arXiv Machine Learning

Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping

The paper introduces AttWarp, a lightweight technique that uses a multimodal large language model’s cross‑modal attention to perform rectilinear warping of input images at test time. By reallocating spatial resolution toward query‑relevant regions without altering model weights or architecture, AttWarp preserves global context while making small objects and subtle relationships easier for the model to read. Experiments on five benchmarks and four MLLMs show consistent accuracy gains, improved compositional reasoning, and reduced hallucinations compared to baseline image‑manipulation methods.

arXiv Computation and Language
Oct 12

TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs

TokenSwap introduces a method to benchmark and reduce the modality gap in multimodal large language models (MLLMs) by interleaving semantically aligned images with text tokens. By transforming text-only benchmarks like MMLU into image‑interleaved versions (TokenSwap‑Bench), the authors reveal a pervasive performance drop—averaging 19.6%—across 42 MLLMs, with reasoning models showing smaller gaps. Training with TokenSwap mitigates this gap while maintaining strong text‑only and vision‑language performance.

By Andong Hua, Colton Bishop, Igor Mordatch, Arian Hosseini, Jindong Gu, Aleksandra Faust, Rebecca Roelofs, Yao Qin
arXiv Computer Vision
Oct 5

Gaze Attention: Query-Adaptive Visual Routing for Efficient Multimodal LLMs

Gaze Attention is a new mechanism for multimodal large language models that selectively focuses on relevant visual regions during each generation step, rather than attending to all visual tokens. By grouping tokens into spatial regions and using learnable context tokens to retain global information, it reduces attention computation and visual key‑value entries by up to 90%. Experiments on 13 image and 6 video benchmarks show that Gaze Attention matches or outperforms dense‑attention baselines while using fewer visual resources.

By Junha Song, Byeongho Heo, Geonmo Gu, Jaegul Choo, Dongyoon Han, Sangdoo Yun