Definition
Plain language
A training trick that throws out training prompts where every attempt got roughly the same score.
As stated in the literature
A prompt-filtering step in RL post-training that ranks prompts by reward variance across rollouts and updates only on high-variance prompts to avoid wasted gradient updates on uniform-reward groups.
Why it matters: Skipping prompts that produce uniform rewards saves a large fraction of training compute while losing almost no learning signal.
For example, if every one of eight rollouts on a prompt gets the same score, the prompt gets dropped from the training step because the gradient would be near zero anyway.
Heard on the show
“They call it SNR-aware filtering.”Episode 010 — When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL