Show HN: We let GPT OSS 120B write and run Python and ARC AGI 2 jumped 4x (github.com)

A new project has been shared that significantly boosts performance for the ARC AGI 2 benchmark by utilizing stateful IPython-based REPLs to enhance the coding capabilities of models, specifically achieving over a fourfold improvement in GPT OSS 120B. This advancement showcases the impact of agentic coding, enabling models to interleave thinking and reasoning, which traditionally has struggled at the infrastructure level. These enhancements also extend to higher model iterations like GPT 5.2, leading to noticeable performance gains. The significance of this development lies in its potential to bridge the gap between artificial and human-like intelligence in complex tasks, particularly through improved reasoning and problem-solving capabilities. The technical implications involve necessary modifications to existing frameworks, like patches to vLLM, to support enhanced processing capabilities. The project provides detailed setup instructions for replicating these results, emphasizing how configurations can leverage various servers and local environments for optimal performance, making it a key resource for developers and researchers in the AI/ML community looking to push the frontiers of model capabilities.