The same system jumped from 44.9% to 75.5% on ARC-AGI-1 and hit 100% on Sudoku
hyperparticle · x · 2026-07-22
The follow-up post adds technical detail on the same system:
- On ARC-AGI-1, it reportedly improved from 44.9% to 75.5% using solutions it invented itself.
- The recipe was transferred to ARC-AGI-2.
- On Sudoku, it supposedly invented Hypothesis-Pinning Search, reached 100% on Sudoku-Extreme, wrote 125k+ lines of code, and ran 684 experiments.
The emphasis is on emergent method invention and the system’s ability to generalize its own recipe across tasks.
Related event: Rekursiv AI Agents Achieve Major Breakthroughs on ARC-AGI(2 posts)→
More from Models
- Fingerprint analysis says Kimi K3 and Fable 5 write more alike than sibling models — alex_verem · 2026-07-23
- Cisco says two small open security models beat GPT-5.5 on vulnerability detection cost — The Decoder · 2026-07-23
- Laguna-S-2.1 stumbles on a 100-meter walk-or-drive sanity check — logic_prevails · 2026-07-23
- BTL-3 launches as a 27B open-weight agent model in an 8.39GB file — QuixiAI · 2026-07-23
- Claude Opus 4.8 now drives 40% of Anthropic usage on OpenRouter — maferase · 2026-07-23
- GLM misses only “Computer locked” in nearly perfect ProgramBench cmatrix run — jyangballin · 2026-07-23