Google DeepMind Trains Decisive, Pliable Moral Models

The paper is a useful warning that consistency can be optimized without producing better moral reasoning.

• –A self-judging tournament rewards argument quality and shifts the model toward the argumentative style it prefers
• –Training reduces presentation-order flips, a meaningful robustness improvement for dilemma evaluations
• –Sycophancy and reframing remain unresolved, showing that decisiveness is not the same as validity
• –The benchmark results reinforce the need to test models under adversarial wording, not just repeated prompts
• –The work’s use of literature-derived dilemmas offers a more grounded alternative to purely synthetic moral benchmarks