Act2Answer - State: leaderboard
Act2Answer (Kachaev, Moskalenko et al., 2026) turns two-option knowledge questions into short Simpler tabletop episodes in which a vision-language-action policy answers by placing a cube on one of two image tiles; each item runs in its original and left/right-swapped layout. This category compares the state of the objects in two photos (for example, the drier grass), from MLLM-CompBench: 150 items, 300 episodes. The score is the percentage of episodes ending on the correct tile (chance 50); higher is better.
Metric: Soft success rate (%; 150 binary-choice items from MLLM-CompBench, 300 episodes; answer by placing a cube on the chosen image tile in a Simpler tabletop episode, original and left/right-swapped layouts, chance 50). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Magma-8B | 59 | |
| 2 | InternVLA-M1 | 53 | |
| 3 | SpatialVLA | 50 | |
| 4 | Xiaomi-Robotics-R0 | 50 | |
| 5 | OpenVLA | 49 | |
| 6 | π0 | 48 |
Interactive version: theaggregate.ai/benchmark?slug=act2answer-state · How It Works · Data refreshed daily, snapshot 2026-11-12.