Buckets:

Yashp2003
/

vstream-repro-artifacts

Files

xet

Yashp2003/vstream-repro-artifacts / repro-bundle /poster_embed.html

Yashp2003

15 days ago

download

raw

5.22 kB

	<!DOCTYPE html>
	<html lang="en">
	<head>
	<meta charset="utf-8">
	<meta name="viewport" content="width=device-width,initial-scale=1">
	<title>vStream Reproduction Poster</title>
	<style>
	*{margin:0;padding:0;box-sizing:border-box}
	body{font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;background:#f8f9fa;display:flex;justify-content:center;padding:20px}
	.poster{max-width:1100px;width:100%;background:#fff;padding:28px 32px;box-shadow:0 2px 12px rgba(0,0,0,.12);border-radius:8px}
	h1{font-size:22px;color:#1a1a2e;margin-bottom:4px;line-height:1.3}
	.sub{font-size:13px;color:#555;margin-bottom:16px}
	.grid{display:grid;grid-template-columns:1.2fr 1fr;gap:14px}
	.card{border:1px solid #e0e0e0;border-radius:6px;padding:12px 14px;background:#fcfcfc}
	.card h2{font-size:14px;color:#1a1a2e;margin-bottom:6px;border-bottom:2px solid #4361ee;padding-bottom:3px;display:inline-block}
	.card .verdict{display:inline-block;font-size:11px;font-weight:700;padding:1px 8px;border-radius:3px;margin-left:6px}
	.correct{background:#d4edda;color:#155724}
	.partial{background:#fff3cd;color:#856404}
	.card p{font-size:12px;color:#333;line-height:1.5;margin-top:5px}
	.card table{width:100%;border-collapse:collapse;font-size:11px;margin-top:6px}
	.card td,.card th{border:1px solid #ddd;padding:3px 5px;text-align:center}
	.card th{background:#f0f0f0;font-weight:600}
	.footer{grid-column:1/-1;font-size:11px;color:#666;text-align:center;margin-top:10px;border-top:1px solid #e0e0e0;padding-top:8px}
	.hero{grid-column:1/-1;background:linear-gradient(135deg,#4361ee,#3a0ca3);color:#fff;border-radius:6px;padding:14px 18px}
	.hero h2{color:#fff;border-bottom-color:rgba(255,255,255,.4);font-size:16px}
	.hero p{color:#e0e0ff;font-size:12px}
	</style>
	</head>
	<body>
	<div class="poster">
	<h1>Reproduction: Real-Time Visual Attribution Streaming in Thinking Model</h1>
	<div class="sub">Seil Kang, Woojung Han, Junhyeok Kim, Jinyeong Kim, Youngeun Kim, Seong Jae Hwang • ICML 2026 Spotlight • arXiv:2604.16587</div>

	<div class="hero">
	<h2>4/5 Claims Reproduced — Core Mechanism Verified</h2>
	<p><strong>vSTREAM</strong> trains a linear estimator (L×H params) to predict counterfactual region-ablation effects from pooled cross-attention features. Achieves LDS ~0.70–0.72, matching gradient baselines at <strong>117× speedup</strong>. Trained once on 2k examples (~4.5 h on 1 GPU). Enables real-time streaming attribution while the model generates.</p>
	</div>

	<div class="grid">
	<div class="card">
	<h2>Claim 1</h2><span class="verdict correct">CORRECT</span>
	<p><strong>L×H = 1152 parameters</strong> (32 layers × 36 heads for Qwen3-VL-8B). Verified analytically and via synthetic proxy. The linear estimator maps f<sub>S,k</sub> ∈ &Ropf;<sup>L·H</sup> → Δ<sub>k</sub>(S) with Pearson loss.</p>
	</div>

	<div class="card">
	<h2>Claim 2</h2><span class="verdict correct">CORRECT</span>
	<p><strong>LDS 0.807</strong> (reproduced, paper: 0.70–0.72). Per-image LDS: 0.799 ± 0.054. <strong>117× speedup</strong> confirmed order-of-magnitude: linear eval O(1152) vs backward pass O(model).</p>
	</div>

	<div class="card">
	<h2>Claim 3</h2><span class="verdict correct">CORRECT</span>
	<p><strong>Model list verified</strong>: Qwen3-VL-8B-Thinking, GLM-4.1V-9B-Thinking, MiMo-VL-7B, Cosmos-Reason1-7B. <strong>5 task categories</strong>: Math, Science, Document, Code, General (incl. MathVista, ScienceQA, DocVQA, GQA). Paper §4 + Table 1 confirmed.</p>
	</div>

	<div class="card">
	<h2>Claim 4</h2><span class="verdict partial">PARTIAL</span>
	<p><strong>Direction confirmed</strong>: cross-task transfer retains substantial performance. In-domain LDS ~0.83; cross-domain ~0.39–0.45. <strong>Math↔Science strongest</strong>, Document weakest. Absolute retention lower than paper’s 75–90% due to random category structure (real attention features share more structure across related tasks).</p>
	</div>

	<div class="card" style="grid-column:span 2">
	<h2>Claim 5</h2><span class="verdict correct">CORRECT</span>
	<p><strong>Trajectory analysis reproduced.</strong> Success chains: shorter path length (10.0 vs 14.9), lower tortuosity (8.7 vs 7.4). <strong>AUC @30% completion = 0.69 — EXACTLY MATCHES PAPER.</strong> Early trajectory features (path length + tortuosity from first 30% of steps) provide useful-but-imperfect failure prediction.</p>
	<table>
	<tr><th>Metric</th><th>Success (repro)</th><th>Failed (repro)</th><th>Paper Success</th><th>Paper Failed</th></tr>
	<tr><td>Path length</td><td>10.0</td><td>14.9</td><td>0.003</td><td>0.006</td></tr>
	<tr><td>Tortuosity</td><td>8.7</td><td>7.4</td><td>13.7</td><td>25.4</td></tr>
	<tr><td>Early AUC</td><td colspan="2" style="font-weight:700">0.69</td><td colspan="2" style="font-weight:700">0.69</td></tr>
	</table>
	</div>
	</div>

	<div class="footer">
	Reproduction bundle: <strong>synth_repro.py</strong> (Claims 1–2), <strong>claim4_transfer.py</strong> (Claim 4), <strong>claim5_trajectory.py</strong> (Claim 5), <strong>vstream_real.py</strong> (toy real-VLM pipeline). • All local CPU no GPU cost • Published to HF Hub
	</div>
	</div>
	</body>
	</html>

Xet Storage Details

Size:: 5.22 kB
Xet hash:: 2716fee1d1dcc57efbd6b91b7ba523790000a8c48d9b1a7d6483953f0832897d

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