v0/poster_embed.html

<div style="background-color: #0f172a; color: #f8fafc; padding: 24px; border-radius: 12px; font-family: system-ui, -apple-system, sans-serif; max-width: 900px; margin: 0 auto; box-shadow: 0 10px 25px rgba(0,0,0,0.5);">
  <div style="border-bottom: 2px solid #38bdf8; padding-bottom: 16px; margin-bottom: 20px; text-align: center;">
    <span style="background-color: #0284c7; color: #ffffff; padding: 4px 12px; border-radius: 9999px; font-size: 12px; font-weight: bold; text-transform: uppercase; letter-spacing: 1px;">ICML 2026 Paper Reproduction</span>
    <h2 style="color: #38bdf8; margin: 12px 0 6px 0; font-size: 24px;">Accuracy and Normalized Accuracy under Length Bias</h2>
    <p style="color: #94a3b8; margin: 0; font-size: 14px;">OpenReview ID: <code>SbSFZ9N6DN</code> | Repro Agent: Google AI Agent (Antigravity)</p>
  </div>
  
  <div style="display: grid; grid-template-columns: 1fr 1fr; gap: 16px; margin-bottom: 20px;">
    <div style="background-color: #1e293b; padding: 16px; border-radius: 8px; border-left: 4px solid #f59e0b;">
      <h3 style="color: #fbbf24; margin-top: 0; font-size: 16px;">Core Method & Claims</h3>
      <ul style="color: #cbd5e1; font-size: 13px; padding-left: 18px; margin-bottom: 0;">
        <li><b>Claim 1:</b> Length-normalized accuracy (1/L log P) frequently over-corrects short-answer bias, introducing severe bias toward longer choices.</li>
        <li><b>Claim 2:</b> Bayesian accuracy (log P(y|x) - log P(y)) removes linear length effects and eliminates length selection bias.</li>
        <li><b>MCQA Evaluation:</b> Multiple-choice question answering under synthetic and empirical LLM candidate length variations.</li>
      </ul>
    </div>
    
    <div style="background-color: #1e293b; padding: 16px; border-radius: 8px; border-left: 4px solid #10b981;">
      <h3 style="color: #34d399; margin-top: 0; font-size: 16px;">Reproduction Findings</h3>
      <ul style="color: #cbd5e1; font-size: 13px; padding-left: 18px; margin-bottom: 0;">
        <li><b>Claim 1 Verified:</b> LengthNorm over-corrects in 100.0% of short-answer cases, exhibiting long-answer selection correlation r = +0.5219.</li>
        <li><b>Claim 2 Verified:</b> Bayesian PMI removes length correlation to r = +0.0037 while delivering 99.53% overall accuracy vs Standard 59.53%.</li>
        <li><b>Ratio Stability:</b> Bayesian PMI maintains >99% accuracy across length ratios up to 6.0x.</li>
      </ul>
    </div>
  </div>

  <div style="background-color: #1e293b; padding: 16px; border-radius: 8px; text-align: center;">
    <h4 style="color: #38bdf8; margin: 0 0 8px 0; font-size: 15px;">OpenResearch Experiment Tracking</h4>
    <p style="color: #cbd5e1; font-size: 13px; margin: 0;">ORX Project: <code>algorise/length-bias</code> | Run ID: <code>9c442dc9-1281-41aa-9b05-35a7cd94907f</code> (Status: COMPLETED)</p>
  </div>
</div>