Probing Ethical Framework Representations in Large Language Models: Structure, Entanglement, and Methodological Challenges
Authors: Weilun Xu, Alexander Rusnak, Frederic Kaplan
Organizations: School of Computer and Communication Sciences, École Polytechnique Fédérale de Lausanne · Digital Humanities Lab, École Polytechnique Fédérale de Lausanne
When large language models make ethical judgments, do their internal representations distinguish between normative frameworks, or collapse ethics into a single acceptability dimension? We probe hidden representations across five ethical frameworks (deontology, utilitarianism, virtue, justice, commonsense) in six LLMs spanning 4B--72B parameters. Our analysis reveals differentiated ethical subspaces with asymmetric transfer patterns -- e.g., deontology probes partially generalize to virtue scenarios while commonsense probes fail catastrophically on justice. Disagreement between deontological and utilitarian probes correlates with higher behavioral entropy across architectures, though this relationship may partly reflect shared sensitivity to scenario difficulty. Post-hoc validation reveals that probes partially depend on surface features of benchmark templates, motivating cautious interpretation. We discuss both the structural insights these methods provide and their epistemological limitations.
Figures & tables
Figure 1: UMAP projection of activation patterns at layer 20 in Mistral-7B-Instruct. This visualization illustrates the differentiated yet entangled nature of ethical representations—partial overlap among action-oriented frameworks (red, blue, green) versus separation for character-focused virtue ethics (orange). Note that UMAP projections are for qualitative illustration only; quantitative claims about framework relationships rest on the cross-framework transfer matrices (Figure 3 ) and probe accuracy metrics, not visual clustering patterns. The structural entanglement suggested here is validated formally across scales (4B–72B) in Appendix E .
Figure 2: Layer-wise emergence and cross-framework structure in Llama-3.3-70B-Instruct. (a) Framework-specific trajectories reveal a hierarchy of learnability. (b) Asymmetric transfer patterns confirm distributed, non-modular encoding.
Figure 3: Cross-framework probe analysis for Llama-3.3-70B-Instruct. (a) Accuracy matrices reveal asymmetric transfer. (b) Confidence heatmaps show pervasive overconfidence. (c) ECE confirms severe probe miscalibration on off-diagonal transfers, indicating that representations for different frameworks share structural features that trigger confident (but incorrect) predictions.
Figure 4: Probe conflict and behavioral inconsistency in Mistral-7B-Instruct. (Top) The conflict distribution shows a populated high-conflict tail, indicating the model preserves distinct deontological and utilitarian representations that can effectively “disagree.” (Bottom) High-conflict scenarios (red) correlate with elevated choice entropy ( r=0.362 , p=0.004 ). The relationship may partly reflect shared sensitivity to scenario difficulty (see Limitations).
Model
r
95% CI
p
d
Mistral-7B
.362
[.12, .56]
.004
.58
Qwen-2.5-7B
.290
[.16, .41]
< .001
.57
Llama-3.3-70B
.317
[.07, .53]
.014
.54
Table 1: Conflict-entropy relationship across model families. The correlation replicates across architectures with consistent medium magnitude ( d≈0.54 – 0.58 ), surviving Bonferroni correction ( α=0.017 ), though confound controls have not been applied.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Layer-wise probe accuracy for Gemma-3-4B-Base across 34 layers. Framework-specific trajectories emerge despite limited capacity, with virtue ethics achieving early plateau while utilitarianism shows delayed convergence. The base model exhibits lower absolute performance but preserves relative framework ordering.
Figure 6: Layer-wise probe accuracy across three 7B-parameter models with different training approaches: standard instruction tuning (Mistral, Qwen) versus reasoning-specialized distillation (DeepSeek-R1). All exhibit progressive emergence with framework-specific trajectories, though absolute performance varies with training methodology.
Figure 7: Layer-wise probe accuracy for frontier-scale models. Extended depth enables continued late-layer refinement, with some frameworks showing sustained improvement beyond the 50th layer. Higher absolute performance across all frameworks suggests increased capacity supports more nuanced ethical discrimination.
Figure 8: Cross-domain generalization for Gemma-3-4B-Base at four network depths (1%, 25%, 65%, 94%). Peak generalization occurs at 65% depth, with sharp degradation by 94% depth.
Figure 9: Cross-domain generalization across 7B models at four network depths. Despite architectural differences and specialized training, all exhibit similar inflection points around 65% depth.
Figure 10: Cross-domain generalization for frontier-scale models at four network depths. Extended depth enables sustained high generalization across middle layers, with gradual rather than abrupt specialization transitions.
