CARVE: Training-Free Token Compression for 3D Medical Volume Understanding
A recent preprint on arXiv presents CARVE (Cross-Slice Anisotropic Reallocation of Visual Evidence), a framework that does not require training to compress visual tokens for understanding 3D medical volumes in slice-based multimodal large language models (MLLMs). Cataloged as arXiv:2608.04515v1, this work tackles the issues related to representing 3D volumes as sequences of 2D slices, which results in an excess of visual tokens. The researchers discovered that overlapping tokens yield diminishing returns; increasing the token budget incurs higher costs while accuracy plateaus. They propose a strategy of selective budget allocation rather than simple expansion. Unlike current techniques for 2D images, CARVE focuses on reducing depth-axis redundancy, reallocating visual evidence across slices to minimize token count without sacrificing performance, thereby improving efficiency in medical imaging.
Key facts
- CARVE is a training-free framework for compressing visual tokens in 3D medical volume understanding.
- The paper is available on arXiv with identifier 2608.04515v1.
- Slice-based MLLMs represent 3D volumes as sequences of 2D slices, generating thousands of visual tokens.
- Scaling analyses on two 3D medical VQA benchmarks show diminishing returns: accuracy saturates while cost rises.
- Improving in-plane resolution is more effective than adding more slices at comparable budgets.
- Existing token compression methods are designed for 2D images or videos, not for depth-axis redundancy.
- CARVE addresses redundancy from near-duplicate content along the depth axis.
- The framework aims to allocate visual token budget more selectively.
Entities
Institutions
- arXiv