dongx1997/TokenSwap-Bench

TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs

Andong Hua¹, Colton Bishop², Igor Mordatch², Arian Hosseini², Jindong Gu², Aleksandra Faust², Rebecca Roelofs², Yao Qin^1,2

¹University of California, Santa Barbara    ²Google DeepMind

NeurIPS 2026

---

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).

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.

---

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) license.