https://taot.github.io/categories/experiments/Recent content in Experiments on Terry's Latent SpaceHugoen-usSat, 10 Oct 2026 15:52:33 -0400https://taot.github.io/posts/2026/10/leworldmodel-pusht-experiment-1/Fri, 09 Oct 2026 00:00:00 +0000https://taot.github.io/posts/2026/10/leworldmodel-pusht-experiment-1/<h1 id="lewm-push-t-my-experiment-vs-paper-arxiv-260319312v3">LeWM Push-T: My Experiment vs Paper (arXiv 2603.19312v3)</h1> <p><strong>Legend:</strong> ✓ = matches · ≈ = close / likely same · ⚠️ = differs, likely affects results · — = not stated in paper</p> <h2 id="1-training-parameters">1. Training parameters</h2> <table> <thead> <tr> <th></th> <th>Mine</th> <th>Paper</th> <th></th> </tr> </thead> <tbody> <tr> <td><strong>Epochs</strong></td> <td><strong>3</strong></td> <td><strong>10</strong></td> <td>⚠️</td> </tr> <tr> <td><strong>Image size</strong></td> <td><strong>112×112</strong></td> <td><strong>224×224</strong></td> <td>⚠️</td> </tr> <tr> <td>Encoder</td> <td>ViT-tiny, patch 14, 12 layers, 3 heads, hidden 192, from scratch</td> <td>ViT-tiny, patch 14, 12 layers, 3 heads, hidden 192, from scratch</td> <td>✓</td> </tr> <tr> <td>Embedding</td> <td>[CLS] → MLP projector (hidden 2048) + BatchNorm, dim 192</td> <td>[CLS] → 1-layer MLP + BatchNorm, dim 192</td> <td>≈</td> </tr> <tr> <td>Predictor</td> <td>transformer, 6 layers, 16 heads, dropout 0.1, AdaLN action conditioning, ~10M params</td> <td>transformer (&ldquo;ViT-S&rdquo;), 6 layers, 16 heads, 10% dropout, AdaLN, ~10M params</td> <td>✓</td> </tr> <tr> <td>History size</td> <td>3</td> <td>3</td> <td>✓</td> </tr> <tr> <td>Sub-trajectory</td> <td>4 frames + 4 blocks of 5 actions</td> <td>4 frames + 4 blocks of 5 actions</td> <td>✓</td> </tr> <tr> <td>Frameskip</td> <td>5</td> <td>5</td> <td>✓</td> </tr> <tr> <td>Batch size</td> <td>128</td> <td>128</td> <td>✓</td> </tr> <tr> <td>SIGReg λ</td> <td>0.09</td> <td>0.1 (λ ablation peaks near 0.09)</td> <td>≈</td> </tr> <tr> <td>SIGReg projections</td> <td>1024</td> <td>1024</td> <td>✓</td> </tr> <tr> <td>SIGReg knots</td> <td>17</td> <td>T nodes in [0.2, 4] (count not stated; insensitive)</td> <td>—</td> </tr> <tr> <td>Optimizer / LR / WD</td> <td>AdamW / 5e-5 / 1e-3</td> <td>—</td> <td>—</td> </tr> <tr> <td>Precision / grad clip</td> <td>bf16 / 1.0</td> <td>—</td> <td>—</td> </tr> <tr> <td>Dataset</td> <td><code>librakevin/lewm-pusht</code></td> <td>DINO-WM Push-T, 20,000 expert episodes</td> <td>≈ (not verified)</td> </tr> <tr> <td>Training seeds</td> <td>1 (seed 3072)</td> <td>3</td> <td>⚠️</td> </tr> </tbody> </table> <h2 id="2-eval-parameters">2. Eval parameters</h2> <table> <thead> <tr> <th></th> <th>Mine</th> <th>Paper</th> <th></th> </tr> </thead> <tbody> <tr> <td>Episodes</td> <td>50</td> <td>50 (same 50 for all models)</td> <td>✓</td> </tr> <tr> <td>Eval budget</td> <td>50 steps</td> <td>50 steps</td> <td>✓</td> </tr> <tr> <td>Goal offset</td> <td>25 steps</td> <td>25 steps</td> <td>✓</td> </tr> <tr> <td>CEM samples</td> <td>300</td> <td>300</td> <td>✓</td> </tr> <tr> <td>CEM iterations</td> <td>30</td> <td>30 (Push-T)</td> <td>✓</td> </tr> <tr> <td>CEM top-k</td> <td>30</td> <td>30</td> <td>✓</td> </tr> <tr> <td>CEM initial variance</td> <td>1.0</td> <td>1</td> <td>✓</td> </tr> <tr> <td>Planning horizon</td> <td>5 (= 25 env steps)</td> <td>5 (= 25 env steps)</td> <td>✓</td> </tr> <tr> <td>Receding horizon</td> <td>5 (execute full plan)</td> <td>5 (execute full plan)</td> <td>✓</td> </tr> <tr> <td><strong>Eval image size</strong></td> <td><strong>112</strong></td> <td><strong>224</strong></td> <td>⚠️</td> </tr> <tr> <td>Eval seed</td> <td>42 (default in <code>config/eval/pusht.yaml</code>)</td> <td>—</td> <td>—</td> </tr> </tbody> </table> <h2 id="3-eval-result-push-t-success-rate">3. Eval Result (Push-T success rate)</h2> <p><strong>All 10 seeds together (same epoch-3 checkpoint, 500 episodes in total):</strong></p>