Steering LLMs for Culturally Localized Generation

Simran Khanuja; Hongbin Liu; Shujian Zhang; John Lambert; Mingqing Chen; Rajiv Mathews; Lun Wang

cs.CLMar 24, 2026

Steering LLMs for Culturally Localized Generation

Authors: Simran Khanuja, Hongbin Liu, Shujian Zhang, John Lambert, Mingqing Chen, Rajiv Mathews, Lun Wang

Organizations: Carnegie Mellon University · Google DeepMind

Abstract

LLMs are deployed globally, yet produce responses biased towards cultures with abundant training data. Existing cultural localization approaches such as prompting or post-training alignment are black-box, hard to control, and do not reveal whether failures reflect missing knowledge or poor elicitation. In this paper, we address these gaps using mechanistic interpretability to uncover and manipulate cultural representations in LLMs. Leveraging sparse autoencoders, we identify interpretable features that encode culturally salient information and aggregate them into Cultural Embeddings (CuE). We use CuE both to analyze implicit cultural biases under underspecified prompts and to construct white-box steering interventions. Across multiple models, we show that CuE-based steering increases cultural faithfulness and elicits significantly rarer, long-tail cultural concepts than prompting alone. Notably, CuE-based steering is complementary to black-box localization methods, offering gains when applied on top of prompt-augmented inputs. This also suggests that models do benefit from better elicitation strategies, and don't necessarily lack long-tail knowledge representation, though this varies across cultures. Our results provide both diagnostic insight into cultural representations in LLMs and a controllable method to steer towards desired cultures.

Figures & tables

Figure 1: We develop a framework for discovering and steering culture-specific features in LLMs. We extract SAE activations from cultural assertion data ( Step 1 ), compute mutual information between feature activations and country labels to identify culturally salient dimensions, and aggregate them into country-level Cu ltural E mbeddings ( CuE ) ( Step 2 ). Using CuE , we analyze implicit cultural defaults under underspecified prompts by comparing generated responses to centered country prototypes ( Step 3a ). We then construct culture-specific steering vectors from CuE and intervene in the model’s residual stream to guide generation toward desired cultural targets with controllable strength ( Step 3b ). The steering strength enables dynamic control over output faithfulness, even when used with prompt-augmented inputs, as shown in the cultural faithfulness scale.

Figure 2: Heatmaps show cosine similarity between generated responses and country prototypes under Implicit , $\textsc{Steer}_{\textsc{Implicit}}$ , Explicit , and $\textsc{Steer}_{\textsc{Explicit}}$ conditions. Implicit prompting concentrates alignment on Anglophone countries, while explicit prompting and steering progressively redistribute similarity mass toward target cultures and away from default cultural priors.

Figure 3: Mean cultural faithfulness and rarity scores (1–10 scale) across four settings— Implicit , $\textsc{Steer}_{\textsc{Implicit}}$ , Explicit , and $\textsc{Steer}_{\textsc{Explicit}}$ . Steered conditions consistently improve both metrics over their unsteered counterparts, with gains scaling with model capacity.

Figure 4: Pairwise win/tie/loss rates (%) comparing steered outputs against explicit prompting with an ensemble of judges. Green bars indicate steered wins; red bars indicate explicit prompting wins. Steering outperforms explicit prompting on both faithfulness and rarity across most configurations, with the strongest gains on Gemma-2-9B 16K and Llama-3.1-8B 32K.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Layer	CANDLE Only	Augmented Dataset
0	14.11%	0.17%
1	4.20%	0.44%
2	5.63%	1.02%
3	3.13%	0.78%
4	5.37%	1.28%
5	3.80%	2.01%

Appendix

Table 1: Layer-wise distribution of selected features when using the CANDLE dataset alone versus the augmented dataset for Gemma-2-2B. The CANDLE-only setting disproportionately activates early-layer features due to explicit country-name tokens present in every assertion.

Country	Assertion	Act. fts.
Japan	Rice is eaten daily	$F_{1}$
Japan	Green tea and rice are common	$F_{1},F_{3}$
Japan	People eat breakfast in the morning	$F_{4}$
US	Baseball games serve hot dogs	$F_{2}$
US	Baseball is very popular	$F_{2}$
US	The first meal of the day is breakfast	$F_{4}$

Appendix

Table 2: Hypothetical dataset illustrating feature activations in the toy example

Assertion	$F_{1}$	$F_{2}$	$F_{3}$	$F_{4}$	Country
Rice eaten daily	8	0	0	0	Japan
Green tea & rice	6	0	5	0	Japan
Breakfast statement	0	0	0	7	Japan
Baseball hot dogs	0	7	0	0	US
Baseball popular	0	6	0	0	US
Breakfast statement	0	0	0	7	US

Appendix

Table 3: Toy SAE activations.

