Sparse autoencoders (SAEs) enable feature-level mechanistic interpretability and activation steering in large language models (LLMs), but SAE-based language control remains unreliable in multilingual settings: most SAEs are trained on English-only data, and steering layers are chosen heuristically. We address these limitations by advancing a principled, mechanistic account of multilingual language steering with SAEs. First, we show that training SAEs on multilingual data consistently strengthens cross-lingual representations and yields more reliable, quality-preserving language control across layers and model families. Second, we introduce an a priori steering layer-selection rule based on the intersection of multilingual alignment and language separability, which predicts effective intervention depths without exhaustive layerwise search. We evaluate our approach on LLaMA-3.1-8B and Gemma-2-9B across machine translation and cross-lingual summarization (CrossSumm), using SpBLEU, ROUGE-L, COMET, and LaSE. Our results show that multilingual SAEs combined with intersection-selected layers stabilize the trade-off between language identification accuracy and generation quality, providing a principled, predictive, representation-level account of multilingual SAE steering.

Large language models (LLMs) are increasingly used across diverse linguistic and cultural contexts, yet their cultural knowledge remains uneven across regions and languages. We present the DFKI-MLT system for SemEval-2026 Task 7 on cultural awareness, where we apply activation steering to multilingual LLMs using language vectors extracted from parallel FLORES data. Our method performs inference-time adaptation by adding language-specific steering vectors to the residual stream at a selected transformer layer, without any parameter updates. We participated in both the short-answer (SAQ) and multiple-choice (MCQ) tracks; however, only our MCQ submission received an official score. In the official MCQ track, we achieved 86.96% accuracy, ranking 7th out of 17 teams. To better understand system behavior, we conduct post-hoc analyses on the shared-task MCQ and SAQ settings. These analyses show that activation steering yields modest and heterogeneous improvements on cultural reasoning: gains are strongly layer-sensitive, vary substantially across language–region pairs (some configurations even degrade performance), and interact with prompt formulation (generic vs. culturally conditioned prompts). Our findings suggest that prompt design and activation steering should be jointly optimized for culturally aware multilingual inference. We release our code and experimental configurations at https://github.com/Yusser96/SemEval-2026-Track7.

Understanding and controlling the behavior of large language models (LLMs) is an increasingly important topic in multilingual NLP. Beyond prompting or fine-tuning, language steering, 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.

Multilingual language models excel across languages, yet how they internally encode grammatical tense remains largely unclear. We investigate how decoder-only transformers represent, transfer, and control tense across eight typologically diverse languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai. We construct a synthetic tense-annotated dataset and combine probing, causal analysis, feature disentanglement, and model steering to LLaMA-3.1 8B. We show that tense emerges as a distinct signal from early layers and transfers most strongly within the same language family. Causal tracing reveals that attention outputs around layer 16 consistently carry cross-lingually transferable tense information. Leveraging sparse autoencoders in this subspace, we isolate and steer English tense-related features, improving target-tense prediction accuracy by up to 11%% in a downstream cloze task.

Large language models (LLMs) exhibit strong multilingual abilities, yet the neural mechanisms behind language-specific processing remain unclear. We analyze language-specific neurons in Llama-3.1-8B, Mistral-Nemo-12B, and Aya-Expanse-8B & 32B across 21 typologically diverse languages, identifying neurons that control language behavior. Using the Language Activation Probability Entropy (LAPE) method, we show that these neurons cluster in deeper layers, with non-Latin scripts showing greater specialization. Related languages share overlapping neurons, reflecting internal representations of linguistic proximity.Through language arithmetics, i.e. systematic activation addition and multiplication, we steer models to deactivate unwanted languages and activate desired ones, outperforming established replacement approaches. These interventions effectively guide behavior across five multilingual tasks: language forcing, translation, QA, comprehension, and NLI. Manipulation is more successful for high-resource languages, while typological similarity improves effectiveness. We also demonstrate that neuron steering enhances downstream performance and reveal internal "fallback" mechanisms for language selection when neurons are progressively deactivated. Our code is made publicly available at https://github.com/d-gurgurov/Language-Neurons-Manipulation.

While recent work has begun to uncover the internal strategies that Large Language Models (LLMs) employ for simple arithmetic tasks, a unified understanding of their underlying mechanisms is still lacking. We extend recent findings showing that LLMs represent numbers in a digit-wise manner and present evidence for the existence of digit-position-specific circuits that LLMs use to perform simple arithmetic tasks, i.e. modular subgroups of MLP neurons that operate independently on different digit positions (units, tens, hundreds). Notably, such circuits exist independently of model size and of tokenization strategy, i.e. both for models that encode longer numbers digit-by-digit and as one token.Using Feature Importance and Causal Interventions, we identify and validate the digit-position-specific circuits, revealing a compositional and interpretable structure underlying the solving of arithmetic problems in LLMs. Our interventions selectively alter the model’s prediction at targeted digit positions, demonstrating the causal role of digit-position circuits in solving arithmetic tasks.

Document-level neural machine translation (NMT) has outperformed sentence-level NMT on a number of datasets. However, document-level NMT is still not widely adopted in realworld translation systems mainly due to the lack of large-scale general-domain training data for document-level NMT. We examine the effectiveness of using Paracrawl for learning document-level translation. Paracrawl is a large-scale parallel corpus crawled from the Internet and contains data from various domains. The official Paracrawl corpus was released as parallel sentences (extracted from parallel webpages) and therefore previous works only used Paracrawl for learning sentence-level translation. In this work, we extract parallel paragraphs from Paracrawl parallel webpages using automatic sentence alignments and we use the extracted parallel paragraphs as parallel documents for training document-level translation models. We show that document-level NMT models trained with only parallel paragraphs from Paracrawl can be used to translate real documents from TED, News and Europarl, outperforming sentence-level NMT models. We also perform a targeted pronoun evaluation and show that document-level models trained with Paracrawl data can help context-aware pronoun translation.