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

2026-07-15

AI Summary

Existing vision-language-action (VLA) models struggle to achieve fine-grained behavioral control, as conventional linear intervention methods are limited by the fact that while behavioral representations are linearly decodable, they are not amenable to linear manipulation. This work proposes DiMaS, a novel approach that introduces distribution matching into VLA behavior modulation for the first time. Instead of applying fixed directional shifts, DiMaS performs transport between representation distributions, thereby overcoming the limitations of linear interventions. Integrated with a flow-matching VLA architecture, representation structure analysis, and cross-task generalization evaluation, DiMaS demonstrates its effectiveness on two state-of-the-art models and systematically characterizes the transfer boundaries of behavioral control as task dissimilarity increases.

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. We further examine the generalizability of this strategy as the tasks it is learned from and evaluated on grow increasingly dissimilar, characterizing where behavioral control transfers and where it weakens. Finally, through an analysis of the representation structure of the action expert, we explain why classical linear steering falls short in the visuomotor setting: behavioral features are linearly decodable but not linearly steerable, which motivates the distribution-matching design of DiMaS. Our code is publicly available at https://github.com/pegah-kh/dimas, with additional results and videos at https://pegah-kh.github.io/dimas/

Problem

• behavioral control
• vision-language-action models
• representation steering
• fine-grained control
• visuomotor policies

Innovation

• Distribution Matching
• Representation Steering
• Vision-Language-Action Models
• Flow Matching
• Behavioral Control