{
  "id": 615041,
  "title": "LB 10.00 with Small Tweaks to 2024ARChitects",
  "content": "<p>My final solution ended up being a lightly modified version of the <a href=\"https://www.kaggle.com/code/dfranzen/arc-prize-2024-solution-by-the-architects\" target=\"_blank\">ARChitects2024approach</a> . None of my other experiments managed to beat it, so huge thanks to the ARChitects team for making their work public. </p>\n<p>Given the instability of the scores, I didn’t pick the one with the best LB result. Instead, I selected two submissions (linked below) that seemed to show more consistent behavior.</p>\n<p>Here are the main tweaks I made:</p>\n<p>1 . Training</p>\n<ul>\n<li>TTT (test-time training) for each task</li>\n<li>num_epoch = 24</li>\n<li>Reduced num_train to 3 for faster training</li>\n<li>batch_size = 4</li>\n<li>grad_accumulation_step = 2</li>\n<li>Removed background lines for faster training in one of the two submissions</li>\n</ul>\n<p>2 . Inference</p>\n<ul>\n<li>n = 1</li>\n<li>Reduced num_train to 2 for faster inference</li>\n<li>Used 2 subprocesses to speed up inference</li>\n<li>top_k = 4 for DFS to reduce search time</li>\n</ul>\n<p>With these hyperparameters, not all tasks finish within the 12-hour limit. I tried lowering num_epoch, but performance dropped a lot, so I stuck with 24.</p>",
  "comments": [],
  "comment_count": 0,
  "messages": [
    {
      "id": 3313003,
      "postDate": "2025-11-08T12:40:57.697Z",
      "content": "<p>My final solution ended up being a lightly modified version of the <a href=\"https://www.kaggle.com/code/dfranzen/arc-prize-2024-solution-by-the-architects\" target=\"_blank\">ARChitects2024approach</a> . None of my other experiments managed to beat it, so huge thanks to the ARChitects team for making their work public. </p>\n<p>Given the instability of the scores, I didn’t pick the one with the best LB result. Instead, I selected two submissions (linked below) that seemed to show more consistent behavior.</p>\n<p>Here are the main tweaks I made:</p>\n<p>1 . Training</p>\n<ul>\n<li>TTT (test-time training) for each task</li>\n<li>num_epoch = 24</li>\n<li>Reduced num_train to 3 for faster training</li>\n<li>batch_size = 4</li>\n<li>grad_accumulation_step = 2</li>\n<li>Removed background lines for faster training in one of the two submissions</li>\n</ul>\n<p>2 . Inference</p>\n<ul>\n<li>n = 1</li>\n<li>Reduced num_train to 2 for faster inference</li>\n<li>Used 2 subprocesses to speed up inference</li>\n<li>top_k = 4 for DFS to reduce search time</li>\n</ul>\n<p>With these hyperparameters, not all tasks finish within the 12-hour limit. I tried lowering num_epoch, but performance dropped a lot, so I stuck with 24.</p>",
      "rawMarkdown": "My final solution ended up being a lightly modified version of the [ARChitects2024approach](https://www.kaggle.com/code/dfranzen/arc-prize-2024-solution-by-the-architects) . None of my other experiments managed to beat it, so huge thanks to the ARChitects team for making their work public. \n\nGiven the instability of the scores, I didn’t pick the one with the best LB result. Instead, I selected two submissions (linked below) that seemed to show more consistent behavior.\n\nHere are the main tweaks I made:\n\n1 . Training\n- TTT (test-time training) for each task\n- num_epoch = 24\n- Reduced num_train to 3 for faster training\n- batch_size = 4\n- grad_accumulation_step = 2\n- Removed background lines for faster training in one of the two submissions\n\n2 . Inference\n- n = 1\n- Reduced num_train to 2 for faster inference\n- Used 2 subprocesses to speed up inference\n- top_k = 4 for DFS to reduce search time\n\nWith these hyperparameters, not all tasks finish within the 12-hour limit. I tried lowering num_epoch, but performance dropped a lot, so I stuck with 24.\n",
      "votes": 11
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-19T21:00:12Z",
  "competition_id": 91496,
  "competition_slug": "arc-prize-2025",
  "team_id": 13596133,
  "solution_url": "/competitions/arc-prize-2025/writeups/lb10-00-with-small-change-to-2024architects"
}