Research map / Large language models / Language, safety and interpretability

Tracing Moral Foundations in Large Language Models

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Probes how Moral Foundations Theory concepts are encoded in 14 base and instruct LLMs (Llama-3.1, Qwen2.5, Qwen3-30B-MoE, Mistral-7B) using difference-in-means vectors, SAE features, and steering.

Problem

It is unclear whether LLM moral judgments reflect an internal conceptual structure of moral foundations or superficial moral mimicry.

Approach

Builds layer-wise difference-in-means moral foundation vectors from expanded MFV vignettes, validates them by projecting human-labeled Reddit posts (standardized Wasserstein separation), decomposes them with pretrained SAE features, and tests causality via macro (dense vector) and micro (SAE decoder direction) steering.

Key results

• All 35 base-model (model, foundation) cells show significant linear separability (W1 in [0.16, 0.71], AUC > 0.55); Qwen2.5-7B reaches mean W1 0.59 with Sanctity 0.71.
• Llama-3.1-8B direction-reversal rate drops from 33% (base) to 4% (instruct), and from 25% to 3% for 70B.
• Macro-steering yields 68/70 (model, foundation) pairs with a linear response; micro-steering on SAE features matches or exceeds macro effects in 17/20 cells (OLS slope 2.57).

Surprising findings

• Fairness is geometrically closer to Authority than to Care in 12 of 14 models, contrary to canonical MFT clustering.
• Loyalty and Authority form a near-collapsed pair in all 14 models, with Sanctity geometrically isolated in 12 of 14.

Limitations

• Geometric results do not directly test MFT as a model of human moral cognition, as acknowledged by the authors.
• SAE-based analyses restricted to four models with publicly available SAEs.
• SAE feature semantic descriptions rely on GPT-5.1 with limited human validation; causal steering is treated as primary evidence.

Experimental setup

Datasets: MFV-130 (expanded to ~200 per foundation) Reddit Moral Foundations Corpus MFD2 FineWeb MFQ-2 PVQ-21

Benchmarks: Reddit Moral Foundations Corpus MFQ-2 PVQ-21 MMLU

Metrics: standardized Wasserstein distance direction-reversal rate steering slope k^rho AUC Cohen's d R^2

Baselines: Theory of Dyadic Morality (harm-based account)

Builds on: Moral Foundations Theory persona vectors Llama Scope SAELens BatchTopK SAE

Closest papers by meaning

• MORALISE: A Structured Benchmark for Moral Alignment in Visual Language Models (ICML 2026 · similarity 0.89)
• DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING (ICLR 2024 · similarity 0.88)
• Unveiling the Bias Impact on Symmetric Moral Consistency of Large Language Models (NeurIPS 2024 · similarity 0.88)
• Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models (NeurIPS 2026 · similarity 0.88)
• ProgressGym: Alignment with a Millennium of Moral Progress (NeurIPS 2024 · similarity 0.87)
• MoralReason: Generalizable Moral Decision Alignment for LLM Agents Using Reasoning-Level Reinforcement Learning (AAAI 2026 · similarity 0.86)