The Answer Stays Right as the Reasoning Fails
TL;DR for operators A reasoning-capable AI assistant can produce the correct final answer while following an invalid reasoning path. A new staged backdoor study deliberately creates this condition: at its third activation stage, incorrect reasoning with correct answers occurs with mean success rates of 93.33% on PrOntoQA and GSM8K, and 77.78% on EntailmentBank, averaged across three model families. ...