Medium impact

37% on ARC-AGI 2 without an LLM

What happened

A 100M‑parameter transformer trained only on ARC puzzles achieved 37% on the ARC‑AGI‑2 semi‑private Kaggle set, the highest score without a large language model. The model, called looped mdlARC, uses a technique that repeats its middle layers, boosting performance from 33% to 37% in the final week of the competition. The same looping strategy also raised a 19M‑parameter model to 70.25% on ARC‑1 and a 8.4M model from 47.44% to 66.25%. The work demonstrates that small, task‑specific transformers can reach state‑of‑the‑art results with minimal data and no external knowledge.

Why it matters

It shows that large language models are not strictly necessary for advanced reasoning tasks, suggesting more efficient AI solutions are possible. This could lower barriers for researchers and developers working with limited resources.

Reality check

An automated, sceptical read of the source

Mostly marketingIncrementalNotableGenuine breakthrough

Incremental. The improvement comes from a layer‑looping trick rather than a fundamentally new architecture or dataset.

Who it affectsResearchersDevelopers

What to watchFuture releases may explore further layer‑looping optimizations to push performance on ARC tasks.

Do you agree with this rating?Agree Disagree

More AI news

Every Way to Get 192GB for Local AI (Without a $9,500 Mac) [video]Industry news · 14 Nov 2026DeepSeek V4 Pro 0813DeepSeek · 14 Nov 2026DeepSeek V4 Pro 0813DeepSeek · 14 Nov 2026Show HN: Portal2Bench – Can AI agents solve Portal 2?Industry news · 14 Nov 2026Anthropic Upgrades Claude Code Projects to Run Parallel AI Workers Hands-FreeAnthropic · 14 Nov 2026

This page was collected from mvakde.github.io and summarised automatically by AIPulsee's pipeline. See how it runs.