SE-RRM paper hits ARC-AGI with just 2M params amid GPT-6 recurrent depth speculation
gklambauer · x · 2026-09-04
- Researcher Günter Klambauer speculates (unverified) that GPT-6 Astra may use "recurrent depth": transformer blocks with shared weights applied multiple times, the RRM idea his team proposed earlier this year.
- The new SE-RRM paper enforces permutation equivariance at the architectural level, guaranteeing identical solutions under symbol/color permutations without costly data augmentation.
- Results: outperforms prior RRMs like HRM/TRM on 9x9 Sudoku; generalizes from 9x9 training to 4x4, 16x16 and 25x25; competitive on ARC-AGI-1/2 with only 2M parameters. Code released.
Related event: GPT-6 Astra Rumored to Use Recurrent Depth as Tiny Model Matches ARC-AGI(2 posts)→
More from Research
- GPT-6 'Astra' Smashes ARC-AGI Records: 62.7% on Standard Harness, 99.9% on ARC-AGI-3 — repligate · 2026-09-04
- Embodied AI dataset ACE-Data-0 hits HF trending with ~30K downloads one week after release — liuziwei7 · 2026-09-04
- One-Shot OPD: a single query keeps on-policy distillation improving for hundreds of steps — ParshinShojaee · 2026-09-04
- Autoformalization nears economics: researcher bets on a Millennium Problem solved by 2027 — Afinetheorem · 2026-09-04
- From unique factorization to elliptic curve crypto: the 2,000-year math behind LLM security — anselm · 2026-09-04
- Beyond scalar scores: Nature paper backs psychometric profiles for AI evals — gleech · 2026-09-04