mathvista

Independent re-measurement of two ZDTaichu5.0-9B model-card numbers on the full sets: CV-Bench 87.26% [85.93, 88.51] against a card of 86.82, MathVista 82.40% [79.90, 84.71] against 84.50 — and a verdict that turns on how 37 truncated generations are counted.

Vision-language pipeline for visual math QA: custom vision-to-text adapter, MPS adapter training, MathVista evaluation (Qwen2.5-VL 0.717, +20.7pp visual ablation)

Open multimodal evaluation harness for Xiaomi MiMo

Unofficial reproduction of VisRef (arXiv 2603.00207) on InternVL3.5-1B × MathVista testmini: TSR gain matches the paper, VisRef's extra gain not confirmed