Research map / Large language models / Alignment and preferences

Training-Free Cultural Alignment of Large Language Models via Persona Disagreement

NeurIPS 2026 · Large language models › Alignment and preferences · approach: Steering and correcting LLM behavior · 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.

Proposes DISCA, a training-free, black-box inference-time method that uses disagreement among four WVS-grounded persona agents per country to apply a bounded, loss-averse, variance-shrunk logit correction for cultural moral alignment on MultiTP.

Problem

LLMs give moral judgements that are culturally misaligned with country-level human preferences (Moral Machine/MultiTP AMCEs), and existing fixes (fine-tuning, reward-guided decoding, activation steering) need per-country data, reward models, or white-box access.

Approach

Treat within-country persona disagreement, not consensus, as the steering signal: aggregate persona logit gaps into a consensus correction via loss-averse (Prospect-Theory) importance sampling, then shrink it by a dual-run inter-run variance gate so contested scenarios get less correction.

Key results

• On the Phi-4 20-country grid, DISCA reaches macro MIS 0.346 versus 0.454 for vanilla decoding and 0.384 for the best training-free baseline (PRISM-style prompt).
• DISCA reduces macro MIS by 10-24% on six backbones of at least 3.8B parameters; Phi-4 (14B) with DISCA beats vanilla Llama-3.3-70B in absolute MIS.
• Gains on the smallest 2B model are 3.4% on binary dilemmas; open-ended scenarios improve by 2-7%.

Surprising findings

• A 14B model with DISCA achieves lower absolute misalignment than a vanilla 70B model.
• Removing the persona component hurts 16-18 of 20 countries, while the method does not help factual cultural QA (BLEnD).

Limitations

• Requires decision-token logprobs, which some black-box APIs do not expose.
• Aligns to AMCE survey statistics, not to legitimacy or perceived appropriateness.
• Does not transfer to factual cultural QA; gains are small on smallest model and open-ended tasks.

Experimental setup

Datasets: MultiTP World Values Survey microdata Moral Machine AMCE vectors BLEnD

Benchmarks: MultiTP BLEnD

Metrics: MIS (L2 AMCE distance) Pearson r Jensen-Shannon divergence win count

Baselines: Vanilla decoding WVS Profile Prompt PRISM-style cultural prompt Fixed logit offset Activation steering MC-Dropout Margin scaling Temperature scaling

Builds on: Moral Machine MultiTP Prospect Theory value function self-consistency PRISM

Closest papers by meaning

• Language Model Alignment in Multilingual Trolley Problems (ICLR 2025 · similarity 0.89)
• Unveiling the Bias Impact on Symmetric Moral Consistency of Large Language Models (NeurIPS 2024 · similarity 0.88)
• ProgressGym: Alignment with a Millennium of Moral Progress (NeurIPS 2024 · similarity 0.87)
• Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values (NeurIPS 2025 · similarity 0.87)
• Position: A Roadmap to Pluralistic Alignment (ICML 2024 · similarity 0.87)
• The PRISM Alignment Dataset: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models (NeurIPS 2024 · similarity 0.86)