AdaMerge: tuning-free patch compression for multi-vector visual document retrieval finds each doc's merge cliff
_reachsumit · x · 2026-09-22
A paper proposes AdaMerge, a tuning-free compression method for multi-vector visual document retrieval (VDR).
- Background: ColPali-style VDR stores hundreds of patch embeddings per document at high storage/latency cost; prior SOTA merging (PtM) requires per-dataset grid-searched cluster budgets.
- Key observation: the hierarchical-merge cosine sequence shows a sharp cliff separating redundancy from salient signal, consistently located in a narrow band across 11,000+ documents from 14 datasets—so the boundary can be detected per document.
- Method & results: AdaMerge detects each document's cliff via gap analysis and builds attention-weighted cluster centroids; on ViDoRe-V2 it significantly outperforms tuned PtM (p < 10^-4), fully plug-and-play with no training.
More from Research
- Google's ScientistTwo solves 80.4% of 107 top-venue ML problems autonomously — thisdudelikesAI · 2026-09-22
- Tencent Hunyuan's WebCraftBench tests web apps like software, matching human preference 85.3% — TencentHunyuan · 2026-09-22
- Sarah Hooker proposes $1 paper submission fee to stress-test auto research agents — MannyKayy · 2026-09-22
- OpenAI's new model reportedly solved 100+ open math problems in 24 days of training — BilelKort · 2026-09-22
- Boltzbit previews paper claiming BAST lets LLMs learn up to 1,000x faster than SOTA training — jmhernandez233 · 2026-09-22
- MiMo's GRS and GAR: scoring what makes an RL answer genuinely good, not just passing — tokenbender · 2026-09-22