VGenST-Bench - Direction Estimation: leaderboard

VGenST-Bench (2026) tests spatio-temporal reasoning of multimodal LLMs on short videos synthesized from validated scene graphs with text-to-image and image-to-video models. Direction Estimation task: after a first-person walk through a room, give the direction of the starting point as a clock position (vista scale, egocentric, static). Accuracy in percent on multiple-choice questions under circular evaluation (right under every rotation of the options), 8 sampled frames; higher is better.

Metric: Direction Estimation (vista scale, egocentric, static) accuracy (%) on the base multiple-choice questions of VGenST-Bench under circular evaluation (correct only if right under every cyclic permutation of the options), macro-averaged over the task's QA types, 8 uniformly sampled frames from about 100 synthesized videos per task (generated with text-to-image and image-to-video models from validated scene graphs); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.476.4#138
2Gemini 3 Flash73.4#180
3Kimi K2.673.1#197
4Gemini 3.1 Flash Lite66.9#377
5GPT-5.4 Mini60#409
6GPT-5.4 Nano59.4#653
7Gemma 4 31B (IT)59.3#354
8InternVL3.5-30B-A3B59#1269
9GLM-4.6V-Flash56.6#1049
10Qwen 3.5 27B54.4#380
11Gemma 4 26B A4B (IT)54.3#563
12Qwen 3.5 9B52#838
13Qwen 3.5 4B50.4#1270
14InternVL3.5-8B46.6#1618
15InternVL3.5-4B41.4#1946

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=vgenst-bench-direction-estimation · How It Works · Data refreshed daily, snapshot 2026-11-22.