# Can Mamba Replace the Transformer in VLA Action Experts?

A new paper published on arXiv today answers a question that has been quietly nagging [Vision-Language-Action model](https://humanoidintel.ai/glossary/vision-language-action-model) engineers for the past year: when you need to stretch the action horizon to make real-time robot deployment feasible, does the Transformer architecture hold up — or is there a better option? The answer from researchers Farida Mohsen, Thowayba Elkaffash, Mohammad Reza Chalak Qazani, Mohamed Mabrok, Nader Meskin, and Ali Safa is unambiguous. At an execution horizon of N=50 actions before replanning — the regime that actually maps to feasible real-time deployment — a Mamba-based selective state-space model (SSM) action expert outperforms a Transformer-based counterpart by **7.8%** on the LIBERO benchmark suite. At N=25, the margin is **3.7%**. And at per-action replanning (N=1), Mamba matches Transformer accuracy while cutting total model parameter count by **24%**.

The study targets SmolVLA, currently one of the more widely referenced reference models in the VLA space for balancing accuracy with low computational overhead. These results matter beyond academic benchmarking: they directly address the wall every VLA deployment team hits when they try to move from lab-rate inference to hardware-rate control.

---

## The Accuracy-Complexity Tradeoff That Keeps Engineers Up at Night

Every VLA team deploying on physical humanoid hardware confronts the same painful tradeoff. Running a full VLA inference per action step (N=1) gives you maximum accuracy — the model sees fresh sensory data before every motor command — but the compute latency makes it impractical for real-time control loops. The robot is effectively frozen between decisions.

The industry workaround is action chunking: generate a sequence of N actions in a single forward pass and execute them open-loop before replanning. This is the mechanism used in architectures like ACT and diffusion policy variants. Larger N reduces per-action compute cost, enabling the policy-call frequency that physical hardware demands. The problem, well-documented in the literature, is that success rates degrade as N grows: you are committing the robot to a longer trajectory without mid-course correction from the model.

This paper's core contribution is demonstrating that the architecture of the action expert — the module that converts VLA latent representations into motor commands — significantly mediates this degradation. The hypothesis: Transformer self-attention, while powerful, may not be the ideal inductive bias for modeling temporal dependencies across long action sequences. Mamba's selective SSM, which propagates a compressed hidden state rather than attending over the full token context, may be better suited to retaining task-relevant information across extended execution horizons.

The empirical results across three execution horizons (N∈{1, 25, 50}) on the LIBERO benchmark suite support this hypothesis decisively.

---

## What the LIBERO Results Actually Show

The LIBERO benchmark is the right venue for this evaluation — it is the most widely adopted standardized suite for tabletop manipulation tasks and enables meaningful apples-to-apples comparison of policy architectures. The authors evaluated both Mamba- and Transformer-based action experts within SmolVLA across all three horizon conditions.

The key finding is directional, not just magnitude: **the Mamba advantage grows with N.** At N=1, the two architectures match on mean success rate, with Mamba winning purely on parameter efficiency (24% reduction). At N=25, Mamba pulls ahead by 3.7%. At N=50 — the practically relevant operating point for real-time deployment — the gap reaches 7.8%.

This scaling pattern is the analytically important result. It suggests Mamba's selective state propagation is specifically better at preserving task-relevant context across longer open-loop execution, rather than being generically better. The Transformer's quadratic attention over action sequences is computationally expensive and may be over-attending to irrelevant temporal context at long horizons — whereas Mamba's gating mechanism selectively retains what matters.

From a deployment standpoint, a 7.8% success-rate improvement at the horizon length that actually makes real-time control feasible is not a marginal gain. In [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) tasks, the difference between ~85% and ~93% task completion can determine whether a system is commercially viable or not.

---

## Skeptical Read: What This Paper Doesn't Resolve

Before operators and AI stack developers update their architecture preferences, some important caveats deserve attention.

