{"slug": "37-on-arc-agi-2-kaggle-without-llms", "title": "37% on ARC-AGI-2 Kaggle Without LLMs", "summary": "Mithil Vakde reported scoring 37% on the ARC-AGI-2 Kaggle competition with a 100M-parameter looped transformer trained from scratch, without using an LLM, peaking at 6th place and beating last year's winning LLM entry. Vakde said adding looping in the final week lifted the score from 33% to 35% to 37%, and that at equal depth looping beat non-looped stacks with up to 4x the parameters by 2.3-7.6%. Vakde also reported 44% on ARC-AGI-1 for 67 cents, trained in 2 hours on a 5090, matching TRM and beating HRM, and argued the non-LLM approach matters because LLM sample efficiency on ARC cannot be measured.", "body_md": "Mithil Vakde on X: \"I got 37% on ARC-AGI-2 Kaggle without LLMs!\n100M looped transformer trained from scratch\n- Likely highest score without an LLM\n- Peaked at 6th place\n- Beats last year's winning LLM\n- Looping, added in the last week: 33% → 35% → 37%\nthanks to @arcprize @fchollet @mikeknoop and @GregKamradt for the competition!\nLink: https://t.co/iDBUrHhmdI\n(writeup for the paper track)\"\n\nI got 37% on ARC-AGI-2 Kaggle without LLMs!\n100M looped transformer trained from scratch\n- Likely highest score without an LLM\n- Peaked at 6th place\n- Beats last year's winning LLM\n- Looping, added in the last week: 33% → 35% → 37%\nthanks to @arcprize@fchollet@mikeknoop and @GregKamradt for the competition!\nLink: mvakde.github.io/blog/looped-md…\n(writeup for the paper track)\n\nI got 37% on ARC-AGI-2 Kaggle without LLMs!\n100M looped transformer trained from scratch\n- Likely highest score without an LLM\n- Peaked at 6th place\n- Beats last year's winning LLM\n- Looping, added in the last week: 33% → 35% → 37%\nthanks to @arcprize@fchollet@mikeknoop and @GregKamradt for the competition!\nLink: mvakde.github.io/blog/looped-md…\n(writeup for the paper track)\n\nWhat's new and different:\nLooping the middle layers\nAll demo pairs in context\nHub-and-spoke sparse attention (to fit in 4xL4s)\nCustom synthetic data (inspired by NVARC by @JFPuget and Ivan Sorkin)\n\nLooping is a huge win:\nAt equal depth, it beat my non-looped stacks with up to 4× the params, by 2.3–7.6% (even at equal depth)\nEven sub 1M models perform very well!\nRoom for growth -- I ran out of time, so final model is severely undertrained. Performance can go much higher\n\nHow does the looping differ from HRM/TRM:\n1) They claim 7M/25M params, but actually train 430M+ params. mdlARC beats their scores at TRUE 1M params\n2) mdlARC is autoregressive (1 pixel at a time), so justification for looping is MDL. HRM/TRM use loops to rewrite drafts of the whole answer (diffusion like)\n3) No hierarchical looping needed. Didnt help at all\n4) MUCH cheaper and faster to train\n\nWhy do it without an LLM?\nBecause I care about sample efficiency. LLMs see so much data, ARC puzzles included, that you can't measure it. We now know that with enough data, ARC-1/2 become trivial\nSo the interesting question is how far a model gets on a few ARC puzzles alone\n\n44% on ARC-AGI-1 in 67 cents!\nTrained from scratch in 2hrs on a 5090\nMatches TRM, beats HRM and is way faster & cheaper\nNo recursion, just a transformer\nAlso, 7% on ARC-2 🧵", "url": "https://wpnews.pro/news/37-on-arc-agi-2-kaggle-without-llms", "canonical_source": "https://twitter.com/evilmathkid/status/2120581574584582301", "published_at": "2026-11-12 00:58:20+00:00", "updated_at": "2026-11-12 01:23:10.454344+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-research", "large-language-models"], "entities": ["Mithil Vakde", "ARC-AGI-2", "Kaggle", "ARC Prize", "François Chollet", "Mike Knoop", "Greg Kamradt", "ARC-AGI-1"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/37-on-arc-agi-2-kaggle-without-llms", "markdown": "https://wpnews.pro/news/37-on-arc-agi-2-kaggle-without-llms.md", "text": "https://wpnews.pro/news/37-on-arc-agi-2-kaggle-without-llms.txt", "jsonld": "https://wpnews.pro/news/37-on-arc-agi-2-kaggle-without-llms.jsonld"}}