ActionPiece Tokenizer Hits 94.8% LIBERO on Qwen3-VL-4B

Pointwise reconstruction error hides a failure mode in VLA action tokenizers. After compression, similar actions still cluster around a representative motion, but the adjustments needed for different contexts get diminished, distorted, or even reversed.

That is the setup for ActionPiece, an arXiv paper posted this week by Shijie Lian, Bin Yu and colleagues that introduces both a tokenizer and a metric to score it. The metric is physical rank consistency, or PRC, and grades how well tokenization "preserves local physical distance rankings after reconstruction." The tokenizer stacks two objectives on top of reconstruction: physical rank preservation on encoder and quantized feature distances, and quantization regularization on codeword assignments. Output is discrete tokens that feed standard autoregressive policy training and run through a frozen decoder at execution.

Under the same Qwen3-VL-4B policy training setup, the paper reports 94.8% on LIBERO, 68.8% on unseen LIBERO-Plus, 71.9% on SimplerEnv, and 51.5% across VLA-Arena L0-L2. "Component ablations show that the two objectives jointly improve PRC and policy success," the authors write, "demonstrating the value of physical relationship supervision for action tokenization."

The abstract does not compare these numbers against a named prior tokenizer, and the arXiv page lists no author affiliations. It lands the same week as CMU's ModAR and AGIBOT's HarnessVLN on our robotics tracker.