SYLIN

mdlARC

by mvakde · Model · Language & code models

Small autoregressive transformer (not an LLM) trained from scratch in 1.5 hours on one RTX 5090 scores 44% on ARC-AGI-1 public eval

AccessOpen codeSize0.0189B parametersLicenseMIT-0

Versions

VersionReleasedRank
mdlARC 3LATESTSep 1
mdlARC 2Sep 1

Evidence

  1. Sep 1
    Commenters question the benchmark specificity: it only solves ARC puzzles, not general tasks; others dispute 'training on test' since eval puzzle labels are unused
  2. Sep 1
    Claimed to match TRM and HRM on ARC-AGI-1, compared only against models using similar test-time training, including CompressARC
  3. Sep 1
    Repo's v3 looped model: 18.9M params reaches 68.4% top-2 on ARC-AGI-1; 8.4M model about 3 hours on one H100