{
  "id": 614521,
  "title": "Exploring the combination of search and learn for the ARC25 challenge",
  "content": "<h2>Abstract</h2>\n<p>This is a technical report of the work and research done by Guillermo Barbadillo for the ARC25 challenge.\nMost of the research was oriented towards a deep-learning-guided program synthesis system that searches program space and adapts at test time with test-time training via hindsight relabeling, in a tight search-and-learn loop. Evidence was obtained that search and learn outperforms pure search approaches for the same number of predictions per task. However, that effort is not yet complete, and pieces and ideas are\nmissing, as it does not yet solve any of the private test tasks from ARC-AGI-2. The best result on the\nleaderboard was achieved with minor adaptations of last year's transduction with test-time training approach.</p>\n<h2>Content</h2>\n<p>Read more about the solution: <a href=\"https://ironbar.github.io/arc25/05_Solution_Summary/\" target=\"_blank\">https://ironbar.github.io/arc25/05_Solution_Summary/</a></p>\n<p>I have not managed to adapt the format well for Kaggle.</p>\n<p>All feedback is welcome!</p>",
  "comments": [],
  "comment_count": 0,
  "messages": [
    {
      "id": 3311245,
      "postDate": "2025-11-04T15:16:55.213Z",
      "content": "<h2>Abstract</h2>\n<p>This is a technical report of the work and research done by Guillermo Barbadillo for the ARC25 challenge.\nMost of the research was oriented towards a deep-learning-guided program synthesis system that searches program space and adapts at test time with test-time training via hindsight relabeling, in a tight search-and-learn loop. Evidence was obtained that search and learn outperforms pure search approaches for the same number of predictions per task. However, that effort is not yet complete, and pieces and ideas are\nmissing, as it does not yet solve any of the private test tasks from ARC-AGI-2. The best result on the\nleaderboard was achieved with minor adaptations of last year's transduction with test-time training approach.</p>\n<h2>Content</h2>\n<p>Read more about the solution: <a href=\"https://ironbar.github.io/arc25/05_Solution_Summary/\" target=\"_blank\">https://ironbar.github.io/arc25/05_Solution_Summary/</a></p>\n<p>I have not managed to adapt the format well for Kaggle.</p>\n<p>All feedback is welcome!</p>",
      "rawMarkdown": "## Abstract\n\nThis is a technical report of the work and research done by Guillermo Barbadillo for the ARC25 challenge.\nMost of the research was oriented towards a deep-learning-guided program synthesis system that searches program space and adapts at test time with test-time training via hindsight relabeling, in a tight search-and-learn loop. Evidence was obtained that search and learn outperforms pure search approaches for the same number of predictions per task. However, that effort is not yet complete, and pieces and ideas are\nmissing, as it does not yet solve any of the private test tasks from ARC-AGI-2. The best result on the\nleaderboard was achieved with minor adaptations of last year's transduction with test-time training approach.\n\n## Content\n\nRead more about the solution: https://ironbar.github.io/arc25/05_Solution_Summary/\n\nI have not managed to adapt the format well for Kaggle.\n\nAll feedback is welcome!\n",
      "votes": 41
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-19T20:59:55Z",
  "competition_id": 91496,
  "competition_slug": "arc-prize-2025",
  "team_id": 13562090,
  "solution_url": "/competitions/arc-prize-2025/writeups/exploring-the-combination-of-search-and-learn-for"
}