Definition
Plain language
A reward that pays an AI agent more when it finishes a task in fewer steps than its peers.
As stated in the literature
In ToolCUA's RL recipe, a group-relative reward term granting a linear bonus for sub-average step counts and applying an exponentially decaying penalty above average, asymmetric to encourage shorter trajectories.
Why it matters: Asymmetric step-count incentives push agents toward shorter, more reliable plans without harshly punishing ones that legitimately need to recover.
For example, an agent that books a flight in 5 steps gets a bonus if the group average was 8, while one that took 12 steps gets a sharper penalty.
Heard on the show
“It's a path-efficiency reward, and it's where GRPO becomes load-bearing.”Episode 066 — Why Giving an AI Agent More Tools Can Make It Worse at Using a Computer