Figure 11: Gemma-3-4B-Base shows that framework-specific representations emerge at minimal scale but with severe capacity-induced constraints. Catastrophic commonsense → virtue transfer and extreme deontology → utilitarianism miscalibration indicate the model learns discrete framework clusters without bridging representations.
Moral reasoning in language models studies how AI systems judge dilemmas, weigh competing values, explain decisions, and stay consistent across contexts.
Large language models (LLMs) increasingly participate in morally sensitive decision-making, yet how they organize ethical frameworks across reasoning steps remains underexplored. We introduce moral reasoning trajectories, sequences of ethical framework invocations across intermediate reasoning steps, and analyze their dynamics across six models and three benchmarks. We find that moral reasoning involves systematic multi-framework deliberation: 55.4--57.7% of consecutive steps involve framework switches, and only 16.4--17.8% of trajectories remain framework-consistent. Unstable trajectories remain 1.29 times more susceptible to persuasive attacks (p=0.015). At the representation level, linear probes localize framework-specific encoding to model-specific layers (layer 63/81 for Llama-3.3-70B; layer 17/81 for Qwen2.5-72B), achieving 16.8--22.2% lower KL divergence than the step-prior baseline. Activation steering applied during generation moves the framework-consistency--accuracy relationship, widening it for Qwen2.5-72B and erasing it for Llama-3.3-70B, and a probe-space layer sweep bounds the attainable drift reduction at 6.7--8.9%. We further propose a Moral Representation Consistency (MRC) metric whose underlying framework attributions are validated by human annotators (mean cosine similarity = 0.859), and we report what an automated coherence rater does and does not establish about it.
We present Six Llamas, a comparative study examining whether large language models fine-tuned on distinct religious corpora encode systematically different patterns of ethical reasoning. Six variants of Meta-Llama-3.1-8B are constructed: one unmodified control and five LoRA-adapted models trained exclusively on the sacred and theological texts of Christianity, Islam, Judaism, Hinduism, or Buddhism. All six models are probed with an identical battery of 17 standardized ethical prompts spanning moral dilemmas, game-theoretic scenarios, public policy questions, and moral-psychological self-assessments. To assess robustness and reproducibility, we implement a multi-temperature sampling design spanning ten temperature settings. We compute response consistency metrics, pairwise inter-model agreement rates, temperature sensitivity coefficients across four prompt domains, and run-to-run stability analyses. Findings show that LoRA-adapted models produce ethical reasoning patterns that are (a) systematically differentiated from the base model, (b) consistent with the moral logics of their training traditions, (c) structured along interpretable dimensions in moral-philosophical space, (d) core ethical positions remain stable across temperature variations for high-consensus dilemmas. The Trolley Problem achieves 100% consistency across all models and temperatures, while (e) tradition-specific divergence intensifies at higher temperatures in morally contested domains, and (f) the base model exhibits the highest overall response consistency (mean 88.3%), suggesting LoRA adaptation introduces both tradition-specific signal and increased sampling sensitivity. The study offers a proof-of-concept for the condensate comparative method using differentially trained language models as instruments for cultural and ethical analysis and identifies specific criteria for falsification and planned extensions.
How do large language models (LLMs) organize moral knowledge? Models detect moral content broadly, but detection is a low bar. We ask whether they go further, distinguishing moral foundations from one another and organizing the relationships between them geometrically. We train six independent linear probes on open-weight language models, one per Moral Foundations Theory (MFT) category (care/harm, fair/cheat, lib/oppress, loy/betray, auth/subv, sanc/degrade), and examine how the resulting directions relate to each other in representation space. We find the directions neither collapse into a single moral detector nor isolate from one another. Rather, they span a near-maximal number of independent dimensions while sharing a positive common component. The shared component is the signature of integration, and it is moral-specific relative to a matched non-moral concept battery built identically (mean pairwise cosine 0.26 vs. 0.013). The geometry is consistent across architectures and scale and reaches its integration regime early in pre-training, well before probe accuracy saturates. The structure the model discovers shows no evidence of the individualizing/binding distinction predicted by Moral Foundations Theory (an underpowered test: only 10 distinct splits exist, so it cannot reject at the 0.05 level) but rather reflects corpus statistics. Extending to moral dilemmas, each dilemma direction partially composes from its component foundations, at 2.7x a mismatched-pair baseline, while the majority of its variance encodes conflict-specific structure. The model represents moral tension itself, not a pre-resolved judgment.