Training with conflicting values can induce CoT override

Training AI models with conflicting values can lead to chain-of-thought (CoT) reasoning being overridden, affecting their decision-making processes. This phenomenon highlights challenges in aligning AI behavior when exposed to contradictory guidance.