mlsanityData to 14 Oct 2026

Emergent Compositional Skills in Mixture-of-Experts VLAs

arXiv 2607.20771 ↗First posted 22 Jul 2026cs.RO cs.AI cs.LGcited by 0

On the map: Mixture-of-experts → MoE in reinforcement learning and robotics

Abstract (arXiv)

We consider the problem of learning compositional robot policies end-to-end from expert demonstrations, without any pre-specified notion of task decomposition or hierarchy. We ask whether a VLA trained with a simplified Mixture-of-Experts (MoE) action head can emergently learn to decompose tasks into reusable, interpretable primitives. We find that learned experts are heavily reused across tasks and consistently correspond to qualitatively distinct low-level behaviors, suggesting that the router implicitly learns to perform high-level sequencing while experts serve as compositional primitives. Our MoE matches the task performance of a monolithic baseline while demonstrating meaningful expert specialization, a step toward modular, interpretable robot policies that emerge from data alone.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in MoE in reinforcement learning and robotics, one of 12 directions in Mixture-of-experts. Builds on 4 of its 5 references that are in the map. First posted on arXiv 22 Jul 2026.

Builds on

4 of 5 references.

References · 4

  1. 2021Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsitycited by 3,073
  2. 2025HAMSTER: Hierarchical Action Models For Open-World Robot Manipulationcited by 130
  3. 2026RL Token: Bootstrapping Online RL with Vision-Language-Action Modelscited by 49
  4. 2025Fast-in-Slow: A Dual-System Foundation Model Unifying Fast Manipulation within Slow Reasoningcited by 31

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2019Plan Arithmetic: Compositional Plan Vectors for Multi-Task Control
  2. 2026Semantically Structured Mixture-of-Experts for Compositional Robotic Manipulation ↗
  3. 2025Expertise need not monopolize: Action-Specialized Mixture of Experts for Vision-Language-Action Learning
  4. 2022Policy Architectures for Compositional Generalization in Control
  5. 2026Abstracting Robot Manipulation Skills via Mixture-of-Experts Diffusion Policies
  6. 2025Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following
  7. 2026Same Scene, Different Task: Skill Alignment for Compositional Generalization in VLAs
  8. 2026AC-VLA: Robust Out-of-Distribution Action Execution via Compositional Learning