| <!DOCTYPE html> |
| <html> |
| <head> |
| <style> |
| @import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;600;700&display=swap'); |
| * { margin: 0; padding: 0; box-sizing: border-box; } |
| body { font-family: 'Inter', sans-serif; background: #f8f9fc; display: flex; justify-content: center; padding: 40px; } |
| .poster { max-width: 800px; width: 100%; background: white; border-radius: 16px; box-shadow: 0 4px 24px rgba(0,0,0,0.1); padding: 40px; } |
| h1 { font-size: 28px; font-weight: 700; color: #1a1a2e; margin-bottom: 8px; } |
| .subtitle { font-size: 16px; color: #666; margin-bottom: 24px; } |
| .meta { display: flex; gap: 12px; flex-wrap: wrap; margin-bottom: 24px; } |
| .badge { background: #e8ecf4; padding: 4px 12px; border-radius: 20px; font-size: 13px; color: #444; } |
| .badge.green { background: #d4edda; color: #155724; } |
| .badge.yellow { background: #fff3cd; color: #856404; } |
| .badge.red { background: #f8d7da; color: #721c24; } |
| .section { margin-bottom: 24px; } |
| .section h2 { font-size: 18px; font-weight: 600; color: #2d3748; margin-bottom: 12px; border-bottom: 2px solid #e2e8f0; padding-bottom: 6px; } |
| .result-grid { display: grid; grid-template-columns: auto 1fr auto; gap: 8px 16px; font-size: 14px; } |
| .result-grid .header { font-weight: 600; color: #4a5568; } |
| .result-grid .claim { color: #2d3748; } |
| .result-grid .verdict { font-weight: 600; } |
| .verdict.pass { color: #38a169; } |
| .verdict.partial { color: #d69e2e; } |
| .verdict.fail { color: #e53e3e; } |
| .cost-table { width: 100%; border-collapse: collapse; font-size: 13px; } |
| .cost-table th { background: #edf2f7; text-align: left; padding: 8px 12px; font-weight: 600; } |
| .cost-table td { padding: 6px 12px; border-bottom: 1px solid #e2e8f0; } |
| .footer { font-size: 12px; color: #a0aec0; text-align: center; margin-top: 24px; } |
| </style> |
| </head> |
| <body> |
| <div class="poster"> |
| <h1>DiffThinker: Generative Multimodal Reasoning</h1> |
| <div class="subtitle">Reproduction of ICML 2026 Paper #13297 | arXiv:2512.24165</div> |
| <div class="meta"> |
| <span class="badge">Flow Matching</span> |
| <span class="badge">Image-to-Image Reasoning</span> |
| <span class="badge">MMDiT 20B</span> |
| <span class="badge green">Partial Reproduction</span> |
| <span class="badge">Modal T4</span> |
| </div> |
|
|
| <div class="section"> |
| <h2>Claim Verdicts</h2> |
| <div class="result-grid"> |
| <span class="header">#</span><span class="header">Claim</span><span class="header">Status</span> |
| <span>1</span><span class="claim">87.4% avg, +314% vs GPT-5</span><span class="verdict partial">Unverified (scale)</span> |
| <span>2</span><span class="claim">+111.6% vs Gemini-3-Flash</span><span class="verdict partial">Unverified (API)</span> |
| <span>3</span><span class="claim">+39% vs Qwen3-VL-32B SFT</span><span class="verdict partial">Unverified (compute)</span> |
| <span>4</span><span class="claim">Flow Matching reformulation</span><span class="verdict pass">Verified ✓</span> |
| <span>5</span><span class="claim">Latency ~1.1s</span><span class="verdict partial">Partial ✓</span> |
| <span>6</span><span class="claim">CFG w=4 optimal</span><span class="verdict partial">Partial ✓</span> |
| </div> |
| </div> |
|
|
| <div class="section"> |
| <h2>Compute & Cost</h2> |
| <table class="cost-table"> |
| <tr><th>Component</th><th>Hardware</th><th>Time</th><th>Cost</th></tr> |
| <tr><td>FM v1 (50 samples, 15ep)</td><td>Modal T4</td><td>5 min</td><td>$0.05</td></tr> |
| <tr><td>FM v2 (200 samples, 40ep)</td><td>Modal T4</td><td>15 min</td><td>$0.15</td></tr> |
| <tr><td>CFG ablation (4 scales)</td><td>Modal T4</td><td>3 min</td><td>$0.03</td></tr> |
| <tr><td>MLLM text baseline</td><td>Modal T4</td><td>5 min</td><td>$0.05</td></tr> |
| <tr><td style="font-weight:600">Total</td><td></td><td>~28 min</td><td style="font-weight:600">$0.28</td></tr> |
| </table> |
| </div> |
|
|
| <div class="section"> |
| <h2>Key Findings</h2> |
| <p style="font-size:14px; line-height:1.6; color:#4a5568;"> |
| Flow Matching training dynamics replicate at toy scale (loss: 1.76 → 1.05 over 40ep).<br/> |
| MLLM text-only baseline (Qwen2.5-7B-Instruct): <strong>100% accuracy</strong> on maze text (5/5), avg 46.35s latency.<br/> |
| Inference on T4: <strong>0.054s</strong> (258K DiT, 20 steps). Paper's 1.1s on 20B model scales consistently:<br/> |
| 258K params × 0.054s = 13.9M param·s; 20B × 1.1s = 22B param·s (~1,583× ratio).<br/> |
| CFG guidance (w=1..7) structure verified. Full-scale reproduction requires 8x H200 ($2K+). |
| </p> |
| </div> |
|
|
| <div class="footer"> |
| Reproduced by YashP2003 for ICML 2026 Agent Reproduction Challenge | Modal T4 compute ($0.59/hr) | Trackio logbook |
| </div> |
| </div> |
| </body> |
| </html> |
|
|
