# Geometric Cross-Modal Token Selection for Latency-Constrained Multimodal Token Communication

Source: [Geometric Cross-Modal Token Selection for Latency-Constrained Multimodal Token Communication](<https://www.sota2.com/research/paper/6a72c382bc6f0f0705846e6f>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-08-06T05:00:05.119Z

Authors: Joohyuk Park, Junyong Shin, Yongjeong Oh, Jihong Park, Yo-Seb Jeon

Published: 2026-08-04T05:14:08.000Z

## Abstract

This paper proposes a geometry-based joint cross-modal token selection framework for latency-constrained multimodal token communications. To capture cross-modal token dependencies, we leverage the cross-attention mechanism to project modality-specific tokens into a shared query-key space, where the modality with the fewest tokens serves as the anchor modality and the others as non-anchor modalities. Inspired by germ-grain models, we define an angular-distance metric and construct semantic grain regions around anchor queries. Based on this geometric representation, we identify cross-modal evidence shared across multiple anchor queries in this space and develop an intersection-based token selection (IBS) strategy that prioritizes non-anchor tokens whose key are covered by multiple grain regions. We further develop an erasure-aware extension, termed robust-IBS (R-IBS), for token-wise erasure channels using an expected angular-distance formulation. In both IBS and R-IBS, the grain regions are optimized for individual queries under a latency constraint, using block coordinate descent and a low-complexity greedy algorithm. Simulations corroborate the effectiveness of IBS and R-IBS under latency-constrained and token-wise erasure channels, achieving up to 31.6% and 29.2% task accuracy gains on visual question answering (VQA) and audio-visual question answering (AVQA) tasks, respectively, over existing token selection baselines.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2608.03141>)

[PDF](<https://arxiv.org/pdf/2608.03141.pdf>)