CaST-Bench (Full Set) - Open-Ended: leaderboard

Open-ended justification quality on the full CaST-Bench set (2026) of 2,066 six-option video questions over 1,015 videos, where the model first outputs a grounded spatio-temporal causal chain and then answers: an LLM judge rates answer correctness, logical consistency, evidence coverage and overall justification on a 0-10 scale. Higher is better.

Metric: Open-ended justification score (0-10): an LLM judge rates the free-form answer and causal chain against the ground truth for answer correctness, logical consistency, evidence coverage and overall justification, all 2,066 six-option questions over 1,015 videos; every model must first output a grounded spatio-temporal causal chain (evidence segments with per-second boxes) and then answer; native video input except the GPT-5 models, which get 1 fps frames; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-54.3#178
2GPT-5 Mini3.98#364
3Gemini 2.5 Pro3.85#276
4Gemini 2.5 Flash3.83#463
5MiMo-VL-7B-RL-25082.84#823
6GLM-4.1V-9B (Thinking)2.66
7InternVL2.5-8B2.59#2370
8Qwen 3 VL 4B Instruct2.56#1332
9InternVL3.5-8B2.52#1604
10InternVL3.5-14B2.5#1504
11InternVL3.5-30B-A3B2.49#1267
12Qwen 3 VL 8B (Thinking)2.15#910
13Qwen 2.5 VL 7B Instruct1.94#1902
14Qwen 3 VL 8B Instruct1.93#1029
15LLaVA-NeXT-Video-34B1.39#4307

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=cast-bench-full-set-open-ended · How It Works · Data refreshed daily, snapshot 2026-11-21.