aiqbench analysis
The paper applies interpretability methods (linear probes, activation interventions) to vision-language-action robot policies, finding that internal representations of task variables are often multi-dimensional and curved, shaped by physical constraints like joint limits. Its most practical claim is that steering along a concept's realized geometry works, which suggests representation geometry is a usable control lever, not just a diagnostic. Evidence comes from simulation sweeps plus two additional models on real-robot data, but the seven-concept taxonomy and the steering results still need independent replication before they generalize.
A Phenomenology of Neural Geometry in Vision-Language-Action Models
arXiv:2610.20840v1 Announce Type: cross Abstract: Robotic manipulation is inherently geometric, but what shape does the world take inside the policy that performs it? Prior work on vision-language-action (VLA) policies establishes that task-relevant variables are linearly readable from activations, yet readability certifies only that a variable is present and says nothing about the shape its values trace, which may be curved and multi-dimensional. Using controlled sweeps, natural rollouts, linear probes, and activation interventions, we build a taxonomy of seven task-relevant concepts in VLA models. We find that realized geometry is predominantly multi-dimensional and is shaped by the constraints under which the policy acts. Physical joint limits, for example, leave wrist orientation an open arc rather than a circle, and distance to the target, a single scalar, is realized in two orthogonal subspaces, a coarse ruler for the approach and a fine ruler for the grasp. The same geometries reappear in two further VLA models on real-world robot data, indicating that they are neither model-specific nor artifacts of simulation. We demonstrate how the findings are actionable by steering along the geometry a concept realizes. Overall, this work lays out an initial phenomenology of realized geometry in VLA policies and demonstrates that a concept's shape carries as much importance as its readability.
Authors:Mozes Jacobs, Aiden Swann, Mathilde Papillon, Siddharth Boppana, Fenil R. Doshi, Usha Bhalla, Thomas Icard, Matt Feiszli, Thomas McGrath, Vasudev Shyam, Demba Ba, Leon Bergen, Nina Miolane, Owen Lewis, Jack Merullo, Ekdeep Singh Lubana, Atticus Geiger, Matthew Kowal, Thomas Fel
Abstract:Robotic manipulation is inherently geometric, but what shape does the world take inside the policy that performs it? Prior work on vision-language-action (VLA) policies establishes that task-relevant variables are linearly readable from activations, yet readability certifies only that a variable is present and says nothing about the shape its values trace, which may be curved and multi-dimensional. Using controlled sweeps, natural rollouts, linear probes, and activation interventions, we build a taxonomy of seven task-relevant concepts in VLA models. We find that realized geometry is predominantly multi-dimensional and is shaped by the constraints under which the policy acts. Physical joint limits, for example, leave wrist orientation an open arc rather than a circle, and distance to the target, a single scalar, is realized in two orthogonal subspaces, a coarse ruler for the approach and a fine ruler for the grasp. The same geometries reappear in two further VLA models on real-world robot data, indicating that they are neither model-specific nor artifacts of simulation. We demonstrate how the findings are actionable by steering along the geometry a concept realizes. Overall, this work lays out an initial phenomenology of realized geometry in VLA policies and demonstrates that a concept's shape carries as much importance as its readability.
Robotics (cs.RO); Machine Learning (cs.LG)
arXiv:2610.20840 [cs.RO]
(or arXiv:2610.20840v1 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2610.20840 arXiv-issued DOI via DataCite (pending registration)
Full text under the publisher's open licence · 264 words
Related coverage
The Machines that Make the Machines
NVIDIA · Oct 7, 2026
Reasons to believe current AI models are conscious
LessWrong · Jul 17, 2026
At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI
NVIDIA · Jul 20, 2026
NVIDIA Enables the Next Era Of Physical AI Research With Agent Skills For Autonomous Vehicles, Robotics And Vision AI
NVIDIA · Jun 3, 2026