Definition
Plain language
A training signal that pays an agent only when its extra work actually moved the answer toward correct.
As stated in the literature
A reward formulation that compares an agent's full trajectory against a counterfactual shadow pass without certain interventions, granting credit only where the intervention is causally responsible for improved outcomes.
Why it matters: It stops agents from being rewarded for elaborate behaviors that didn't actually change the outcome, sharpening what the policy learns to value.
For example, an agent that retrieved a document and answered correctly only gets credit if the same model without that retrieval would have failed.
Heard on the show
“And in the paper this becomes what they call a contrastive reward.”Episode 051 — Why Parallel Sampling Plateaus, And What Evidence Graphs Do Instead