Baidu ERNIE's HiUni Matches Base VLM With One Shared Backbone
TL;DR
- HiUni averages 75.56 across seven understanding benchmarks, close to the 75.61 of the Qwen3-VL-4B-Instruct VLM it was initialized from.
- On Qwen-Image-Bench, HiUni scores 47.98 Chinese and 50.00 English, described as the highest Chinese score among unified multimodal models.
- In controlled ablations, HiUni-Dense-4B leads understanding at 65.04 while the MoT-2B variant edges OneIG generation 0.345 to 0.314.
HiUni, from the ERNIE team at Baidu, averages 75.56 across seven understanding benchmarks, close to the 75.61 of the Qwen3-VL-4B-Instruct VLM it was initialized from, according to the arxiv paper.
The architecture leans on what the abstract calls 'extensive parameter sharing.' HiUni 'reuses the same language-aligned ViT and dense Transformer backbone for both capabilities, while confining task-specific computation to the output side,' with a Pixel Diffusion Head bolted to the output and a 'lightweight HiGate pathway' feeding fine spatial detail back in near the output stage.
On Qwen-Image-Bench, HiUni's Chinese score of 47.98 is 'the highest Chinese score among UMMs,' the paper reports. The English score is 50.00.
A controlled comparison of shared-dense versus task-routed backbones cuts less cleanly than the single headline suggests. HiUni-Dense-4B leads on understanding average (65.04 against the MoT-2B's 59.23), while the MoT variant edges it on OneIG generation (0.345 against 0.314). The authors' reading: 'a larger shared dense backbone can benefit both capabilities.'
This is our tenth Baidu alert of the last 90 days.
Originally reported by arxiv.org
Read the original article →Original headline: Baidu ERNIE's HiUni Shows a Shared ViT Backbone Can Serve Understanding and Generation at Once