-Action Models: From Modules to Milestones and Challenges
  6. 2026Beyond English: Uncovering the Multilingual Gap in Vision-Language-Action Models
  7. 2026Do Vision-Language-Action Models Understand Instructions? A Mechanistic Interpretability Study on Language Grounding
  8. 2026Restoring Linguistic Grounding in VLA Models via Train-Free Attention Recalibration