EDCT-Bench (Rubric-Graded): leaderboard
EDCT-Bench (2026) measures vision-language model explanation faithfulness through explanation-driven counterfactual image edits over 300 questions from OK-VQA, DriveLM and 3DSRBench. This board is the rubric-graded consistency score, an LLM-judge rating from one to five of the post-edit answer and explanation rescaled to percent, higher is better.
Metric: Counterfactual Consistency Score, rubric-graded (%; LLM-judge 1-5 faithfulness rating of the answer and explanation after verified minimal image edits rescaled to 0-1, k=2, Qwen3-235B-A22B judge, FLUX.2 Max editor). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash (Thinking) | 60 | #452 (Gemini 2.5 Flash) |
| 2 | GLM-4.6V (Reasoning) | 59 | #698 (GLM-4.6V) |
| 3 | Qwen 3 VL 8B Instruct | 57 | #982 |
| 4 | Gemma 3 27B | 55 | #1638 |
| 5 | Pixtral-12B | 52 | #2493 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=edct-bench-rubric-graded · How It Works · Data refreshed daily, snapshot 2026-11-12.