Pith. sign in

REVIEW 11 cited by

Mechanistic interpretability for steering vision-language-action models

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2509.00328 v1 pith:XPATLXYN submitted 2025-08-30 cs.RO cs.LG

Mechanistic interpretability for steering vision-language-action models

classification cs.RO cs.LG
keywords modelssteeringvlascontrolmechanisticroboticsbehaviorembodied
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

Vision-Language-Action (VLA) models are a promising path to realizing generalist embodied agents that can quickly adapt to new tasks, modalities, and environments. However, methods for interpreting and steering VLAs fall far short of classical robotics pipelines, which are grounded in explicit models of kinematics, dynamics, and control. This lack of mechanistic insight is a central challenge for deploying learned policies in real-world robotics, where robustness and explainability are critical. Motivated by advances in mechanistic interpretability for large language models, we introduce the first framework for interpreting and steering VLAs via their internal representations, enabling direct intervention in model behavior at inference time. We project feedforward activations within transformer layers onto the token embedding basis, identifying sparse semantic directions - such as speed and direction - that are causally linked to action selection. Leveraging these findings, we introduce a general-purpose activation steering method that modulates behavior in real time, without fine-tuning, reward signals, or environment interaction. We evaluate this method on two recent open-source VLAs, Pi0 and OpenVLA, and demonstrate zero-shot behavioral control in simulation (LIBERO) and on a physical robot (UR5). This work demonstrates that interpretable components of embodied VLAs can be systematically harnessed for control - establishing a new paradigm for transparent and steerable foundation models in robotics.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 11 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Present but Not Remembered: Auditing How Frozen VLAs Encode, Deploy, and Steer Visual History

    cs.CV 2026-07 conditional novelty 7.0

    VLAs linearly encode past frames as near-redundant copies of the present and deploy them only as architecture-conditional fallback or standing use, so steerability tracks deployment regime not encoding.

  2. Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies

    cs.RO 2026-05 conditional novelty 7.0

    Event-grounded SAE analysis in VLA policies produces stronger causal effects on robot behavior than standard methods by anchoring features to clustered end-effector keyframes across simulations and real-robot tests.

  3. Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control

    cs.LG 2026-04 conditional novelty 7.0

    Local linearity of LLM layers enables LQR-based closed-loop activation steering with theoretical tracking guarantees.

  4. What Do They See? Interpreting Complex Road Scenarios Through the Eyes of Vision-Language-Action Models for Safe and Trustworthy Autonomous Vehicle Learning

    cs.CV 2026-07 conditional novelty 6.0

    Erasing objects from front-camera images shows Alpamayo 1's trajectories depend most on large vehicles, pedestrians, and traffic lights, but attributions are seed-unstable and some effects reach the output without tou...

  5. Inference-Time Robot Behavior Steering through Physically-Aware Reconfiguration of Task-Structure

    cs.RO 2026-06 unverdicted novelty 6.0

    ReStruct steers robot policies at inference time by reconfiguring task structure with neural automata and synchronous products, claiming up to 25% gains over VLA models in success and preference adherence.

  6. Learning What to Say to Your VLA: Mostly Harmless Vision Language Action Model Steering

    cs.RO 2026-06 unverdicted novelty 6.0

    A search-and-distill framework with conformalized improvement head produces a language feedback policy that boosts frozen VLA performance by 24.7% in simulation and 65% on hardware while guaranteeing harmlessness on p...

  7. Interpreting Video Representations with Spatio-Temporal Sparse Autoencoders

    cs.CV 2026-04 conditional novelty 6.0

    Spatio-temporal contrastive SAEs recover temporal coherence lost by hard TopK, improve action probes by +3.9% and retrieval by up to 2.8× R@1, and expose a monosemanticity metric artifact.

  8. Open-Loop Planning, Closed-Loop Verification: Speculative Verification for VLA

    cs.RO 2026-04 unverdicted novelty 6.0

    SV-VLA uses infrequent heavy VLA planning of action chunks plus a lightweight closed-loop verifier to achieve both efficiency and robustness in dynamic robot control.

  9. Behavior Uncloning: Distilling Mode Redirection into Policy Weights without Inference-Time Steering

    cs.RO 2026-06 unverdicted novelty 5.0

    MoRE improves robot policy success rates by 44 percentage points by distilling mode redirection into weights, matching filtered retraining performance without inference overhead.

  10. Act on What You See: Unlocking Safe Social Navigation in Vision-Language-Action Models

    cs.RO 2026-06 unverdicted novelty 5.0

    SALSA aligns social features and adds future-risk signals in VLA models to cut near-collisions by 86.4% and raise social accuracy from 53% to 93% on SCAND and real robots.

  11. Early Warning Signals for OpenVLA Failure under Visual Distribution Shift

    cs.CV 2026-06 conditional novelty 4.0

    OpenVLA layer-16 activations allow a logistic probe to predict failure within 15 steps under occlusion (AUROC 0.972) better than baselines, with some transfer to camera jitter.