# Does Representation Steering Work for Robot Manipulation Policies?

Linear steering fails for flow-matching [Vision-Language-Action Models](https://humanoidintel.ai/glossary/vision-language-action-model), and a new paper from researchers Pegah Khayatan, Sara Meziane, Jayneel Parekh, and Matthieu Cord explains precisely why — and proposes a distribution-matching alternative called DiMaS that actually works. Posted to arXiv on July 17, 2026 (arXiv:2607.14280), the work targets one of the most commercially urgent problems in humanoid robotics: fine-grained behavioral control without full retraining. The core finding is deceptively simple but has significant architectural implications — behavioral features in VLA action experts are **linearly decodable but not linearly steerable**. That asymmetry explains why techniques that work in pure language models break down when you try to apply them to visuomotor policies. DiMaS addresses this by transporting between representation distributions rather than nudging activations along a fixed linear direction. The authors demonstrate the method across two state-of-the-art flow-matching VLAs and characterize how transfer degrades as the tasks used to learn the steering vectors grow increasingly dissimilar from evaluation tasks.

This matters to anyone deploying humanoid manipulation stacks: the ability to govern *how* a robot executes a task — speed, compliance, grip style — without retraining the full policy is a prerequisite for practical industrial deployment.

---

## Why Linear Steering Breaks in Visuomotor Policies

Representation steering has a well-established track record in large language models and vision-language models. The standard recipe: identify a linear direction in activation space corresponding to a behavioral feature, then shift activations along that direction at inference time. It is cheap, interpretable, and surprisingly effective for text-based systems.

The DiMaS authors show this recipe fails in the visuomotor setting, and they provide a structural explanation grounded in the architecture of flow-matching VLA action experts. The key distinction they draw is between linear *decodability* and linear *steerability*. A feature being linearly decodable means a probe can read it out from the representation — the information is there, organized in a way a linear classifier can access. Linear steerability means you can *write* to that feature by shifting along a direction in the same space. In VLA action experts, the former holds but the latter does not.

The authors argue this is not a quirk of a particular model but a structural property of how flow-matching policies encode behavior. Flow-matching VLAs learn to transport a noise distribution to an action distribution conditioned on visual and language inputs; the internal representations encode trajectory distributions, not point estimates. Shifting along a fixed linear direction in that space changes the mean of the distribution without correctly reshaping its structure, producing degraded or unstable behavior.

This finding has direct implications for the interpretability tooling the humanoid industry is starting to build. Teams at companies developing large-scale [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) policies — whether based on flow matching or diffusion — should treat linear probing results with caution when the goal is intervention rather than diagnosis.

---

## What DiMaS Actually Does

DiMaS — Distribution Matching Steering — replaces the fixed-direction shift with an optimal transport-style approach that moves the entire representation distribution from a source behavioral regime to a target one. Rather than asking "what direction encodes this behavior," it asks "what transformation maps the distribution of representations under behavior A to the distribution under behavior B."

The practical upshot is that DiMaS can modulate *how* a robot performs a task — the paper frames this as fine-grained behavioral control — without touching model weights. The steering is learned from examples and applied at inference time by intervening on internal representations.

The authors test generalizability explicitly, evaluating how performance degrades as the tasks used to learn the steering become more dissimilar to evaluation tasks. This is the right question for deployment: operators need to know whether steering learned from, say, a pick-and-place task will transfer to an assembly task, or whether they need task-specific steering data. The paper characterizes where transfer holds and where it weakens, though the source text does not specify the exact task pairs or quantitative transfer gaps — readers should consult the full paper and supplementary videos for those details.

Code is publicly available at the URL provided in the abstract, with additional results and videos hosted separately.

---

## Why This Matters for Humanoid Deployment

The commercial pressure here is real. Operators deploying humanoid manipulation systems in variable environments — automotive assembly, electronics, logistics — need behavioral tuning without the cost and downtime of policy retraining. Retraining a large VLA on new behavior specifications typically requires new demonstration data, compute, and validation cycles measured in days to weeks.

Inference-time steering sidesteps that loop. If DiMaS's approach generalizes robustly, it opens a path to operator-accessible behavioral customization: a factory floor engineer adjusting grip force characteristics or motion speed profiles through a steering interface rather than a robotics ML pipeline.

The skeptical read: the paper demonstrates this on two VLAs in what the abstract describes as a research setting, and transfer weakens as task dissimilarity grows. The boundary conditions of that weakening will determine how broadly useful DiMaS is in production. "Increasingly dissimilar" tasks need precise operational definition before any deployment team treats this as a solved problem. The authors deserve credit for characterizing the failure mode rather than papering over it — that kind of honest transfer analysis is rarer than it should be in the VLA literature.

From a broader trajectory standpoint, DiMaS is part of a larger movement toward mechanistic interpretability and controllability for [Physical AI](https://humanoidintel.ai/glossary/physical-ai) systems. As flow-matching VLAs mature into production-grade manipulation policies, the ability to inspect and steer their internal representations — safely, without retraining — becomes an engineering requirement, not just a research curiosity.

---

## Key Takeaways

- **Linear steering fails in VLAs** because behavioral features are linearly decodable but not linearly steerable — DiMaS identifies this as a structural property, not a model-specific bug.
- **DiMaS transports between representation distributions** using a distribution-matching approach, enabling fine-grained behavioral control at inference time without weight updates.
- **Validated on two flow-matching VLAs**, with explicit analysis of how transfer degrades as source and evaluation tasks diverge.
- **The asymmetry between decodability and steerability** is the key theoretical contribution — it reframes how practitioners should interpret linear probing results for visuomotor policies.
- **Production applicability is promising but bounded** — task-dissimilarity limits on transfer need quantification before deployment teams rely on it.
- **Code and supplementary videos are publicly available**, lowering the barrier for the research community to validate and extend the approach.

---

## Frequently Asked Questions

**What is DiMaS and what problem does it solve?**
DiMaS (Distribution Matching Steering) is a method for controlling the behavior of flow-matching Vision-Language-Action models at inference time — without retraining. It solves the problem of fine-grained behavioral control, letting you govern *how* a robot performs a task by intervening on its internal representations rather than modifying model weights.

**Why doesn't standard linear steering work for VLA models?**
The DiMaS authors show that behavioral features in VLA action experts are linearly decodable (a probe can read them out) but not linearly steerable (shifting along a linear direction does not correctly reshape the representation distribution). Flow-matching policies encode trajectory distributions, not point estimates, so a fixed-direction shift degrades rather than redirects behavior.

**What types of VLA models does DiMaS apply to?**
The paper specifically targets flow-matching-based VLA models and demonstrates the approach on two state-of-the-art VLAs. Whether it extends to diffusion-based or autoregressive VLA architectures is not addressed in this paper.

**Does DiMaS transfer across different tasks?**
The authors explicitly test this: transfer holds to a degree but weakens as tasks used to learn the steering become more dissimilar to evaluation tasks. The paper characterizes where transfer breaks down, which is essential information for any deployment team considering this approach.

**How does this relate to LLM interpretability and steering research?**
Representation steering is a well-established technique in language model interpretability. DiMaS extends the paradigm to visuomotor policies, but demonstrates that the methods cannot be directly ported — the structural differences between language models and VLA action experts require a fundamentally different steering design.