CLaS-Bench: A Cross-Lingual Alignment and Steering Benchmark

2026-01-13

AI Summary

This work addresses the absence of standardized benchmarks for evaluating the efficacy of internal representation steering in multilingual reasoning within large language models. The authors propose CLaS-Bench, a lightweight parallel question benchmark spanning 32 languages, which establishes the first systematic protocol for multilingual steering evaluation and introduces a harmonic mean score that jointly accounts for language control and semantic relevance. Through comprehensive experiments employing diverse steering methods—including DiffMean on residual streams, probing directions, language-specific neurons, PCA/LDA, sparse autoencoders, and prompting—the study demonstrates that DiffMean consistently achieves the best performance across all languages. Analyses further reveal that steering directions cluster by language family and that language-specific structures predominantly emerge in the model’s later layers, offering a promising pathway for adapting models to low-resource languages.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Knowledge Representation Languages

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Understanding and controlling the behavior of large language models (LLMs) is an increasingly important topic in multilingual NLP. Beyond prompting or fine-tuning, , i.e.,~manipulating internal representations during inference, has emerged as a more efficient and interpretable technique for adapting models to a target language. Yet, no dedicated benchmarks or evaluation protocols exist to quantify the effectiveness of steering techniques. We introduce CLaS-Bench, a lightweight parallel-question benchmark for evaluating language-forcing behavior in LLMs across 32 languages, enabling systematic evaluation of multilingual steering methods. We evaluate a broad array of steering techniques, including residual-stream DiffMean interventions, probe-derived directions, language-specific neurons, PCA/LDA vectors, Sparse Autoencoders, and prompting baselines. Steering performance is measured along two axes: language control and semantic relevance, combined into a single harmonic-mean steering score. We find that across languages simple residual-based DiffMean method consistently outperforms all other methods. Moreover, a layer-wise analysis reveals that language-specific structure emerges predominantly in later layers and steering directions cluster based on language family. CLaS-Bench is the first standardized benchmark for multilingual steering, enabling both rigorous scientific analysis of language representations and practical evaluation of steering as a low-cost adaptation alternative.

Problem

• cross-lingual alignment
• steering
• large language models
• multilingual NLP
• benchmark

Innovation

• cross-lingual steering
• CLaS-Bench
• DiffMean
• language-specific representations
• multilingual alignment