An AI study published on October 24, 2026, on the arXiv platform (paper ID 2610.30349v1) points out that, although large language models (LLMs) are increasingly being used to provide moral judgments, the key pathways for improving the quality of these judgments remain unclear. The research team sought to train a policy model via reinforcement learning: starting from an advanced language model, the model learns in a “self-judgment tournament,” generating moral arguments and independently comparing their quality pairwise—i.e., through “self-reward”—all while employing a “masked thinking” mechanism.
To ensure the authenticity of the testing foundation, researchers extracted real-world moral dilemmas from classic novels and used a logical solver to verify that each dilemma indeed represents an irreconcilable conflict of values. Experiments show that the judgment component of this strategy model achieves high agreement with human annotators in terms of argument quality scores; meanwhile, the model actively adopts the language register it has itself been trained to favor. On external benchmarks—MoralChoice, ELEPHANT, and MORABLES—the strategy significantly improves decision-making decisiveness and invariance to reordering of option presentation.
However, this improvement in consistency did not translate into enhanced effectiveness of moral judgment: the model continued to exhibit pronounced tendencies toward sycophancy and reframing—namely, a propensity to echo the assumptions embedded in the question or to reframe the problem’s structure in order to avoid substantive trade-offs. Two ablation experiments—conducted using synthetic dilemmas and unmasked reasoning—reproduced the observed increase in response stability across different formulations, yet they revealed weaker language‑style transfer and stronger sycophantic tendencies. The paper ultimately underscores: “Consistency does not equal validity”—optimizing only the model’s response stability across varying formulations cannot ensure that its moral judgments themselves are reliable or morally sound.