analiza aiqbench
This study applies Sparse Autoencoders to Vision-Language-Action models to mechanistically identify interpretable, steerable features that drive generalization across robotic tasks. By distinguishing between reusable motion primitives and episode-specific memorizations, the research provides causal evidence that amplifying these specific features induces coherent behaviors while ablating them degrades performance. These findings offer a concrete pathway for debugging and controlling complex robot policies through direct manipulation of their internal representation spaces.
Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models
arXiv:2603.19183v2 Announce Type: replace Abstract: Vision-Language-Action (VLA) models have emerged as a promising approach for general-purpose robot manipulation. However, little research has mechanistically explored when and why they generalize across objects, scenes, and instructions. To probe internal representations, we train Sparse Autoencoders (SAEs) on the VLA's hidden-layer activations. SAEs learn sparse dictionaries over model activations, often revealing features that correspond to interpretable directions in the model's representation space. We identify SAE features corresponding to motion primitives and semantic concepts, including features that are general across episodes and causally steerable. We propose a metric to categorize features as general transferable primitives or episode-specific memorizations, offering a promising glimpse towards VLA generalization. We validate these findings through steering experiments on both the LIBERO simulation benchmark and on real-world DROID hardware. We find that amplifying general and semantic features induces behaviors consistent with their meanings, whereas ablating them destroys model performance. Furthermore, we demonstrate steering as a way to control behavior in unpromptable directions. Together, these results provide mechanistic evidence that VLAs can learn reusable internal features linking perception, language, and action across tasks and scenes. Our project page is located at https://drvla.github.io
Abstract:Vision-Language-Action (VLA) models have emerged as a promising approach for general-purpose robot manipulation. However, little research has mechanistically explored when and why they generalize across objects, scenes, and instructions. To probe internal representations, we train Sparse Autoencoders (SAEs) on the VLA's hidden-layer activations. SAEs learn sparse dictionaries over model activations, often revealing features that correspond to interpretable directions in the model's representation space. We identify SAE features corresponding to motion primitives and semantic concepts, including features that are general across episodes and causally steerable. We propose a metric to categorize features as general transferable primitives or episode-specific memorizations, offering a promising glimpse towards VLA generalization. We validate these findings through steering experiments on both the LIBERO simulation benchmark and on real-world DROID hardware. We find that amplifying general and semantic features induces behaviors consistent with their meanings, whereas ablating them destroys model performance. Furthermore, we demonstrate steering as a way to control behavior in unpromptable directions. Together, these results provide mechanistic evidence that VLAs can learn reusable internal features linking perception, language, and action across tasks and scenes. Our project page is located at this https URL
Robotics (cs.RO)
arXiv:2603.19183 [cs.RO]
(or arXiv:2603.19183v2 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2603.19183 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 278 cuvinte
Subiecte conexe
A Geometric View for Understanding Concept Learning and Neuron Interpretation in Sparse Autoencoders
arXiv · 8 iun. 2026
Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability
arXiv · 6 iun. 2026
Analyzing Visual Aircraft Representations with Sparse Autoencoders
arXiv · 16 iun. 2026
How Matryoshka Sparse AutoEncoders Recover Feature Hierarchies That Vanilla SAEs Lose
LessWrong · 15 iun. 2026