Maxwell: coding agents re-architect neural nets mid-training, lifting AIME score from 64% to 71%

GabGarrett · x · 2026-11-07

Double AI publishes Maxwell, billed as the first 'Code Execution Model': an AI agent placed inside a training run performs 'mechanistic intervention,' diagnosing what the network fails to learn and rewriting its data, objective, update rule, or architecture on the fly.

Key findings

The agent builds its own toolkit, including the 'phantom gradient,' a measure of unmet learning demands discovered on nanochat, where loss fell below ordinary training after each new code layer.
It grows small programs into the network as 'alien' code layers — e.g. one computes gradients of hidden representations within a single forward pass for self-correction.
On CLRS-Text, transformers trained at certain input sizes fail badly on in-between sizes (33% accuracy, unaffected by 2x training); grown code layers lift that to 79%.
Algorithmic reasoning accuracy rose from 55% to 84%.
Self-modification: DeepSeek-V4.1-Flash acted as its own agent, raising AIME 2026 and HMMT 2026 scores from 64% to 71%.

The blog frames architecture as the last hand-designed piece of deep learning, now made part of training.

Original post →

Maxwell: coding agents re-architect neural nets mid-training, lifting AIME score from 64% to 71%

More from coding & agent