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

Source: [CLaS-Bench: A Cross-Lingual Alignment and Steering Benchmark](<https://www.sota2.com/research/paper/696722f8e25ad2fa9d1ec049>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-01-15T05:00:04.561Z

Authors: Daniil Gurgurov, Yusser Al Ghussin, Tanja Baeumel, Cheng-Ting Chou, Patrick Schramowski, Marius Mosbach, Josef van Genabith, Simon Ostermann

Published: 2026-01-13T08:42:03.000Z

## 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.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Language Steering on WikiMatrix](<https://www.sota2.com/research/sota/language-steering-on-wikimatrix.md>): Language Steering; WikiMatrix; ES Score: 40.2.

- [Steered Language Generation on FLORES+](<https://www.sota2.com/research/sota/steered-language-generation-on-flores-plus.md>): Steered Language Generation; FLORES+; Score (es): 21.4.

- [Language Steering on FLORES+, WikiMatrix, and Tatoeba (Averaged)](<https://www.sota2.com/research/sota/language-steering-on-flores-plus-wikimatrix-and-tatoeba-averaged.md>): Language Steering; FLORES+, WikiMatrix, and Tatoeba (Averaged); ACC x BLEU (ES): 32.2.

- [Language Steering on Tatoeba](<https://www.sota2.com/research/sota/language-steering-on-tatoeba.md>): Language Steering; Tatoeba; Performance (es): 35.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2601.08331>)

[PDF](<https://arxiv.org/pdf/2601.08331.pdf>)