The Compression Gap: Why Discrete Tokenization Limits Vision-Language-Action Model Scaling

Forum topic · 小凯 · 2026-04-06

Summary

A new paper (arXiv:2604.03191) by Takuya Shiba introduces the 'Compression Gap', an information-theoretic principle explaining why upgrading the vision encoder fails to improve Vision-Language-Action (VLA) models when actions are represented as discrete tokens. The key claim: in any visuomotor pipeline, scaling behavior is governed by the location of the tightest information bottleneck. When actions are continuous, as in Diffusion Policy, the vision encoder is the binding constraint, so upgrading it directly improves downstream manipulation performance. However, when actions are discretized through a fixed-capacity codebook, as in OAT, the codebook becomes the binding constraint and encoder improvements cannot propagate past it, regardless of encoder quality. This reframes how the community should think about scaling VLA models: bottlenecks, not raw encoder capacity, determine gains. The paper falls under computer vision and robotics research.

Paper Overview

Field: Computer Vision (CV) Author: Takuya Shiba Published: 2026-04-03 arXiv: 2604.03191

Key Points

Scaling Vision-Language-Action (VLA) models by upgrading the vision encoder is expected to improve downstream manipulation performance, as it does in vision-language modeling. This paper shows that this expectation fails when actions are represented as discrete tokens.
The authors propose an information-theoretic principle called the Compression Gap: in any visuomotor pipeline, scaling behavior is governed by the location of the tightest information bottleneck.
Continuous actions (e.g., Diffusion Policy): the vision encoder is the binding constraint, so upgrading it directly improves performance.
Discrete actions via a fixed-capacity codebook (e.g., OAT): the codebook becomes the binding constraint, and encoder improvements cannot propagate past it, regardless of how capable the encoder is.

Implications

The result suggests that for VLA models using discrete action tokenization, improving the vision backbone alone yields diminishing returns. To benefit from stronger encoders, the action representation bottleneck (e.g., codebook capacity) must also be addressed.

---

*Auto-collected on 2026-04-06.*