Pooling-aware training cuts retrieval vectors 4x while boosting nDCG to 97.8% of unpooled
lateinteraction · x · 2026-11-11
tonywu71 highlights that token pooling already achieves extreme compression for late-interaction multimodal retrieval (ColPali, NeoMME), and pooling-aware training can push it further. Quoting tomaarsen's experiment: two ModernBERT-base models trained on the same 120k MS MARCO triplets, changing only the loss — pooling-aware training drops vectors per document from 193 to 49 while raising NanoBEIR@13 mean nDCG@10 from 49.03 to 51.23, retaining 97.8% of the unpooled score vs 94.6% with ordinary training.
More from Multimodal
- Fable 5.2 impresses with pixel art gaussian splats, posts go viral — Flomerboy · 2026-11-12
- Seedream 5.5 nails a consistent character sticker set at just $0.02 per image — azed_ai · 2026-11-12
- Ant Group's open-source Ming-3.0-tts-0.5B tops Open TTS Leaderboard with 0.88 WER — NielsRogge · 2026-11-12
- Viggle Remake recreated in ComfyUI: swap any character in your videos with MiniMax H3 — TheLocalLab · 2026-11-12
- Pika Integrates Seedream 5.5: Split Images Into 100+ Layers for Precise Multi-Turn Edits — aziz4ai · 2026-11-12
- Llama app adds simultaneous voice, image and video input, running locally on LFM3-Omni — mervenoyann · 2026-11-12