mlsanityData to 14 Oct 2026

Steering Robustness into World Action Models via Mechanistic Interpretability and Optimal Control

arXiv 2607.14943 ↗First posted 16 Jul 2026cs.RO cs.AI cs.LGcited by 3

On the map: World models → Embodied and robot world models

also in 1 more
Abstract (arXiv)

World Action Models (WAMs) enable semantically- and physically-informed control but are brittle under distribution shift. In this work, we use mechanistic interpretability to study how robustness-relevant perturbations are represented in WAM activation space. Comparing activations across successful and unsuccessful rollouts, we find some WAM architectures exhibit low-dimensional linear separability for robustness-critical features, while others do not. This motivates the use of contrastive activation directions for training-free WAM steering. We also show that local linearity in WAM activation dynamics enables efficient feedback steering via model-based optimal control, yielding World-Action Linear Quadratic Regulator (WA-LQR), a minimally-invasive reduced-order LQR controller. Via mechanistic evaluations, we predict strong steerability in the Cosmos-Policy and DiT4DiT models but weak steerability in LingBot-VA, consistent with steering intervention results. On Cosmos-Policy and DiT4DiT, WA-LQR generalizes contrastive directions to new tasks and improves robustness to camera, gripper, and visual-noise perturbations over unsteered and prompt steering baselines.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Embodied and robot world models, one of 11 directions in World models. Builds on 62 of its 68 references that are in the map. First posted on arXiv 16 Jul 2026.

Builds on

62 of 68 references.

Closest ancestors · 6

Same direction, newest first

  1. 2026World Action Models: The Next Frontier in Embodied AI12 May
  2. 2026Do World Action Models Generalize Better than VLAs? A Robustness Study23 Mar
  3. 2026GigaWorld-Policy: An Efficient Action-Centered World--Action Model18 Mar
  4. 2026Fast-WAM: Do World Action Models Need Test-time Future Imagination?17 Mar
  5. 2026World Action Models are Zero-shot Policies17 Feb
  6. 2026Causal World Modeling for Robot Control29 Jan

Other references · 56

  1. 2022Scalable Diffusion Models with Transformerscited by 6,054
  2. 2023RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Controlcited by 3,080
  3. 2024OpenVLA: An Open-Source Vision-Language-Action Modelcited by 2,887
  4. 2024π0: A Vision-Language-Action Flow Model for General Robot Controlcited by 2,407
  5. 2022RT-1: Robotics Transformer for Real-World Control at Scalecited by 2,061
  6. 2024Octo: An Open-Source Generalist Robot Policycited by 1,484
  7. 2023LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learningcited by 1,410
  8. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  9. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  10. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  11. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  12. 2025Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Successcited by 883
38 more
  1. 2023Open X-Embodiment: Robotic Learning Datasets and RT-X Modelscited by 862
  2. 2019Plug and Play Language Models: A Simple Approach to Controlled Text Generationcited by 799
  3. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
  4. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
  5. 2025FAST: Efficient Action Tokenization for Vision-Language-Action Modelscited by 665
  6. 2023Learning Universal Policies via Text-Guided Video Generationcited by 522
  7. 2024Evaluating Real-World Robot Manipulation Policies in Simulationcited by 495
  8. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
  9. 2025LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Modelscited by 303
  10. 2026Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planningcited by 300
  11. 2024Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representationscited by 289
  12. 2024Programming Refusal with Conditional Activation Steeringcited by 206
  13. 2024ReFT: Representation Finetuning for Language Modelscited by 197
  14. 2025Open Problems in Mechanistic Interpretabilitycited by 179
  15. 2025Interactive Post-Training for Vision-Language-Action Modelscited by 135
  16. 2024THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulationcited by 133
  17. 2025Improving Vision-Language-Action Model with Online Reinforcement Learningcited by 116
  18. 2025What Can RL Bring to VLA Generalization? An Empirical Studycited by 108
  19. 2026DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Controlcited by 76
  20. 2024Controlling Language and Diffusion Models by Transporting Activationscited by 68
  21. 2024Aligning Large Language Models with Representation Editing: A Control Perspectivecited by 64
  22. 2025Angular Steering: Behavior Control via Rotation in Activation Spacecited by 53
  23. 2024VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulationcited by 46
  24. 2026World Model for Robot Learning: A Comprehensive Surveycited by 46
  25. 2024VLATest: Testing and Evaluating Vision-Language-Action Models for Robotic Manipulationcited by 45
  26. 2023What's the Magic Word? A Control Theory of LLM Promptingcited by 41
  27. 2025Exploring the Limits of Vision-Language-Action Manipulations in Cross-task Generalizationcited by 33
  28. 2025Mechanistic interpretability for steering vision-language-action modelscited by 31
  29. 2026VP-VLA: Visual Prompting as an Interface for Vision-Language-Action Modelscited by 17
  30. 2026Observing and Controlling Features in Vision-Language-Action Modelscited by 16
  31. 2026Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Modelscited by 16
  32. 2025Activation Steering with a Feedback Controllercited by 15
  33. 2025Mechanistic Finetuning of Vision-Language-Action Models via Few-Shot Demonstrationscited by 15
  34. 2026Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Modelscited by 15
  35. 2025LinEAS: End-to-end Learning of Activation Steering with a Distributional Losscited by 14
  36. 2025Video Unlearning via Low-Rank Refusal Vectorcited by 13
  37. 2025Video2Act: A Dual-System Video Diffusion Policy with Robotic Spatio-Motional Modelingcited by 13
  38. 2026Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Controlcited by 9

50 most cited of 56 All 56 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026FBFM: A Training-Free Asynchronous Feedback Mechanism for Flow-Matching in World-Action Models Executioncited by 3
  2. 2026What do VLM-Based Vision-Language Navigation Models Rely on: Interpreting and Steering Policy Behaviorcited by 0
  3. 2026RolloutFaith: Auditing Persistent Internal Interventions in Visual World Modelcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026MARS: Systematic Multi-Level Acceleration for World Action Models
  2. 2026HarmoWAM: Harmonizing Generalizable and Precise Manipulation via Adaptive World Action Models
  3. 2026Copper-Policy: Focus on the Representation for Robust Robot Manipulation
  4. 2026HarnessWAM: Bridging Prediction and Deliberation in World Action Models
  5. 2026Enhancing Policy Learning with World-Action Model
  6. 2026Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination
  7. 2026World-Action Models for Robot Learning and Control: A Survey
  8. 2026MV-WAM: Manifold-Aware World Action Model with Value Augmentation