TokenSwap-Bench (Image-Interleaved): leaderboard
TokenSwap-Bench (2026) replaces visualizable words in 1,516 MMLU questions with matching images to measure the modality gap of multimodal LLMs. Accuracy in percent on the image-interleaved questions; higher is better.
Metric: Accuracy (%; 1,516 MMLU questions in which visualizable concepts (4.58 per question on average) are replaced by semantically aligned images inside the text, four-option multiple choice, standard prompt, answer letter extracted by regular expression). Source: arxiv.org. Saturation forecast: Estimated already saturated. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 92.6 | #102 |
| 2 | Gemini 3.1 Pro (Preview) | 91.4 | #94 |
| 3 | Claude Sonnet 4.6 | 90.6 | #161 |
| 4 | GPT-5.1 | 90.3 | #229 |
| 5 | Gemini 2.5 Pro | 88.8 | #266 |
| 6 | GPT-5.2 | 88.7 | #176 |
| 7 | Gemini 2.5 Flash | 88.5 | #453 |
| 8 | Gemini 3 Flash | 88.4 | #164 |
| 9 | GPT-5 | 85.6 | #167 |
| 10 | Claude Haiku 4.5 | 83.3 | #568 |
| 11 | Qwen 3 VL 8B (Thinking) | 80.8 | #887 |
| 12 | Qwen 3 VL 4B (Thinking) | 78 | #1296 |
| 13 | Gemini 2.5 Flash Lite | 76.6 | #958 |
| 14 | GPT-4.1 | 76.1 | #495 |
| 15 | Gemini 3.1 Flash Lite | 75.9 | #364 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tokenswap-bench-image-interleaved · How It Works · Data refreshed daily, snapshot 2026-11-13.