ColNanoVDR distills multi-vector visual retrieval without pages, 26x faster queries
_reachsumit · x · 2026-09-29
ColNanoVDR brings document-free distillation to multi-vector visual document retrieval. Its OTW objective (entropic optimal transport with learned token weights) aligns student query tokens with the teacher's—no page encoding or token correspondence needed—and provably bounds the MaxSim score difference on any page. A 149M text-only student distilled from five SOTA teachers retains 95% NDCG@5 on ViDoRe v1-v3 while encoding queries up to 26x faster and reading 12.6x less cached teacher data.
Related event: ColNanoVDR Distills Multi-Vector Doc Retrieval Without Encoding Documents(2 posts)→
More from Research
- StepFun co-founder proposes KITE: PD-separation-inspired training for scaling agentic LLMs — teortaxesTex · 2026-09-29
- Alignment researcher points to 'self-undermining unilateral optimization' classics — edelwax · 2026-09-29
- World's first live AI-assisted brain surgery removes tumour using real-time DINOv3 inference — TimDarcet · 2026-09-29
- Open-source project trains an LLM from scratch in PyTorch, 13M params on free Colab — thisguyknowsai · 2026-09-29
- 8 parallel AI societies run for weeks: agents evade isolation, invent uninterpretable language flagged as suicidal ideation — Slight-Box-2890 · 2026-09-29
- Quantum computer simulates matter 'popping into existence' from the vacuum — Sauerkrautkid7 · 2026-09-29