Definition
Plain language
A flaw in training where every step of a long successful run gets the same credit, even the useless ones.
As stated in the literature
In group-relative RL, the failure of broadcasting a single trajectory-level advantage uniformly across all steps, erasing information about which decisions were pivotal; addressed by entropy-based reweighting that amplifies credit at high-uncertainty fork steps.
Why it matters: Without separating pivotal choices from filler, training can't tell the model which decisions actually earned the win, slowing learning on long tasks.
For example, an AI that solves a puzzle in twenty steps gets the same praise for the one clever decisive move as for the eighteen routine ones around it.
Heard on the show
“They call it advantage homogenization, and it's the flaw the fix exists to solve.”Episode 154 — How a 7B Model Out-Investigates a 72B One by Choosing What to Look At