Tag
The paper introduces CHASE, a counterfactual-guided method for harness evolution in AI agent evaluation that mitigates benchmark shortcuts, enhancing reliability through constraint generation and validity checks.