ARCgentica

What it can do

• Solve ARC-AGI grid puzzles autonomously

  ARC-AGI input-output grid examples → Predicted output grids for test inputs

• Synthesize Python transformation programs

  Grid pattern examples showing input-output relationships → Executable Python code that transforms input grids to output grids

• Analyze abstract visual patterns in grids

  ARC-AGI puzzle grid data → Identified transformation rules and pattern descriptions

• Evaluate generated programs against test inputs

  Python transform programs and test grid inputs → Evaluated solutions with pass/fail scoring results

• Run multi-agent reasoning pipelines

  ARC-AGI puzzle and selected LLM backend (OpenAI, Anthropic, or OpenRouter) → Coordinated sub-agent analysis and solution attempts

• Score and benchmark solution performance

  Completed puzzle run results → Accuracy scores and performance metrics across puzzle sets

• Generate detailed run logs for AI reasoning research

  Executed ARC-AGI solving sessions → Full logs of agent reasoning steps, program attempts, and evaluations

Why it made the leaderboard

If you're researching symbolic AI, program synthesis, or multi-agent LLM architectures, ARCgentica gives you a fully open-source, auditable implementation that achieves one of the highest publicly reported scores on ARC-AGI-2 — a benchmark specifically designed to resist LLM shortcuts. The sub-agent architecture that writes and evaluates Python transformation programs is a concrete, reusable pattern for anyone building reasoning-heavy agent systems.

Comments (0)

No comments yet