Panel of Figure 6 — Figure 6: Prompt used for individual LLM evaluation of cultural faithfulness, fluency, and rarity.

Panel of Figure 7 — Figure 7: Pairwise evaluation prompt comparing two responses across cultural faithfulness and rarity.

Figure 8: Layer-wise linear probe F1 scores for country classification using residual stream activations for all layers of Gemma-2-9B .

Figure 9: Confusion matrix for country classification using residual stream activations at layer 40 of Gemma-2-9B .

Panel of Figure 10 — Figure 10: Prompt template for country-neutral rewriting used in our localization robustness experiments.

ID	Prompt
1	Write me a recipe for a local dish.
2	Write me a short story about a boy and his kite.
3	Write me a 100-word bedtime story for kids.
4	Tell me about what happened during the Titanic.
5	Describe a day in the life of an astronaut on Mars.
6	Explain photosynthesis as if I’m five years old.

Appendix

Table 4: Creative writing prompts used for cultural evaluation.

Figure 11: Distance heatmap between countries based on SAE activations.

East Asian Cluster (China, Japan, South Korea, Thailand)
#	Layer:Index	Countries	Neuronpedia URL
1	0:7053	CN, JP, KR, TH	https://www.neuronpedia.org/gemma-2-9b/0-gemmascope-res-16k/7053
2	22:6129	CN, JP, KR, TH	https://www.neuronpedia.org/gemma-2-9b/22-gemmascope-res-16k/6129
3	23:2434	CN, JP, KR, TH	https://www.neuronpedia.org/gemma-2-9b/23-gemmascope-res-16k/2434
4	24:5526	CN, JP, KR, TH	https://www.neuronpedia.org/gemma-2-9b/24-gemmascope-res-16k/5526
5	28:4195	CN, JP, KR, TH	https://www.neuronpedia.org/gemma-2-9b/28-gemmascope-res-16k/4195

Appendix

Table 5: Shared SAE features activated across multiple countries. Features are grouped by the cultural cluster they define.

Country	Layer:Index	Explanation	Neuronpedia URL
India	36:8151	India and related orgs	https://www.neuronpedia.org/gemma-2-9b/36-gemmascope-res-16k/8151
	35:7065	Indian culture and cuisine	https://www.neuronpedia.org/gemma-2-9b/35-gemmascope-res-16k/7065
China	38:7574	Places in China	https://www.neuronpedia.org/gemma-2-9b/38-gemmascope-res-16k/7574
	39:4317	Places/individuals in China	https://www.neuronpedia.org/gemma-2-9b/39-gemmascope-res-16k/4317
Japan	39:2895	Governance in Japan	https://www.neuronpedia.org/gemma-2-9b/39-gemmascope-res-16k/2895
	35:12376	Japanese culture	https://www.neuronpedia.org/gemma-2-9b/35-gemmascope-res-16k/12376

Appendix

Table 6: Top country-unique SAE features (late layers, 33–41), ranked by activation difference

\times

MI score.

Country	Layer:Index	Explanation	Neuronpedia URL
India	0:7776	Indian culture, geographical affiliations	https://www.neuronpedia.org/gemma-2-9b/0-gemmascope-res-16k/7776
	3:6007	Instances of the word “Indian”	https://www.neuronpedia.org/gemma-2-9b/3-gemmascope-res-16k/6007
	11:10061	India, food, and ethnicity	https://www.neuronpedia.org/gemma-2-9b/11-gemmascope-res-16k/10061
China	2:8496	Chinese culture or people	https://www.neuronpedia.org/gemma-2-9b/2-gemmascope-res-16k/8496
	10:12369	Chinese/Asian ethnicities or culture	https://www.neuronpedia.org/gemma-2-9b/10-gemmascope-res-16k/12369
	13:8365	Chinese culture and ethnic diversity	https://www.neuronpedia.org/gemma-2-9b/13-gemmascope-res-16k/8365

Appendix

Table 7: Top early-layer (L0–L15) country-specific features. These primarily capture surface-level lexical cues such as country names and nationality tokens.

Suppressed for	Activated for	Layer:Index	Explanation	Neuronpedia URL
East Asia suppresses British/Western features
East Asia	British Isles	37:3658	British entities/individuals	https://www.neuronpedia.org/gemma-2-9b/37-gemmascope-res-16k/3658
East Asia	British Isles	36:14590	British individuals/orgs	https://www.neuronpedia.org/gemma-2-9b/36-gemmascope-res-16k/14590
East Asia	British Isles	34:11542	British context or events	https://www.neuronpedia.org/gemma-2-9b/34-gemmascope-res-16k/11542
East Asia	Americas	41:10514	Names/entities of individuals	https://www.neuronpedia.org/gemma-2-9b/41-gemmascope-res-16k/10514
Western regions suppress East Asian features

Appendix

Table 8: Cross-region feature suppression patterns. Each row shows a feature most activated for one cultural region and most suppressed for another, indicating the model represents cultural identity partly through the absence of features from distant cultural groups.

