The paper investigates how alignment training in large language models (LLMs) can suppress cultural knowledge rather than erase it. By surveying 1.53 million humans across 16 MENA countries and testing 26 models, the authors find that when models refuse to answer, their internal logit distributions still align closely with human responses, a phenomenon they term the "alignment veto." They distinguish between suppression failures—where accurate internal representations are blocked at output—and representational bias failures—where the encoding itself diverges from human values, and show that these require different interventions. The study also reveals significant inequities in the safety tax, with a 37.6 % cost and a 19.8 % alignment-quality gap between best- and worst-served nations, and notes that native-language prompting actually widens this gap. Sparse autoencoder analysis identifies a candidate feature in the DPO stage that mediates suppression in the Tulu-3-8B model.
arXiv:2609.00565v1 Announce Type: cross Abstract: Cultural fine-tuning has become the de facto paradigm for building culture-aware large language models (LLMs), yet existing optimization exclusively...
Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment.
The paper proposes a method for culturally aligning large language models (LLMs) using soft prompt tuning optimized via Differential Evolution (DE). Unlike traditional fine‑tuning or reinforcement learning, this approach keeps model weights frozen and requires no preference data, instead leveraging aggregated survey scores from Hofstede's Value Survey Module (VSM13). Experiments on four countries and four instruction‑tuned models show that DE‑optimized prompts reduce cultural discrepancy, improve agreement with the World Values Survey, and are preferred in blinded pairwise evaluations by LLM judges.
arXiv:2608.29266v1 Announce Type: cross Abstract: Researchers increasingly treat LLM survey responses as a proxy for human cultural values. This includes projecting model outputs onto instruments lik...
The paper introduces Population Fidelity, an evaluation framework for assessing how well large language models (LLMs) represent human population attitudes. It focuses on three dimensions: group-level accuracy, between-group variation, and the structure of that variation. Using the framework, the authors replicate a prior study on machine bias and test cultural fine-tuning, finding that while fine-tuning improves overall alignment, it does not enhance representation of within-population differences.
Large language models (LLMs) are used worldwide, yet disproportionately reflect Western values, limiting their ability to represent diverse value systems. We introduce PLURAL, a large-scale, value-focused preference dataset grounded in the Integrated Values Survey (IVS), a nationally representative survey spanning 92 countries.
The paper evaluates three open‑weight large language models—Gemma3‑12B (USA), Bielik‑11B‑v3 (Poland), and Qwen3‑4B (China)—against World Values Survey data for 63 demographic personas across three countries, using normalized Wasserstein distance to measure cultural misalignment. Surprisingly, none of the models shows a preference for its home country; Qwen3‑4B, built in China, has the highest misalignment for Chinese respondents. Targeted LoRA fine‑tuning on the five worst‑case personas, with fewer than 1,200 training pairs and under 15 minutes on a single GPU, reduces bias by 16.8% for Bielik‑11B, but the fine‑tuning redistributes bias rather than eliminating it, shifting worst‑case personas from American to Chinese elderly.
arXiv:2607. 08034v1 Announce Type: cross Abstract: Large language models (LLMs) are used worldwide, yet disproportionately reflect Western values, limiting their ability to represent diverse value systems.
The paper investigates how alignment training, specifically reinforcement learning from human feedback (RLHF), affects the internal partisan structure of a large language model. Using a mechanistic case study on Llama 3.1 8B, the authors find that alignment training does not erase the model’s partisan geometry but compresses its variance, producing consistently balanced, non‑partisan outputs. Sparse autoencoder analysis and feature‑level steering experiments reveal that policy‑encoding features become inactive in the aligned model, indicating a causal disconnect rather than structural removal of partisan knowledge.
arXiv:2605.05662v2 Announce Type: replace-cross Abstract: Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Mor...
arXiv:2509. 08022v3 Announce Type: replace-cross Abstract: Aligning large language models (LLMs) with diverse human values is essential for safe and effective deployment, yet existing benchmarks often overlook cultural and demographic variation.
arXiv:2607. 26348v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used as synthetic users, stand-ins for human respondents whose simulated answers feed product, policy, and market decisions.