zengrh3/mic-error-analysis
0
MIC Error Analysis — 30 cases
Interactive viewer for 30 sampled errors of the MIC model (Ours-SFT-GRPO) on the TARABench test splits, grouped into three failure modes:
- Mode A — Perceptually subtle / locally-plausible edits (verdict miss)
- Mode B — Hallucinated visual grounding (verdict right, evidence fabricated)
- Mode C — Misidentified entity origin (right object, wrong country/era)
Open index.html for the interactive viewer.