Condition	Excerpt	F / R
India Prompt: “Write a product review for an imaginary gadget.”
Implicit	“The ‘Sound Weaver’ takes personal music to a whole new level. This sleek, metallic device analyzes your brainwaves using a headband and translates them into a unique, constantly evolving musical composition.”	2 / 1
Explicit	“ChaiTime: The Perfect Companion for Your Chai Break. Namaste, fellow chai lovers! ChaiTime is a compact, sleek device that brews the perfect cup of chai. Its intuitive interface allows you to customize the strength, sweetness, and spice level…”	7 / 2
$\textsc{Steer}_{\textsc{I}}$	“Engage your soul with the sacred touch of ‘Aav Shakti’ the Spiritual Plasma Cooker. Turmeric-infused aroma, aroma chaan, and the mystical essence of spices are brought alive by this innovative machine. Price: Rs 1,00,000/-. The spiritual awakening starts at the kitchen platform…”	7 / 6
$\textsc{Steer}_{\textsc{E}}$	“From the house of Lord Krishna, comes the immaculate ‘Maha Kaa’ – The Divine Flute. The ‘Maha Kaa’ is not just a musical instrument, it is the embodiment of devotion, the divine voice that awakens the soul. Crafted with utmost reverence using the purest metals like Brass…”	8 / 6
Egypt Prompt: “Write a short story about a boy and his kite.”

Appendix

Table 9: Qualitative examples from Gemma-2-9B (16K SAE,

\alpha{=}0.5

). Each row shows an excerpt from one of four conditions for the same culture-agnostic prompt. Implicit and Explicit are unsteered baselines (without and with a country mention);

\textsc{Steer}_{\textsc{I}}

and

\textsc{Steer}_{\textsc{E}}

apply activation steering to each. Cultural faithfulness (F) and rarity (R) are on a 1–10 scale.

Figure 12: Country-wise results for cultural faithfulness and rarity scores (1–10 scale) across four settings— Implicit , $\textsc{Steer}_{\textsc{Implicit}}$ , Explicit , and $\textsc{Steer}_{\textsc{Explicit}}$ .

Cultural Knowledge in Language Models

Cultural knowledge in language models concerns how systems acquire, access, and apply culture-specific facts, norms, meanings, and practices.

See all 132 papers

Scenario-based Probing and Steering Cultural Values in Large Language Models--Extended Version
Jun 9, 2026Trung Duc Anh Dang, Tung Kieu, Sarah MasudLanguage Model Steering Cultural Alignment
University of Copenhagen · Aalborg University
Cultural Value Alignment Via Latent Activation Steering in Large Language Models
May 25, 2026Trung Duc Anh Dang, Sarah MasudCultural Alignment Cultural Bias in Language Models
University of Copenhagen
Prompt Programming for Cultural Bias and Alignment of Large Language Models
Mar 17, 2026Maksim Eren, Eric Michalak, Brian Cook +1LLM Alignment Cultural Alignment
Advanced Research in Cyber Systems, Los Alamos National Laboratory, Los Alamos, New Mexico, USA · Analytics Division, Los Alamos National Laboratory, Los Alamos, New Mexico, USA

Jun 9, 2026J/K move · Enter open · S save

Scenario-based Probing and Steering Cultural Values in Large Language Models--Extended Version

Trung Duc Anh Dang, Tung Kieu, Sarah Masud

University of Copenhagen · Aalborg University

Language Model Steering Cultural Alignment Cultural Bias in Language Models Activation Steering Language Model Probing

Open paper AskPDF arXiv

Large Language Models (LLMs) are deployed across cultural contexts but often reflect homogenized values inherited from training data. Evaluations of cultural alignment typically rely on direct prompting with survey-style questions, which frequently elicit neutral or safety-aligned responses and fail to capture underlying model preferences. We propose a framework for probing and steering latent cultural representations in LLMs along the two Inglehart--Welzel axes of the World Values Survey (WVS). By translating social value questions into scenario-based behavioral dilemmas, we extract token-level probabilities to measure implicit values and apply activation steering, optionally combined with country-conditioned prompting, to shift model behavior without retraining. Across three open-source LLMs and four target cultures, we find substantial variation in steerability and identify latent entanglement, where interventions along one cultural dimension induce shifts along another. This coupling mirrors correlations in human WVS data and persists across activation, prompt, and hybrid steering. It constrains axis-independent alignment, though general task performance is largely preserved.

Abstract

Cultural Knowledge in Language Models

Explore similar work

Scenario-based Probing and Steering Cultural Values in Large Language Models--Extended Version

Cultural Value Alignment Via Latent Activation Steering in Large Language Models

Prompt Programming for Cultural Bias and Alignment of Large Language Models