| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| pretty_name: TokenSwap-Bench |
| size_categories: |
| - 1K<n<10K |
| task_categories: |
| - visual-question-answering |
| - multiple-choice |
| tags: |
| - multimodal |
| - modality-gap |
| - interleaved |
| - mmlu |
| configs: |
| - config_name: img |
| data_files: |
| - split: test |
| path: img/test-*.parquet |
| - config_name: text |
| data_files: |
| - split: test |
| path: text/test-*.parquet |
| --- |
| |
| <h1 align="center">TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs</h1> |
|
|
| <p align="center"><a href="https://dongxzz.github.io/website/">Andong Hua</a><sup>1</sup>, Colton Bishop<sup>2</sup>, Igor Mordatch<sup>2</sup>, Arian Hosseini<sup>2</sup>, Jindong Gu<sup>2</sup>, Aleksandra Faust<sup>2</sup>, Rebecca Roelofs<sup>2</sup>, Yao Qin<sup>1,2</sup></p> |
|
|
| <p align="center"><sup>1</sup>University of California, Santa Barbara <sup>2</sup>Google DeepMind</p> |
|
|
| <p align="center"><b>NeurIPS 2026</b></p> |
|
|
| <p align="center"> |
| <a href="https://dongxzz.github.io/TokenSwap/"><img src="https://img.shields.io/badge/Project%20Page-363636?style=for-the-badge&logo=data:image/svg%2Bxml;base64,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%2BPC9zdmc%2B&logoColor=white" alt="Project Page"></a> |
| <a href="https://arxiv.org/abs/2607.28640"><img src="https://img.shields.io/badge/Paper-363636?style=for-the-badge&logo=data:image/svg%2Bxml;base64,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%3D&logoColor=white" alt="Paper"></a> |
| <a href="https://github.com/DongXzz/TokenSwap"><img src="https://img.shields.io/badge/Code-363636?style=for-the-badge&logo=github&logoColor=white" alt="Code"></a> |
| <a href="#bibtex"><img src="https://img.shields.io/badge/BibTeX-363636?style=for-the-badge&logo=data:image/svg%2Bxml;base64,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%3D&logoColor=white" alt="BibTeX"></a> |
| </p> |
|
|
| --- |
|
|
| <p align="center"><img src="assets/tokenswap_teaser.gif" width="60%" alt="Same question as text vs. with images: the model answer flips from right to wrong"></p> |
|
|
| <p align="center"><img src="assets/fig1_overview.png" width="80%"></p> |
|
|
| Across **42 MLLMs**, we observe a pervasive modality gap, with performance decreasing by **4.2% to 47.4%** when moving from text-only to image-interleaved inputs. |
|
|
| --- |
|
|
| ## TokenSwap-Bench |
|
|
| TokenSwap operates at the **concept level**: it replaces individual textual concepts with semantically aligned natural images while preserving the surrounding context and structure. We apply TokenSwap to MMLU to construct TokenSwap-Bench, which contains **1,516 samples** with **6,946 image replacements**, averaging 4.58 replacements per sample. |
|
|
| | Config | Input | |
| |---|---| |
| | `text` | Text-only (original MMLU) | |
| | `img` | Image-interleaved | |
|
|
| The two configs are paired by `doc_id`. **Modality gap** = Acc(`text`) − Acc(`img`). |
|
|
| ```python |
| from datasets import load_dataset |
| |
| text = load_dataset("dongx1997/TokenSwap-Bench", "text", split="test") |
| img = load_dataset("dongx1997/TokenSwap-Bench", "img", split="test") |
| ``` |
|
|
| | Field | Description | |
| |---|---| |
| | `question`, `choices` | In `img`, concepts are replaced by `<image N>` placeholders (in the question and/or the choices) | |
| | `answer` | Index of the correct choice (0–3) | |
| | `subject` | MMLU subject | |
| | `doc_id` | Index in the MMLU test set | |
| | `replaced_words` | `img` only: the original concept for `<image 1>`, `<image 2>`, … | |
| | `image_1` … `image_20` | `img` only: `image_N` fills `<image N>`; unused slots are `None` | |
|
|
| For evaluation with lmms-eval, see the [code](https://github.com/DongXzz/TokenSwap). |
|
|
| --- |
|
|
| ## BibTeX |
|
|
| ```bibtex |
| @inproceedings{hua2026tokenswap, |
| title = {TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs}, |
| author = {Hua, Andong and Bishop, Colton and Mordatch, Igor and Hosseini, Arian and |
| Gu, Jindong and Faust, Aleksandra and Roelofs, Rebecca and Qin, Yao}, |
| booktitle = {NeurIPS}, |
| year = {2026} |
| } |
| ``` |
|
|
| --- |
|
|
| ## License |
|
|
| This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license. |
|
|
