Multilingual Steering by Design: Multilingual Sparse Autoencoders and Principled Layer Selection

• URL: http://arxiv.org/abs/2605.23036v1
• Date: Thu, 21 May 2026 21:00:32 GMT
• Title: Multilingual Steering by Design: Multilingual Sparse Autoencoders and Principled Layer Selection
• Abstract summary: We show that training SAEs on multilingual data consistently strengthens cross-lingual representations.<n>We introduce an empha priori steering layer-selection rule based on the intersection of multilingual alignment and language separability.<n>Our results show that multilingual SAEs combined with intersection-selected layers stabilize the trade-off between language identification accuracy and generation quality.
• Score: 16.857692737043198
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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 \emph{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.

Related papers

• Strengthening Target-Language Features: SAE-Based Steering for Multilingual Inference [10.43186930809146]
  We propose an inference-time multilingual steering method that uses pretrained sparse autoencoders to identify and strengthen target-language-related features.<n>Experiments with Gemma-3-12B-it show average accuracy improvements of 10.9 percentage points on XCOPA, 5.3 points on XNLI, and 1.9 points on MGSM.
  arXiv  Detail & Related papers  (2026-08-05T14:32:14Z)
• Multilingual Refusal Alignment for Safer Large Language Models [53.64286756804503]
  We investigate whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general knowledge capabilities.<n>We introduce RefusEU, a novel refusal alignment dataset covering 12 European languages, including a dedicated test set for evaluating current state-of-the-art models.<n>Our controlled Direct Preference Optimization (DPO) experiments provide two key insights: aligning models exclusively in English is insufficient to ensure cross-lingual safety, even for the same harm categories, whereas training on multilingual datasets can improve safety without degrading general performance, as measured by the Global MMLU benchmark.
  arXiv  Detail & Related papers  (2026-04-24T09:29:14Z)
• CLaS-Bench: A Cross-Lingual Alignment and Steering Benchmark [21.574271160875046]
  We introduce CLaS-Bench, a benchmark for evaluating language-forcing behavior in large language models (LLMs) across 32 languages.<n>We find that across languages simple residual-based DiffMean method consistently outperforms all other methods.
  arXiv  Detail & Related papers  (2026-01-13T08:42:03Z)
• LangGPS: Language Separability Guided Data Pre-Selection for Joint Multilingual Instruction Tuning [49.22807995935406]
  Joint multilingual instruction tuning is a widely adopted approach to improve the multilingual instruction-following ability and downstream performance of large language models (LLMs)<n>Existing selection methods, often based on features like text quality, diversity, or task relevance, typically overlook the intrinsic linguistic structure of multilingual data.<n>We propose LangGPS, a lightweight two-stage pre-selection framework guided by language separability which quantifies how well samples in different languages can be distinguished in the model's representation space.
  arXiv  Detail & Related papers  (2025-11-13T12:02:32Z)
• Language-Specific Layer Matters: Efficient Multilingual Enhancement for Large Vision-Language Models [60.39744129890118]
  Large vision-language models (LVLMs) have demonstrated exceptional capabilities in understanding visual information with human languages.<n>In this work, we identify a salient correlation between the multilingual understanding ability of LVLMs and language-specific neuron activations in shallow layers.<n>We introduce PLAST, a training recipe that achieves efficient multilingual enhancement for LVLMs by Precise LAnguage-Specific layers fine-Tuning.
  arXiv  Detail & Related papers  (2025-08-25T18:15:25Z)
• Causal Language Control in Multilingual Transformers via Sparse Feature Steering [7.754609745940422]
  We investigate whether sparse autoencoder features can be leveraged to steer the generated language of multilingual language models.<n>We achieve controlled language shifts with up to 90% success, as measured by FastText language classification.<n>Our analysis reveals that language steering is most effective in mid-to-late transformer layers.
  arXiv  Detail & Related papers  (2025-07-17T06:49:16Z)
• Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders [41.1110443501488]
  We introduce a novel metric to assess the monolinguality of features obtained from SAEs.<n>We show that ablating these SAE features only significantly reduces abilities in one language of LLMs, leaving others almost unaffected.<n>We leverage these SAE-derived language-specific features to enhance steering vectors, achieving control over the language generated by LLMs.
  arXiv  Detail & Related papers  (2025-05-08T10:24:44Z)
• Enhancing Multilingual ASR for Unseen Languages via Language Embedding Modeling [50.62091603179394]
  Whisper, one of the most advanced ASR models, handles 99 languages effectively.<n>However, Whisper struggles with unseen languages, those not included in its pre-training.<n>We propose methods that exploit these relationships to enhance ASR performance on unseen languages.
  arXiv  Detail & Related papers  (2024-12-21T04:05:43Z)
• Lens: Rethinking Multilingual Enhancement for Large Language Models [70.85065197789639]
  We propose Lens, a novel approach to enhance multilingual capabilities in large language models (LLMs)<n>Lens operates on two subspaces: the language-agnostic subspace, where it aligns target languages with the central language to inherit strong semantic representations, and the language-specific subspace, where it separates target and central languages to preserve linguistic specificity.<n>Lens significantly improves multilingual performance while maintaining the model's English proficiency, achieving better results with less computational cost compared to existing post-training approaches.
  arXiv  Detail & Related papers  (2024-10-06T08:51:30Z)
• Soft Language Clustering for Multilingual Model Pre-training [57.18058739931463]
  We propose XLM-P, which contextually retrieves prompts as flexible guidance for encoding instances conditionally. Our XLM-P enables (1) lightweight modeling of language-invariant and language-specific knowledge across languages, and (2) easy integration with other multilingual pre-training methods.
  arXiv  Detail & Related papers  (2023-06-13T08:08:08Z)
• FILTER: An Enhanced Fusion Method for Cross-lingual Language Understanding [85.29270319872597]
  We propose an enhanced fusion method that takes cross-lingual data as input for XLM finetuning. During inference, the model makes predictions based on the text input in the target language and its translation in the source language. To tackle this issue, we propose an additional KL-divergence self-teaching loss for model training, based on auto-generated soft pseudo-labels for translated text in the target language.
  arXiv  Detail & Related papers  (2020-09-10T22:42:15Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.