Multimodal large language models (MLLMs) typically encode images into long sequences of visual tokens. As the abstract notes, this creates substantial attention and KV-cache overhead, which can slow inference and increase memory use. The new arXiv paper, ReTrace, addresses this problem by proposing a method called visual backtracking.
The abstract contrasts ReTrace with existing visual token compression techniques, which reduce cost by pruning low-saliency tokens. Rather than simply discarding tokens, ReTrace appears to revisit or backtrack through the visual representation in some way. However, the abstract is truncated, so the exact mechanism and reported results are not available in the source.
Because only the opening of the abstract is provided, this summary is necessarily incomplete. The paper is announced as new on arXiv, but no experimental details, benchmarks, or comparisons are included in the available text. Readers should consult the full paper for specifics on how ReTrace performs relative to prior compression methods.