Generalist robot policies offer strong behavioral priors, but reaching deployment-level reliability often requires reinforcement learning (RL) fine-tuning. Online interaction is expensive, and the paper argues that prior methods adapting frozen flow-matching policies by treating noise space as a continuous action space are inefficient because that space is behaviorally unstructured.

AnchorQ instead introduces a structured, refinable latent action representation. At each decision step, flow reversal maps reference behaviors to state-conditioned latent anchors, then samples local particles around those anchors and applies value-guided sampling to explore promising refinements. This supports a coarse-to-fine policy improvement process.

The authors analyze trade-offs: anchors reduce identification complexity but introduce support bias, particle refinement trades that bias against action-class complexity, and critic accuracy determines the reliability of KL-regularized, value-guided selection. Across simulated and real-world manipulation tasks, AnchorQ outperformed state-of-the-art post-training baselines, delivering a 37% improvement in real-world success rate and 97% higher simulated task throughput under the same interaction budget.