{
  "id": 614983,
  "title": "My attempt at ARC-AGI-2",
  "content": "<p>To be upfront, my accuracy score on the evaluation set is <a href=\"https://www.kaggle.com/code/huikang/arc-agi-2-code-approach/\" target=\"_blank\">zero</a>.\nI also do not claim to have improved performance even on the data I have trained on.</p>\n<p>In any case, these are my contributions, in two html files</p>\n<ul>\n<li>Golden solutions partial dataset - <a href=\"https://arc.huikang.dev/\" target=\"_blank\">arc.huikang.dev</a></li>\n<li>Finetuning dashboard - <a href=\"https://traces.huikang.dev/\" target=\"_blank\">traces.huikang.dev</a></li>\n</ul>\n<h2>What I am betting on</h2>\n<p>This is what I observe</p>\n<ul>\n<li>LLMs are already good at LeetCode problems.\nIn addition, it seems that LLMs can solve LeetCode even without a dedicated thought process. </li>\n<li>The code solutions to ARC-AGI problems is not much harder than LeetCode problems.\nThere are no advanced data structures.\nThere is no need to optimize for time complexity.\nThere are no edge cases that you find out and think through.\nMost of the time, you just need to put for loops in the correct places.\nFor half of the ARC-AGI-2 eval set problems, I think I can write the code in under 10 minutes.</li>\n</ul>\n<p>I thought that I could easily achieve 6% just by lightly aligning the LLMs. As you can see, this is much harder than expected.</p>\n<p>I see writing or following a procedure to be the way to solve ARC-AGI. As a human, you solve ARC-AGI with a process. You copy the input, change the grid size, change some colors - you follow some procedure. The procedure can be codified, with code.</p>\n<h2>General approach</h2>\n<p>This is my approach</p>\n<ul>\n<li>I create a set of \"golden\" solutions</li>\n<li>I use the golden solutions to guide the reinforcement learning</li>\n</ul>\n<p>The full writeup is at <a href=\"https://blog.huikang.dev/2025/11/04/arc-agi-2-attempt.html\" target=\"_blank\">https://blog.huikang.dev/2025/11/04/arc-agi-2-attempt.html</a></p>",
  "comments": [
    {
      "id": 3312923,
      "author_name": "CPMP",
      "author_url": "",
      "post_date": "2025-11-08T09:03:41.160000",
      "content": "<p>It looks like you should have entered the google code competition <a href=\"https://www.kaggle.com/competitions/google-code-golf-2025\" target=\"_blank\">https://www.kaggle.com/competitions/google-code-golf-2025</a></p>\n<p>It was about writing code to solve arc agi puzzles.</p>",
      "votes": 2,
      "replies": []
    }
  ],
  "comment_count": 1,
  "messages": [
    {
      "id": 3312776,
      "postDate": "2025-11-07T23:53:13.570Z",
      "content": "<p>To be upfront, my accuracy score on the evaluation set is <a href=\"https://www.kaggle.com/code/huikang/arc-agi-2-code-approach/\" target=\"_blank\">zero</a>.\nI also do not claim to have improved performance even on the data I have trained on.</p>\n<p>In any case, these are my contributions, in two html files</p>\n<ul>\n<li>Golden solutions partial dataset - <a href=\"https://arc.huikang.dev/\" target=\"_blank\">arc.huikang.dev</a></li>\n<li>Finetuning dashboard - <a href=\"https://traces.huikang.dev/\" target=\"_blank\">traces.huikang.dev</a></li>\n</ul>\n<h2>What I am betting on</h2>\n<p>This is what I observe</p>\n<ul>\n<li>LLMs are already good at LeetCode problems.\nIn addition, it seems that LLMs can solve LeetCode even without a dedicated thought process. </li>\n<li>The code solutions to ARC-AGI problems is not much harder than LeetCode problems.\nThere are no advanced data structures.\nThere is no need to optimize for time complexity.\nThere are no edge cases that you find out and think through.\nMost of the time, you just need to put for loops in the correct places.\nFor half of the ARC-AGI-2 eval set problems, I think I can write the code in under 10 minutes.</li>\n</ul>\n<p>I thought that I could easily achieve 6% just by lightly aligning the LLMs. As you can see, this is much harder than expected.</p>\n<p>I see writing or following a procedure to be the way to solve ARC-AGI. As a human, you solve ARC-AGI with a process. You copy the input, change the grid size, change some colors - you follow some procedure. The procedure can be codified, with code.</p>\n<h2>General approach</h2>\n<p>This is my approach</p>\n<ul>\n<li>I create a set of \"golden\" solutions</li>\n<li>I use the golden solutions to guide the reinforcement learning</li>\n</ul>\n<p>The full writeup is at <a href=\"https://blog.huikang.dev/2025/11/04/arc-agi-2-attempt.html\" target=\"_blank\">https://blog.huikang.dev/2025/11/04/arc-agi-2-attempt.html</a></p>",
      "rawMarkdown": "To be upfront, my accuracy score on the evaluation set is [zero](https://www.kaggle.com/code/huikang/arc-agi-2-code-approach/).\nI also do not claim to have improved performance even on the data I have trained on.\n\nIn any case, these are my contributions, in two html files\n\n- Golden solutions partial dataset - [arc.huikang.dev](https://arc.huikang.dev/)\n- Finetuning dashboard - [traces.huikang.dev](https://traces.huikang.dev/)\n\n\n## What I am betting on\n\nThis is what I observe\n\n- LLMs are already good at LeetCode problems.\n    In addition, it seems that LLMs can solve LeetCode even without a dedicated thought process. \n- The code solutions to ARC-AGI problems is not much harder than LeetCode problems.\n    There are no advanced data structures.\n    There is no need to optimize for time complexity.\n    There are no edge cases that you find out and think through.\n    Most of the time, you just need to put for loops in the correct places.\n    For half of the ARC-AGI-2 eval set problems, I think I can write the code in under 10 minutes.\n\nI thought that I could easily achieve 6% just by lightly aligning the LLMs. As you can see, this is much harder than expected.\n\nI see writing or following a procedure to be the way to solve ARC-AGI. As a human, you solve ARC-AGI with a process. You copy the input, change the grid size, change some colors - you follow some procedure. The procedure can be codified, with code.\n\n\n## General approach\n\nThis is my approach\n\n- I create a set of \"golden\" solutions\n- I use the golden solutions to guide the reinforcement learning\n\n\nThe full writeup is at https://blog.huikang.dev/2025/11/04/arc-agi-2-attempt.html\n",
      "votes": 10
    },
    {
      "id": 3312923,
      "postDate": "2025-11-08T09:03:41.160Z",
      "content": "<p>It looks like you should have entered the google code competition <a href=\"https://www.kaggle.com/competitions/google-code-golf-2025\" target=\"_blank\">https://www.kaggle.com/competitions/google-code-golf-2025</a></p>\n<p>It was about writing code to solve arc agi puzzles.</p>",
      "rawMarkdown": "It looks like you should have entered the google code competition https://www.kaggle.com/competitions/google-code-golf-2025\n\nIt was about writing code to solve arc agi puzzles.",
      "votes": 2
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-19T21:00:09Z",
  "competition_id": 91496,
  "competition_slug": "arc-prize-2025",
  "team_id": 14347062,
  "solution_url": "/competitions/arc-prize-2025/writeups/my-attempt-at-arc-agi-2"
}