Stratification Replicates Sparks of Weight-Space Programming on GPT 6 Astra and DeepSeek-V4.1-Flash

An independent replication and stratification study of the paper Sparks of Weight-Space Programming examines the reasoning models GPT-6 Astra and DeepSeek-V4.1-Flash on whether they can identify games from model weights. The author finds that while both models perform well on the surface, stratification experiments expose that Flash relies heavily on prompt metadata and channel labels, whereas Astra can recognize signals more robustly from partial weight inputs. Open article →