LaViSA - Anaphora: leaderboard
LaViSA (2026) shows a vision-language model a structurally ambiguous sentence and a generated image, and asks which of two or three disambiguated readings the image depicts. This board is per-trial accuracy in percent on the anaphora category (chance 50); higher is better.
Metric: Per-trial accuracy (%) on the anaphora category; task: pick which disambiguated reading of a structurally ambiguous sentence a generated image depicts (2 or 3 readings per sentence, 100 sentences per category); chance 50; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 90 | #94 |
| 2 | GPT-5.2 | 90 | #176 |
| 3 | Qwen 3 VL 32B Instruct | 84.6 | #585 |
| 4 | Qwen 3 VL 32B (Thinking) | 84.6 | #559 |
| 5 | Gemini 3.1 Flash Lite | 82.6 | #364 |
| 6 | Qwen 3 VL 4B Instruct | 81.6 | #1279 |
| 7 | Qwen 3 VL 8B Instruct | 80.1 | #981 |
| 8 | Gemma 3 27B (IT) | 76.1 | #1289 |
| 9 | Gemma 3 12B (IT) | 75.1 | #1837 |
| 10 | Qwen 3 VL 8B (Thinking) | 74.1 | #887 |
| 11 | Qwen 3 VL 4B (Thinking) | 62.2 | #1296 |
| 12 | Gemma 3 4B (IT) | 52.7 | #3831 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=lavisa-anaphora · How It Works · Data refreshed daily, snapshot 2026-11-13.