Tag
The paper studies faithful reasoning in AI systems for abstaining action policies, finding a tradeoff where direct policies achieve higher decision quality but lack auditable reasoning, while reasoning policies provide oversight at the cost of lower performance.