arcagi2/

arcagi2-agentic-coding-publication

This dataset accompanies our blog post Agentic coding improves ARC AGI 2 performance across models. This contains the complete outputs from the relevant experiment runs; including the full prompts, llm responses and tool calls.

We provide data from experiment run using three models, in the folder interleaved_thinking_vs_plain_cot:

gpt_5_2_xhigh gpt_oss_120b_high minimax_m2_1

For each model, unless otherwise noted, we provide data for

baseline: for each… See the full description on the dataset page: https://huggingface.co/datasets/arcagi2/arcagi2-agentic-coding-publication.