OpenBMB's Trident fixes multimodal retrieval's text bias
Even small 2B-scale retrievers can double mixed-corpus retrieval quality once modality bias is addressed in training, which matters for anyone building RAG over documents with screenshots, tables or charts. The paper, code and project page are open, so teams can adopt the Multi-Positive View InfoNCE recipe or swap Trident-Qwen3VL-2B into their existing embedding stack.
OpenBMB identified a failure in multimodal retrievers: they work well on text-only or image-only corpora but degrade sharply when text and images are mixed, a phenomenon the team calls "Chaos in the Text." The root cause is a systematic Modality Preference, where text receives higher similarity scores than semantically aligned images. The team introduced Trident, which jointly trains Text, Image and Fused views with a Multi-Positive View InfoNCE objective. On 6 visual-document retrieval benchmarks, Trident-Qwen3VL-2B raised Mix Avg NDCG@10 from 40.12 to 73.90.
🔱 Why can irrelevant text outrank an image that actually contains the answer?
We find that multimodal retrievers can perform well on text-only and image-only corpora, yet degrade sharply when text and images are mixed.
We call this phenomenon "Chaos in the Text."
Our analysis reveals a systematic Modality Preference: text often receives higher similarity scores than semantically aligned images. We introduce Trident, which jointly trains Text / Image / Fused views with Multi-Positive View InfoNCE.
On 6 visual-document retrieval benchmarks, Trident-Qwen3VL-2B improves Mix Avg NDCG@10 from 40.12 → 73.90.
📄 Paper/Project: https://t.co/d9sAJY3UKe 💻 Code: https://t.co/H1L0QvkuUi
#MultimodalRetrieval #MultimodalRAG #RAG #VLM #Embedding #Trident #ChaosInTheText #OpenBMB