XModBench

Benchmarking Cross-Modal Capabilities and Consistency in Omni-Language Models

🎉 Accepted at ICLR 2026

What is XModBench?

XModBench is the first tri-modal (audio / vision / text) multiple-choice QA benchmark explicitly designed to measure cross-modal consistency — does an omni-language model give the same correct answer when the same semantic content is presented in different modalities?

Each item is a 4-choice question with a <context> (question stem) and four <candidates> (options). By permuting which modality carries the context vs. the candidates, every question is instantiated in six modality configurations, so no single modality is privileged.

Samples	61,320 QA pairs
Task families	5 — Perception, Spatial, Temporal, Linguistic, Knowledge
Subtasks	17
Modality configs	6 — A→T, A→V, T→A, T→V, V→A, V→T
Lite split	6,000 — balanced 5 families × 6 configs × 200
Languages	English, Chinese (speech translation)

Repository layout

RyanWW/XModBench/
├── data/            # 10 JSONL files, one per raw modality combination
│   ├── audio_text.jsonl  text_audio.jsonl  audio_image.jsonl  ...
├── data_lite/       # 6 JSONL — XModBench-Lite (a2t,a2v,t2a,t2v,v2a,v2t)
├── Data.zip         # ALL media (audio/image/video) — download + unzip → Data/
├── tasks/           # original per-subtask task definitions (JSON)
└── eval_logs/       # released per-sample model outputs (reproduced via lmms-eval)
    └── <model>/<lite|full>/  samples_*.jsonl + summary.json

Media live in Data.zip. The JSONL question files (data/, data_lite/) reference media by repo-relative paths like Data/vggss_audio_bench/xxx.wav. Download and unzip Data.zip once so those paths resolve. (Data.zip was rebuilt with Chapter-stripped emotions/ clips — a fix for a moviepy parsing crash; see Changelog.)

Loading the data

1. Get the media (one-time, ~30 GB):

huggingface-cli download RyanWW/XModBench Data.zip \
    --repo-type dataset --local-dir .
unzip Data.zip          # → ./Data/...   (matches the JSONL paths)

2. Load the questions:

from datasets import load_dataset

# one modality configuration (full set)
ds = load_dataset("RyanWW/XModBench", "audio_text", split="train")

# XModBench-Lite (balanced 6k)
lite = load_dataset("RyanWW/XModBench", "lite_a2t", split="train")

# media path for the first item (resolve against the unzipped Data/)
print(ds[0]["conditions"]["input"])   # e.g. Data/vggss_audio_bench/....wav

The lmms-eval port handles the download + path resolution automatically — no manual unzip needed there.

Sample schema

{
  "index": 1,
  "subtask": "01_perception/finegrained",
  "question": "Listen to this audio clip. Which text description best matches the sound you hear? Answer with A, B, C, or D",
  "conditions": { "modality": "Audio", "input": "Data/vggss_audio_bench/ymuNh7Cwhrs_000040.wav" },
  "options": {
    "A": { "modality": "Text", "input": "dog howling" },
    "B": { "modality": "Text", "input": "chicken clucking" },
    "C": { "modality": "Text", "input": "alligators, crocodiles hissing" },
    "D": { "modality": "Text", "input": "cuckoo bird calling" }
  },
  "correct_answer": "A",
  "category": "Animal Sounds"
}

• conditions.input / options[*].input are repo-relative media paths (Data/...) for non-text modalities, or the literal text for Text.
• correct_answer ∈ {A, B, C, D}; subtask is NN_family/subtask.

Modality configurations

Code	Context → Candidates
A→T	Audio → Text
A→V	Audio → Vision (image/video)
T→A	Text → Audio
T→V	Text → Vision
V→A	Vision → Audio
V→T	Vision → Text

data/ keeps Image and Video separate (10 files) for efficient loading; data_lite/ merges Vision = Image ∪ Video into the 6 canonical configs.

XModBench-Lite

A 6,000-sample split, balanced across 5 task families × 6 configs × 200, for fast, low-cost evaluation. It tracks full-set model rankings closely (see leaderboard) and is the recommended quick-eval target.

Evaluate with lmms-eval

XModBench is pre-integrated in XingruiWang/lmms-eval; the dataset auto-downloads on first run.

git clone https://github.com/XingruiWang/lmms-eval.git
cd lmms-eval && pip install -e ".[all]"

# XModBench-Lite, all 6 configs (resource-aware GPU profile)
./submit_lite.sh qwen2_5_omni_interleave Qwen/Qwen2.5-Omni-7B qwenomni3

# Level-2 metrics: by-config / by-family / disparity / imbalance
python lmms_eval/tasks/xmod_bench/summarize.py \
    --logs logs/xmod_bench_lite/results_qwen2_5_omni_interleave/

Per-sample model outputs we reproduced are released here under eval_logs/.

Leaderboard — XModBench-Lite (reproduced via lmms-eval)

By-config accuracy (%); Avg. is the mean over the six configs.

Model	A→T	A→V	T→A	T→V	V→A	V→T	Avg.
Qwen3-Omni-30B	71.6	52.0	62.5	67.0	55.6	83.1	65.3
Qwen2.5-Omni-7B	63.1	49.8	59.2	62.5	50.3	76.4	60.2
Baichuan-Omni-1.5	52.5	32.0	47.6	56.6	47.0	77.7	52.2
OmniVinci	62.2	—	—	—	—	78.8	—

Qwen2.5-Omni matches its full-set paper numbers within 5 points on every configuration. Full-set numbers for all 14 paper models are on the project website.

Changelog

• 2026-05: Data.zip rebuilt — the emotions/ MELD clips had MP4 Chapter metadata that crashed moviepy's parser (used by some evaluation backends). All emotion clips were re-muxed with ffmpeg -map_chapters -1 (video/audio streams untouched). Frame content is identical; only the Chapter atom was removed. No other media changed.

License

Released under the MIT License. Media are redistributed for research use; please also respect the licenses of the underlying source datasets (VGG-Sound, STARSS23, GTZAN, URMP, MELD, URBANSAS, and others).

Citation

@inproceedings{wang2026xmodbench,
  title     = {XModBench: Benchmarking Cross-Modal Capabilities and Consistency in Omni-Language Models},
  author    = {Wang, Xingrui and Liu, Jiang and Huang, Chao and Yu, Xiaodong and Wang, Ze and Sun, Ximeng and Wu, Jialian and Yuille, Alan and Barsoum, Emad and Liu, Zicheng},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2510.15148}
}