新智元 reported that DoubleAI introduced “mechanistic intervention,” a method that places programming agents into model training and inference so they can generate and insert Python code layers. DoubleAI’s reported experiments include a 20-line layer that enabled gradient descent during one forward pass, while DeepSeek-V4.1-Flash reportedly used a self-written probe-and-resampling program to score 71.0% on AIME and HMMT 2026, compared with 63.9% for standard long-chain reasoning and 65.4% for length-only resampling. The supplied evidence comes from one secondary account and includes no independent replication or implementation audit.