Photon: Speedup Volume Understanding with Efficient Multimodal Large Language Models

TL;DR

Photon is a novel framework that efficiently processes 3D medical images with large language models by adaptively reducing tokens, leading to faster computation and state-of-the-art accuracy in clinical visual question answering.

Contribution

Photon introduces instruction-conditioned token scheduling and surrogate gradient propagation for adaptive token reduction in 3D medical imaging, improving efficiency without sacrificing accuracy.

Findings

Achieves state-of-the-art accuracy on medical VQA tasks.

Reduces computational resource usage significantly.

Accelerates training and inference processes.

Abstract

Multimodal large language models are promising for clinical visual question answering tasks, but scaling to 3D imaging is hindered by high computational costs. Prior methods often rely on 2D slices or fixed-length token compression, disrupting volumetric continuity and obscuring subtle findings. We present Photon, a framework that represents 3D medical volumes with token sequences of variable length. Photon introduces instruction-conditioned token scheduling and surrogate gradient propagation to adaptively reduce tokens during both training and inference, which lowers computational cost while mitigating the attention dilution caused by redundant tokens. It incorporates a custom backpropagation rule with gradient restoration to enable differentiable optimization despite discrete token drop. To stabilize token compression and ensure reliable use of visual evidence, Photon further applies regularization objectives that mitigate language-only bias and improve reliability. Experiments on diverse medical visual question answering tasks show that Photon achieves state-of-the-art accuracy while reducing resource usage and accelerating both training and inference.