37% on ARC-AGI-2 Kaggle With a 100M Looped Transformer Trained From Scratch, No LLM
rbhar90 · x · 2026-11-12
Independent researcher Mithil Vakde's Looped mdlARC — a 100M autoregressive transformer trained only on ARC puzzles, no LLM — scores 37.08% on ARC-AGI-2 Kaggle, likely the highest non-LLM score and 3x the next best among all 746 ARC Prize 2026 writeups, beating last year's winning LLM (NVARC, fine-tuned 4B Qwen, 27.64%). It peaked at 6th place before a late surge of Qwen entries.
Key points:
- Looping gives extraordinary gains: blocks 3–6 run six times, so 8 blocks yield 28 block passes; adding looping in the final week took the score 33% → 35% → 37%.
- Unlike HRM/TRM, nothing is hidden: an actual 1M looped model beats the older 76M one; uses in-context demo pairs and hub-and-spoke attention.
- The final model is severely undertrained and unoptimized — room to improve.
More from Research
- Deployment is where physical AI grows up — and its most important data source, per Tedrake's CoRL talk — m_wulfmeier · 2026-11-12
- Earlybird: a 29M-parameter predictive speech-to-text model that outputs words 90ms before you finish saying them — DrDatta_AIIMS · 2026-11-12
- A LoRA Is All You Need to Turn a Speech LLM Full-Duplex — kastnerkyle · 2026-11-12
- WICL: robots learn manipulation in-context, 83.4% success on all 50 RoboTwin tasks — _akhaliq · 2026-11-12
- Epoch: AI software progress cuts training compute needs several-fold per year — datagenproc · 2026-11-12
- Blood on the Clocktower benchmark: Kimi K3.1 Pro tops open models, just below GPT 5.5 — eliebakouch · 2026-11-12