**Benchmark scope.** LIBERO is a tabletop manipulation benchmark. The results say nothing directly about locomotion, [whole-body control](https://humanoidintel.ai/glossary/whole-body-control), or the kind of multi-modal long-horizon tasks that define real humanoid deployments in unstructured environments. Whether the Mamba advantage persists when the action space expands to full-body DOF — including legs, torso, and arms simultaneously — is an open question this paper does not address.

**Single base model.** The study is scoped to SmolVLA specifically. SmolVLA is described by the authors as a reference model valued for its accuracy-efficiency profile, but it is one architecture. Whether Mamba action experts improve the tradeoff for larger, more capable VLA backbones (think π0-class models from [Physical Intelligence](https://humanoidintel.ai/companies/physical-intelligence)) requires separate investigation.

**Sim-to-real gap.** LIBERO evaluations are simulation-based. The [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) implications of switching the action expert architecture — particularly whether Mamba's hidden state dynamics introduce different failure modes on physical hardware — are not characterized here.

**No latency profiling.** The paper establishes parameter count reduction (24% at N=1) as a proxy for compute efficiency, but wall-clock inference latency on target deployment hardware is not reported. For teams making architecture decisions, this is the missing number.

---

## Industry Trajectory Implications

The broader significance here is architectural: this paper adds meaningful evidence to a growing literature questioning whether Transformer attention is the right primitive for the action-generation portion of VLA stacks. The language understanding and visual grounding components of VLAs clearly benefit from attention mechanisms trained on massive internet-scale data. But the action expert is a different animal — it operates on embodiment-specific action distributions over short temporal windows, where the Mamba SSM's compression-and-selection dynamics may simply be better matched.

For humanoid platform teams currently building or evaluating VLA stacks — and given the pace of deployment across the industry, that list is long — the practical implication is worth a benchmark run: swapping the Transformer action expert for a Mamba variant costs relatively little to test and could materially improve success retention at the action horizons that actually work on hardware.

The 24% parameter reduction at equivalent accuracy (N=1) is also relevant for edge deployment scenarios where model size directly constrains what can run onboard.

---

## Key Takeaways

- Mamba-based action experts in SmolVLA outperform Transformer baselines by **7.8%** on LIBERO at N=50 action horizon — the regime corresponding to feasible real-time deployment.
- At N=25, the Mamba advantage is **3.7%**; at N=1 (per-action replanning), Mamba matches accuracy while reducing parameter count by **24%**.
- The advantage scales with execution horizon, suggesting Mamba's selective state propagation is specifically better at preserving task-relevant context across long open-loop execution windows.
- Results are scoped to SmolVLA on LIBERO (simulated tabletop tasks) — transfer to full humanoid whole-body control and physical hardware requires further validation.
- Wall-clock latency profiling on target deployment hardware is absent from the paper and represents the key gap for teams making real architecture decisions.

---

## Frequently Asked Questions

**What is the core tradeoff VLA models face in real-time robot control?**
VLA models must choose between per-action replanning (N=1), which maximizes accuracy but is too slow for real-time control, and longer action horizons (N≫1), which reduce compute load but degrade task success rates. This paper specifically targets how action expert architecture affects that degradation.

**What is Mamba and how does it differ from a Transformer in this context?**
Mamba is a selective state-space model that processes sequences by maintaining a compressed, selectively updated hidden state rather than computing full self-attention over all tokens. This makes it more computationally efficient for longer sequences and, according to this paper, better at retaining task-relevant information across extended action horizons.

**What is SmolVLA and why was it chosen for this study?**
SmolVLA is described by the authors as a popular reference model valued for its combination of high accuracy and low computational complexity. It is widely used as a baseline in VLA research, making it a natural testbed for action expert architecture comparisons.

**What benchmark was used and how reliable is it?**
The LIBERO benchmark suite is the most widely adopted standardized evaluation framework for VLA manipulation policies and enables reproducible cross-study comparisons. However, it is simulation-based, so real-hardware success rates may differ.

**Does this result mean teams should immediately switch from Transformer to Mamba action experts?**
Not without further validation. The results are compelling specifically for SmolVLA on LIBERO. Teams running larger VLA models, full humanoid embodiments, or real hardware should treat this as strong motivation to run their own ablations rather than a drop-in prescription.