3rd-place ARC-AGI-2 submission open-sourced, using Gemini to turn historic puzzle books into specs

GregKamradt · x · 2026-11-07

Greg Kamradt highlighted that the 3rd-place ARC-AGI-2 Public Leaderboard submission by @gonedarragh and @JFPuget has been open-sourced.

Key details from the writeup:

• Gemini Flash first scouts each book-page image for mechanisms that could become grid puzzles;
• Kept pages are OCR'd by Gemini Pro/Flash, which writes up to four specs in NVARC's summary format (rule, grid limits, input generation, output relationship, solution steps, key insight and concepts);
• The pipeline keeps NVARC's 2025 stages (summaries, mixing, input code, output code) and mixes in ARC-AGI-2 training, ConceptARC, MINI-ARC and RE-ARC augmented sets;
• Each book spec is paired with 20–46 ARC task summaries — the book supplies the rule, ARC supplies the idiom of grids, colours and objects.

Kamradt noted he didn't know the team had pulled in puzzle data from 1588–1949.

Related event: ARC-AGI-2 Third-Place Solution Open-Sourced: Gemini Turns Century-Old Puzzle Books into Training Data(2 posts)→

Original post →

More from Research

• Near-identical neural net weights still diverge on ~3% of inputs after training — bravo_abad · 2026-11-08
• Removing negative-advantage samples drops robot zip-tie success from 86% to 60% — micoolcho · 2026-11-08
• DECIDE benchmark tests voice-agent decisions on 4,867 questions from 1,081 real calls, scoring 20 models — Aiden_Tech_Ai · 2026-11-08
• KV Cache Quantization Tested Across 35 Quants: 8-bit Is Essentially Free for Qwen3.8 27B — Hefty_Wolverine_553 · 2026-11-08
• YC amplifies demo of models handling semi-serious X-band phased-array design — ycombinator · 2026-11-08
• Analysis of 250k ICLR reviews finds 65% AI-written, scores collapsing to middle — TuhinChakr · 2026-11-08

Research channel →