PearlVLA: Progressive Embodied Action-Plan Refinement in Latent Space

arXiv 2026 53.3 method

TLDR

PearlVLA refines action plans in latent space using a frozen world model and RL, achieving state-of-the-art on LIBERO.

Reasoning

The paper introduces a novel latent-space deliberation method that separates visual grounding from iterative plan refinement using a frozen world model, reducing latency. However, it is only evaluated on the simulated LIBERO benchmark, lacking real-world validation.

Read-first score

Read-first score 53.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 32.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

94.5

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.94526979

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,evaluation

Topical relevance 29%

45.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 406.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

generative world model

world simulator

interactive world model

video world model

Deep Analysis

Innovations

Moving deliberation into the latent space of a vision-language model (VLM) to avoid costly explicit reasoning
Separating VLM meta-query representations into a fixed visual grounding branch and an iterative latent plan branch
Plan-conditioned world query that probes a frozen latent world model for action-free future observation latents
Future-guided RefineNet with scheduled residual updates for progressive coarse-to-fine latent action plan refinement
Causal Refinement-Grouped Process-Reward RL to optimize the latent refinement process using rewards from longer-horizon imagined futures

Methodology

PearlVLA separates VLM meta-query representations into a fixed visual grounding branch and an iterative latent plan branch. At each refinement round, a plan-conditioned world query probes a frozen latent world model for a future observation latent, which is fed back to guide plan refinement via a future-guided RefineNet that applies scheduled residual updates. After K rounds, the refined latent plan is decoded in parallel into an action chunk for low-latency execution. Training employs Causal Refinement-Grouped Process-Reward RL with rewards from longer-horizon imagined futures induced by latent plan edits.

Key Results

Empirical evaluations on the LIBERO benchmark demonstrate that PearlVLA achieves state-of-the-art performance among existing methods.

Resources

Paper