Double AI Research: Early Signs LLMs Can Redesign Their Own Architecture During Training via Mechanistic Intervention

Details

Double AI introduces 'mechanistic intervention,' where AI agents read internal network states and insert 'code layers' or modify architecture in real-time during the training loop, using a tool called 'phantom gradient' to map unmet demand in representations.
In the 'Closing the loop' experiment, DeepSeek-V4.1-Flash (763B MoE) acted as its own agent, designing a code layer to read its hidden states and reject incorrect answers, improving AIME 2026 and HMMT 2026 accuracy from 64% to 71% without thinking mode.
Timing is critical: inserting code layers mid-training yields better results than including them from the start, as early insertion causes 'shortcut learning' where the network relies on the layer instead of building its own circuits.
Code layers solve 'jaggedness' (failure on input sizes not seen in training) more effectively than doubling training time; on CLRS-Text, growing code layers improved gap problem accuracy from 33% to 79%, while standard training plateaued at 55%.
The approach allows for 'alignment through code' since inserted logic is human-readable, and enables recursive self-improvement where models accumulate probes and code layers as a tool library to refine their own learning process.

Key point