mlsanityData to 14 Oct 2026

Cultural Binding Heads in Language Models

arXiv 2605.28543 ↗First posted 27 May 2026cs.AI cs.CL cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

LLMs often default to equal treatment across cultural groups, even though context warrants differentiation: this is a lack of difference awareness. Using mechanistic interpretability and a factorial design on the N4 cultural appropriation benchmark from Wang et al. (2025), we identify 2-3 mid-layer attention heads per model that contribute causally to cultural binding across eight models (base and instruct versions of four architectures). Cultural binding is the process of associating a cultural item with its related identity. Knockout of the identity-to-item edges on these heads lowers the binding strength by 9-23%. The identified heads transfer from instruct to base models, suggesting that cultural binding is created during pre-training. An α-scaling shows a graded dose-response. Moderate amplification steering at generation (α = 2-3) increases cultural differentiation accuracy by 1-3 pp while leaving reasoning on culturally neutral questions mostly intact. A knowledge probing task shows that models know 3-6 times more than they act upon, indicating that the bottleneck lies in routing and not knowledge.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 15 of its 23 references that are in the map. First posted on arXiv 27 May 2026.

Builds on

15 of 23 references.

References · 15

  1. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  2. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  3. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  4. 2023Steering Language Models With Activation Engineeringcited by 922
  5. 2023The Hydra Effect: Emergent Self-repair in Language Model Computationscited by 156
  6. 2023How do Language Models Bind Entities in Context?cited by 108
  7. 2023Discovering Variable Binding Circuitry with Desideratacited by 34
  8. 2025Query-Focused Retrieval Heads Improve Long-Context Reasoning and Re-rankingcited by 31
  9. 2025Fairness through Difference Awareness: Measuring Desired Group Discrimination in LLMscited by 20
  10. 2025Localized Cultural Knowledge is Conserved and Controllable in Large Language Modelscited by 20
  11. 2025Isolating Culture Neurons in Multilingual Large Language Modelscited by 4
  12. 2026Steering LLMs for Culturally Localized Generationcited by 4
3 more
  1. 2025Finding Culture-Sensitive Neurons in Vision-Language Modelscited by 3
  2. 2025BiasGym: A Simple and Generalizable Framework for Analyzing and Removing Biases through Injectioncited by 1
  3. 2025Neuron-Level Analysis of Cultural Understanding in Large Language Modelscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025CulTrace: Tracing Internal Cultural Reasoning in Large Language Models
  2. 2025CURE: Cultural Understanding and Reasoning Evaluation - A Framework for "Thick" Culture Alignment Evaluation in LLMs
  3. 2026Aligned but Flattened: Analyzing the Trade-off between Cultural Alignment and Diversity in LLMs
  4. 2025From Word to World: Evaluate and Mitigate Culture Bias in LLMs via Word Association Test
  5. 2026The Culture Funnel: You Can't Align What isn't in the Data
  6. 2026Scenario-based Probing and Steering Cultural Values in Large Language Models--Extended Version
  7. 2025Do Large Language Models Truly Understand Cross-cultural Differences?
  8. 2024Investigating Cultural Alignment of Large Language Models