How Moral Concepts Are Organized Inside Language Models

In Tracing Moral Foundations in Large Language Models, researchers use Moral Foundations Theory (MFT) to investigate how moral concepts are encoded, organized and expressed. The study covers 14 base and instruction-tuned large language models from four families—Llama, Qwen2.5, Qwen3-MoE and Mistral—with model sizes ranging from 7 billion to 70 billion parameters.

The researchers combine three methods. First, they analyze representations of MFT concepts across model layers and assess how those representations align with human moral perceptions. Second, they use pretrained sparse autoencoders (SAEs) over the models’ residual streams to identify sparse features that support moral concepts. Third, they test causal steering interventions: they adjust representations using dense vectors associated with MFT and sparse SAE features, then examine whether foundation-related behavior shifts.

The authors report that models represent and distinguish moral foundations in ways that align with human judgments. Their analysis suggests that this moral geometry emerges during pretraining and is selectively rewired by post-training. At a finer level, they find SAE features with semantic links to particular foundations, which they interpret as evidence for partly disentangled mechanisms within shared representations. In other words, some features appear related to particular moral concepts, even though those concepts are represented within broader shared model activity.

The steering results add a test of whether internal representations are connected to model behavior. The authors report that steering with either dense vectors or sparse features produces predictable shifts in foundation-related outputs. Because the researchers intervene on representations, rather than only observing them, these results support a causal connection between the targeted representations and outputs in the tested setup. They do not, by themselves, establish what such representations mean in a human sense or how models use them in other settings.

The study addresses whether moral-sounding responses might reflect more than superficial mimicry. Its findings support the presence of structured, layered patterns associated with moral foundations in the models examined. The authors suggest that pluralistic moral structure can emerge as a latent pattern from language statistics alone. That is a claim about patterns learned by models, not evidence that they hold personal convictions or experience morality as people do.

The results should also be read within the study’s scope: a particular framework, a set of model families and scales, and the analyses and interventions described in the paper. The abstract-level account available here does not establish how broadly the findings apply beyond those conditions. Nor does representational alignment with human perceptions mean that a model’s judgments are consistently correct, or that its internal organization matches human moral thought in every respect.

This article is based on the abstract of an arXiv preprint, rather than a full-text assessment. The source supplied here does not establish whether the paper has undergone peer review. The findings offer evidence about how moral concepts may be represented and how targeted steering can affect related outputs, but they do not settle the broader question of what those patterns amount to.