LLM Architecture & Training Design

All research

Research on novel concepts in LLM architecture and training design. Reinforcement-learning-driven training control, self-discovering architectures, and forensic interpretability of trained transformers.

The position

Most training stacks run on values a human chose. Someone picked the learning rate schedule, the clip threshold, the gate positions and the per-layer widths, defended the choice with intuition, and left it alone. The work in this subject area starts from the opposite assumption, that a decision made by hand is a decision nobody measured.

The results so far are concrete. A small reinforcement-learning policy placed on the training loop's forward-pass control surface beat a hand-tuned baseline by 42 percent at matched compute, matched data and matched architecture. The same primitive applied one level up lets a model discover its own per-layer width profile while training is still running. A four-role taxonomy of layers falls out of the finished checkpoints, computable without rerunning anything.

This research sits upstream of the finance work published elsewhere in the library. It is here because the same discipline governs both. A system that learns its own settings outperforms a system configured by people who were guessing.

Series

Each series is a coherent sequence of papers on a single topic within this subject area.

01 Training Substrate (4 papers) Reinforcement-learning-driven control of the training loop, self-discovering architectures, and the layer-role taxonomy that surfaces when those mechanisms are applied at scale.