LetTokenBurnDataVista: Diagnosing Multimodal LLMs on Data Video Understanding
A benchmark with 961 real-world data videos and 6,775 multiple-choice questions evaluating MLLMs on data perception, temporal reasoning, and narrative understanding.
$ npm i -g lettokenburn-cli$ ltb bench info datavista
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini 3.1 Pro | 70.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5 | 67.4 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Gemini 3 Flash Preview | 66.3 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Claude Sonnet 4.6 | 63.6 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3-VL-32B | 57.3 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-4o | 57.1 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen2.5-VL-32B | 54.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| InternVL3.5-38B | 54.1 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| InternVL3.5-14B | 47.4 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 46.2 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 46.1 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| InternVL3-8B | 45.1 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 43.9 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
DataVista comprises 961 professionally produced data videos from five domains (Economy, Society, Science, Politics, Culture) paired with 6,775 multiple-choice questions. It uses a three-level framework assessing data perception, temporal reasoning, and narrative understanding.
Scoring is based on accuracy across 10 fine-grained question types. Evaluation of 19 MLLMs shows Gemini-3.1-Pro achieves 70.0% accuracy, significantly below human expert performance, with notable weaknesses in causal reasoning and narrative structure.