From VisionXLab/RISEBench datav2_total_w_subtask.json and data.zip with the prompts of PhoenixZ810/RISEBench gpt_eval.py @ aac4c6f: 100 spatial consistency prompts (curly quotes), 175 temporal/causal quality prompts (image’s), 4 spatial reasoning prompts, 4 references and 1 instruction (trailing spaces), the reference images of logical_reasoning_4 and _73 (lossy re-encodings), and the official subtask column.