Skip to content

Author

Shoko Wakamiya

2 papers indexed here

We haven’t gathered this author’s papers yet. Follow them and we’ll fetch their work.

Not the right person? Other researchers publish under this name.

Open access Jul 2026

Evaluating the cultural alignment of multilingual LLMs in typical Japanese workplace scenarios

While current evaluations of LLM cultural alignment predominantly rely on static benchmarks in Western contexts, their ability to navigate generative, high-context socio-pragmatic demands in non-Western environments remains critically underexplored. This study investigates how multilingual LLMs adapt to the Japanese workplace—a stringent stress-test environment characterized by strong high-context communication norms and rigid honorific conventions—using Hofstede’s six cultural dimensions as a heuristic framework. We evaluated five state-of-the-art LLMs (LLM-jp, Phi, Llama, Qwen, and GLM) through a large-scale crowdsourced human evaluation. Based on 1,718 valid evaluator sessions, native Japanese raters assessed model outputs to generate a holistic Japanese Workplace Cultural Alignment Score (JWCAS). To dissect the underlying communicative strategies, we paired this with a three-layer diagnostic sub-score analysis (Linguistic Form, Socio-Cultural Values, and Social Action). Our results reveal that leading multilingual models (Phi and GLM) achieved overall JWCAS scores comparable to, or significantly higher than, the native Japanese model (LLM-jp). Crucially, our sub-score analysis demonstrates that holistic evaluation metrics can obscure deep pragmatic deficits: while LLM-jp overfits to surface-level linguistic politeness (Layer 1), it shows critical weaknesses in socio-cultural values (Layer 2) and context-aware social strategies (Layer 3). In contrast, leading multilingual models demonstrate balanced competence across all layers. These findings suggest that true cultural competence requires moving beyond native linguistic mastery, highlighting the necessity of multi-dimensional diagnostic frameworks for cross-cultural AI alignment.

Zhiwei Gao, Nobuyuki Shimizu, Sumio Fujita et al. · 0 citations
Preprint Aug 2026

What Does Activation Steering Control? Attribution Across Answer Encodings and Output-Sensitive Subspaces

Activation steering is often evaluated under the answer encoding used to construct the direction. A reported gain may reflect the intended judgment or compatibility with answer identifiers seen during construction. We introduce Cross-Encoding Steering Evaluation, which freezes an intervention while re-encoding answers to the same held-out items. On NormBank, after A/B/C identifiers are reassigned, contrastive activation addition (CAA) induces larger target-versus-source score changes for the extraction indices than for the semantic labels under the new mapping. We call this extraction-index following. Varying identifier vocabulary (A/B/C, X/Y/Z, or 1/2/3) and row order shows that the effect tracks extraction index rather than row position. After matching direction norms across layers, extraction-index following emerges mainly at later depths. A low-rank output-sensitive component containing 15.4% of the direction's squared norm retains 96.3% of this effect. An Inference-Time Intervention (ITI)-style method also favors extraction-index over semantic-label following on NormBank in three models. In aggregate, MNLI favors extraction-index following, whereas Social Chemistry 101 (SC101) favors semantic-label following. Multiple-choice and open-ended evaluations can yield different behavioral conclusions. Thus, a steering gain under one answer encoding does not by itself identify what the intervention controls.

Zhiwei Gao, Shaowen Peng, Shoko Wakamiya et al. · 0 citations