doubleAI Says DeepSeek-V4.1-Flash Lifts AIME and HMMT 2026 Accuracy from 64% to 71%

doubleAI says coding agents it placed inside training runs can add code layers to a network while it learns. On CLRS-Text, accuracy at input sizes between those seen in training rises from 33% to 79%. DeepSeek-V4.1-Flash, acting on itself, lifts its AIME and HMMT 2026 accuracy from 64% to 71%.

• In nanochat, the agent inserts three code layers into a 286M-parameter transformer, and validation loss falls after each insertion.
• doubleAI says its agent discovered a tool called the phantom gradient, which measures unmet demand in a live network's representations.
• A lookup layer inserted at step 1,200 raises out-of-window recall to 72%, against 55% when present from step 0, doubleAI reports.