Skip to content
conftrace_
2026 EMNLP EMNLP 2026

MoralTrajectories: Evaluating Moral Steerability in Large Language Models

AbstractLarge language models are increasingly deployed in contexts where contested social and political values are at stake, yet we lack a systematic understanding of how their moral priors shape what they generate. We introduce **‘MoralTrajectories‘**, a benchmark for probing morally grounded generation across contentious political and social topics. The benchmark pairs questions from the AllSides Red Blue Translator with first-person belief statements operationalizing the six foundations of Moral Foundations Theory to construct, for each model and topic, a full *moral trajectory*: seven responses spanning an unconstrained baseline and one condition per foundation. We evaluate eight instruction-tuned models across families and sizes using lexical and semantic similarity, distributional divergence, and token-level uncertainty analysis, supported by a human annotation study. Three findings emerge consistently: unconstrained models default to CARE-oriented language, attributable to the empathy-rewarding objectives of instruction tuning; moral prompting shifts generation in foundation-consistent ways, but the effect is predominantly surface-level; and steerability varies substantially across model families and topics. Together, these results suggest that current alignment training collapses moral diversity into a narrow register, and that recovering genuine moral plurality in generation remains an open challenge.

More from EMNLP 2026

All EMNLP 2026 papers →