d1-Omni proves a strict VLM judge, rejecting bboxes accepted by Gemma and LFM2.5-VL

mervenoyann · x · 2026-11-17

The author used d1-Omni as a judge in a VLM-as-a-judge pipeline to audit bbox labels from Qwen. It rejected (p<0.5) boxes previously accepted by LFM2.5-VL and Gemma-4, and shows higher agreement with human annotations. Screenshots show traffic-sign cases with acceptance probabilities mostly below 0.5.

Original post →

More from Models

Models channel →