Poetiq: SOTA Reasoning on ARC-AGI (github.com)

Poetiq has released a reproducible codebase for its record‑breaking submission to the ARC‑AGI‑1 and ARC‑AGI‑2 benchmarks, letting researchers rerun the exact experiment that achieved state‑of‑the‑art reasoning performance. The repository (and its companion launch post, “Traversing the Frontier of Superintelligence”) documents the Poetiq 3 configuration by default, while exposing alternative configs in config.py so you can vary problem sets, problem counts and other hyperparameters. Running the experiment requires Python 3.11+, model API keys (Gemini, OpenAI, etc.), a virtual environment, installing requirements, creating a .env with GEMINI_API_KEY / OPENAI_API_KEY, and executing python main.py. This is significant because ARC‑AGI is designed to probe long‑chain, abstract problem solving—areas where large models historically struggle—so a reproducible, top‑performing solution helps the community analyze which prompting, chaining, or model‑ensemble techniques actually drive gains. The repo makes the methodology transparent (configs, code paths, and defaults) and encourages citation of the accompanying paper and blog post. Practical implications include easier benchmarking against Poetiq’s approach, extending or stress‑testing its reasoning strategies, and investigating how model choice and configuration affect AGI‑style reasoning performance—though full replication requires access to the underlying commercial models.