5-2B models.
Many contrastive training strategies for multimodal embeddings add little value
A controlled study on the MMEB benchmark reveals that most sophisticated techniques yield marginal gains over a simple baseline, while a combination of three specific strategies delivers a 6% improvement.
Chinese Tech
Duo Li · Haojun Pan · Zuhao Yang · Xintong Wang · Chengfu Huo · Shijian Lu · +1 more
Nanyang Technological University · Alibaba Group · Universität Hamburg · Sichuan University
Research Digest··2 min read
Li et al.
Why this paper
From Alibaba Group and 3 others
In one line
Most contrastive strategies yield marginal gains; combining hard-negative mining, false-negative mitigation, and hard-batch construction works best.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§