phantomcoder1996's picture
download
raw
2.83 kB
<!doctype html>
<html lang="en">
<head><meta charset="utf-8"><title>CrossQ toy reproduction poster</title>
<style>
@page{size:12.5in 8.333in;margin:0}
body{margin:0;background:#0b1020;color:#f7f8ff;font:24px Arial,sans-serif;text-wrap:pretty}
.poster{box-sizing:border-box;width:1200px;min-height:800px;padding:56px;background:linear-gradient(135deg,#111a38,#172b48)}
h1{font-size:56px;margin:0 0 16px;color:#8ee6d2;text-wrap:balance} h2{font-size:30px;color:#ffd166;text-wrap:balance}
.grid{display:grid;grid-template-columns:1fr 1fr;gap:24px;align-items:start}.card{padding:26px;border:2px solid #355070;border-radius:18px;background:#14233d}
.big{font-size:62px;font-weight:700;color:#ffd166}.note{font-size:18px;color:#b9c5d8;line-height:1.45}
a{color:#8ee6d2;text-decoration:none}
</style></head>
<body><main class="poster" data-measure-role="poster">
<h1>CrossQ: conditional quantization for late interaction</h1>
<p class="note">A transparent toy reproduction of the mechanism, not a paper-scale benchmark.</p>
<div class="grid" data-measure-role="column">
<section class="card" data-measure-role="card" data-logbook-target="claim-1-late-interaction-retrievers-like-colbert-achieve-high-quality-but-suffer-from-large-multi-vector-indices"><h2>Claim 1 · index pressure</h2><div class="big">8 × 32</div><p>Each synthetic document stores eight token vectors of dimension 32. Full precision uses 32 bits per scalar, illustrating why multi-vector indices grow quickly.</p></section>
<section class="card" data-measure-role="card" data-logbook-target="claim-2-crossq-adaptively-allocates-precision-by-conditioning-token-codes-on-lightweight-document-context-computed-at-indexing-time"><h2>Claim 2 · conditioning</h2><div class="big">2-bit + context</div><p>Document mean context is computed once at indexing time; token residuals are quantized relative to that context and reconstructed before MaxSim scoring.</p></section>
<section class="card" data-measure-role="card" data-logbook-target="claim-3-achieves-61x-compression-while-narrowing-gap-with-full-precision-colbert-to-just-2-3-mrr-10-loss"><h2>Claim 3 · result</h2><div class="big">9.85×</div><p>Our toy proxy reaches 9.85× conditional index compression. The released checkout has no paper code, weights, or benchmark data, so 61× and 2.3% are not verified.</p></section>
<section class="card" data-measure-role="card"><h2>Reproducibility status</h2><p>Local run: deterministic NumPy script, 120 documents, 48 queries. HF GPU Job: blocked before launch by account credit (HTTP 402). See the linked logbook pages for commands and outputs.</p><p class="note"><a href="https://openreview.net/forum?id=jxHF6z0S1L">OpenReview paper</a> · <a href="https://github.com/Chenruishuo/posterly">Posterly</a></p></section>
</div></main></body></html>

Xet Storage Details

Size:
2.83 kB
·
Xet hash:
1ea2060f2d122c6747e109f6dfdd5193d8541284b6b3693f262f75ad51a14eee

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.