@perdactor: 1/ Meet Argus-Retriever: the first late-interaction visual doc retriever where the document representation adapts to th…

X AI KOLs Following Models

Summary

Argus-Retriever is a new late-interaction visual document retriever that adapts document representation to the query, achieving SOTA performance on ViDoRe benchmarks with a smaller index.

1/👁️ Meet Argus-Retriever: the first late-interaction visual doc retriever where the document representation adapts to the query: D(q). - 86.0 NDCG@5 on ViDoRe V1+V2 (SOTA open model) - 📦 1024-dim head, 4.5× smaller index
Original Article
View Cached Full Text

Cached at: 06/08/26, 07:18 AM

1/👁️ Meet Argus-Retriever: the first late-interaction visual doc retriever where the document representation adapts to the query: D(q).

  • 86.0 NDCG@5 on ViDoRe V1+V2 (SOTA open model)
  • 📦 1024-dim head, 4.5× smaller index

Sumit (@_reachsumit): Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval

Introduces a query-conditioned late-interaction visual document retriever.

📝 https://t.co/oB5oa4bgKc 👨🏽‍💻

1/ Meet Argus-Retriever: the first late-interaction visual doc retriever where the document representation adapts to the query: D(q).

  • 86.0 NDCG@5 on ViDoRe V1+V2 (SOTA open model)
  • 1024-dim head, 4.5× smaller index

2/ The idea: a region-aware Mixture-of-Experts inside the document encoder. A router reads each region’s content, its 2D position, and a pooled query context z_q, then mixes 4 latent experts (+1 shared). The page is now encoded differently per query → D(q). Still MaxSim.

Similar Articles

@mixedbreadai: https://x.com/mixedbreadai/status/2071678747439505816

X AI KOLs Following

Mixedbread AI introduces asymmetric quantization for late interaction retrieval, achieving 32x storage reduction with minimal quality loss by storing document vectors as binary signs while keeping query vectors high-precision, making late interaction practical for billion-scale production systems.

RAGOCR: Optical Compression of Retrieval-Augmented Text via Visual Representation

arXiv cs.CL

RAGOCR is a novel framework that compresses retrieved documents into compact visual representations conditioned on the input query, using query-aware dynamic resolution to balance compression rate and information fidelity. Experiments show it surpasses naive RAG by over 15% accuracy while using only one-eighth the input tokens.