Proforma Global
Proforma Global Research

Training Substrate

All research › LLM Architecture & Training Design

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.

What the series covers

The series follows one idea through four papers. Give the model control of its own training decisions and it outperforms the engineer who would otherwise make them.

It opens with the posture itself, the argument that every hyperparameter, threshold and gate in a training stack should be learned rather than picked. The second paper puts that to the test, where a small reinforcement-learning policy on the training loop's forward-pass control surface beat a hand-tuned baseline by 42 percent at matched compute. The third applies the same primitive one level up the stack, letting a model discover its own per-layer width profile mid-training. The fourth reads the finished checkpoints and finds a four-role taxonomy of layers, hub, damper, passthrough and specialist, computable from any trained transformer without rerunning the training.

Read them in order. Each paper assumes the one before it.

Papers

Listed in reading order.

  1. 01 Research Posture: Why We Let the Model Choose Why every hyperparameter, threshold, and gate in a training stack should be learned, not picked.
  2. 02 Reward-Driven Training Control A small reinforcement-learning policy on the training-loop control surface beat a hand-tuned baseline by 42% at matched compute.
  3. 03 Self-Discovering Architectures The same primitive applied one level up the stack: letting the model discover its own per-layer width profile mid-training.
  4. 04 Emergent Layer Roles and Functional Specialization A four-role taxonomy of layers in trained transformers (hub, damper, passthrough, specialist), computable from any finished checkpoint.

© 2026 Proforma Global. All rights reserved. Permission requests: [email protected].