Act2Intention - Intention Understanding Semantic Accuracy: leaderboard
Act2Intention Bench (arXiv 2608.14132v1, Table 7): mobile GUI action trajectories, in-distribution test set, general models without task fine-tuning. Cosine similarity between embeddings of the predicted and reference intention descriptions, 0-1. Higher is better.
Metric: Embedding cosine similarity to reference intention (0-1). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.5 Pro | 0.32 | #454 |
| 2 | Claude 3.5 Sonnet | 0.29 | #362 |
| 3 | Qwen 2.5 72B | 0.27 | #512 |
| 4 | GPT-4o | 0.25 | #357 |
| 5 | Qwen 2.5 7B | 0.24 | #895 |
| 6 | Claude 3.5 Haiku | 0.24 | #571 |
| 7 | GPT-4o Mini | 0.21 | #598 |
| 8 | GPT-3.5 Turbo | 0.17 | #858 |
| 9 | Llama 3.1 8B | 0.11 | #1130 |
| 10 | Mistral 7B | 0.11 | #1382 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=act2intention-intention-understanding-semantic-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.