DiMaS: Distribution Matching for Steering Vision-Language-Action Models

¹ISIR, Sorbonne Université, Paris, France ²Valeo.ai, Paris, France

^★Equal contribution

{pegah.khayatan, sara.meziane, jayneel.parekh}@sorbonne-universite.fr

Abstract

Flow-matching-based vision-language-action (VLA) models have emerged as powerful policies for robotic manipulation, yet a critical capability remains underexplored: fine-grained behavioral control, the ability to govern how a robot performs a task by intervening on its internal representations. Representation steering is a well-established interpretability tool for language and vision-language models, where behavioral features are typically encoded as linear directions, but we show that these classic methods fall short in VLAs. We propose DiMaS, a Distribution Matching Steering strategy tailored to flow-matching VLAs, which transports between representation distributions rather than shifting along a fixed direction, and show that it effectively controls behavior across two state-of-the-art VLAs.

DiMaS matches action expert activations between two distributions (e.g. slow vs. fast, low vs. high), then steers a VLA policy toward one of the clusters at inference time

We first extract representations from a chosen layer of the flow-matching action expert, across a number of tasks and episodes. We then learn a classifier that separates two user-defined behavior classes (e.g. high vs. low speed), and use a distribution matching algorithm to learn a correspondence between the two representation distributions.

At inference time, we check whether the current representation already lies in the desired distribution. If not, we transport it using the map learned offline. Rather than applying the full transport, we interpolate between the original layer output and the transported representation, giving smooth, controllable steering strength.

Key conclusions

Hidden representations in the action expert layers are linearly separable but not linearly steerable.
DiMaS substantially outperforms linear and prompt-based baselines on the speed and height control trade-offs.
Steering quality depends on the diversity of actions seen during training, which is why we define four different steering setups.

Steering Demonstrations

Baseline vs. DiMaS-steered rollouts on the same task and seed.

Speed steering, high → low

Baseline (unsteered)

Steered (high → low)

The end-effector moves noticeably slower after steering.

Height steering, low → high

Baseline (unsteered)

Steered (low → high)

The end-effector moves higher (larger |Δz| per step) after steering.

@misc{dimas2026, title = {{DiMaS}: Distribution Matching for Steering Vision-Language-Action Models}, author = {Khayatan, Pegah and Meziane, Sara and Parekh, Jayneel and Cord, Matthieu}, year = {2026}, note = {TODO: venue / arXiv id once available}, }