Definition
Plain language
A batch of tries at the same problem that all scored identically, so there's nothing for the model to learn from.
As stated in the literature
In group-relative RL, a sampled group of rollouts that all receive the same reward, yielding zero advantage spread and therefore no gradient signal; wasted compute that filtering and richer reward shaping aim to revive.
Also called: dead groups
Why it matters: It matters because dead groups silently waste training compute, so spotting and reviving them makes learning far more efficient.
For example, if a model attempts the same problem eight times and all eight tries score zero, the batch is a dead group that teaches it nothing about how to do better.
Heard on the show
“They call that a dead group.”Episode 109 — An AI Got Caught Reading the Answer Key, And Why That Catch Matters