ARC Lang Solver

Summary

jerber/arc-lang-public generates instruction programs for ARC-AGI grids, scores them by leave-one-out application on training examples (another LLM call follows the instructions onto the grid), then revises top instructions (StepRevision) or synthesizes a pooled plan (StepRevisionPool) before emitting up to two diverse test-grid guesses. Async orchestration with a monitored API semaphore; presets for Grok, GPT, mini, and open-source providers. Optional Neon/Postgres persistence of instruction scores. Bundles public ARC-Prize JSON under data/. No LICENSE / SPDX on the GitHub API. Python ≥3.10 (Ruff target 3.12).

Keywords

arc-agi · arc-lang · instruction-synthesis · grok · program-synthesis · jeremy-berman

Topics

ARC-AGI, LLM program synthesis, instruction following

Research notes

• Primary: GitHub README + API (no SPDX license, Python, 315 stars / 46 forks at check). Author from GitHub profile Jeremy Berman (@jerber). Discord posted the repo. data/ copies public ARC-Prize challenge JSON rather than a new hosted corpus, so no datasets_local row.