Definition
Plain language
Grading an AI agent on whether each individual step was the right move, instead of only on the final answer.
As stated in the literature
Per-step reward signals that score each agent decision against the correct action at that point (right source, right document, avoiding redundant searches), improving credit assignment and sample efficiency.
Also called: situational reward
Why it matters: By scoring each decision, it tells an agent exactly which moves were good or bad, helping it learn faster and assign credit correctly.
For example, instead of only grading the final answer, the agent gets credit for picking the right document to read at each step along the way.
Heard on the show
“They do the same targeted trick on the task side too — they call them situational rewards.”Episode 104 — How Making a Research Agent Smarter Quietly Makes It Leak Your Secrets