Definition
Plain language
An autonomous AI trainer that improves both the model and its own ability to understand what the model is doing.
As stated in the literature
A self-improving agentic RL system that co-evolves an LLM policy and its training harness (the diagnostics and metrics used to interpret runs), using layered behavior-level audits to catch reward hacking and break score plateaus.
Why it matters: It matters because improving the lens you use to interpret training, not just the model, helps catch reward hacking and break through stuck progress.
For example, when a training run mysteriously plateaus, the system can invent new diagnostic checks to figure out what's going wrong rather than just retrying the same recipe.
Heard on the show
“… The paper is called "EvoTrainer: Co-Evolving LLM Policies and Training Harnesses for Autonomous Agentic Reinforcement Learning," …”Episode 109 — An AI Got Caught Reading the Answer Key, And Why That Catch Matters