EVID-Bench - Causal Inversion: leaderboard

Metric: Point-level accuracy (%) on the 25 causal inversion videos (single-source editing: events reordered to invert cause and effect), using the retrieval-augmented verification pipeline of the paper (chain-of-thought analysis, up to six YouTube search-verify-reflect rounds, 64 frames at 480p, temperature 0); each video has 3 to 5 annotated misinformation points, matched by a majority of three LLM judges; mean of three runs; higher is better. Source: arxiv.org. Saturation forecast: Around June 2028. 9 models tracked.

Top models

#ModelScore
1GPT-5.555.32
2Gemini 3 Flash54.3
3Claude Opus 4.653.19
4Gemini 3.1 Pro (Preview)52.1
5Claude Sonnet 4.650
6GPT-5.445.74
7Qwen 3.5 Plus44.7
8Qwen 3 VL 235B A22B Instruct43.6
9GPT-5.4 Mini41.49

Interactive version: theaggregate.ai/benchmark?slug=evid-bench-causal-inversion · How It Works · Data refreshed daily, snapshot 2026-09-29.