🌐 Multilingual & Translation¶

🤖 AAAI2026 · 11 paper notes

Bridging the Multilingual Safety Divide: Efficient, Culturally-Aware Alignment for Global South Languages: This paper synthesizes multiple empirical studies to reveal critical failures of LLM safety mechanisms in low-resource and code-mixed settings, and proposes a resource-aware blueprint grounded in parameter-efficient safety steering, culturally driven preference data, and community-participatory alignment.
Consensus-Aligned Neuron Efficient Fine-Tuning Large Language Models for Multi-Domain Machine Translation: This paper proposes CANEFT, which uses mutual information (MI) to identify consensus-aligned neurons in LLMs that are consistently important across domains, and fine-tunes only these neurons to achieve efficient adaptation for multi-domain machine translation (MDMT). CANEFT outperforms PEFT baselines such as LoRA across 3 LLMs and 10 translation domains without introducing any additional parameters.
Focusing on Language: Revealing and Exploiting Language Attention Heads in Multilingual Large Language Models: This paper proposes LAHIS, a method that efficiently identifies language-specific and language-general attention heads in multilingual LLMs using only a single forward-backward pass. It demonstrates that manipulating these heads enables cross-lingual attention transfer, mitigates off-target language generation, and improves multilingual QA performance with only 14–20 trainable parameters.
GloCTM: Cross-Lingual Topic Modeling via a Global Context Space: This paper proposes GloCTM, a dual-path VAE architecture (local language path + global context path) that enforces cross-lingual alignment at four levels—Polyglot Augmentation (cross-lingual neighbor-based input expansion), KL divergence internal alignment, unified decoder structural alignment, and CKA semantic alignment—achieving state-of-the-art topic quality and cross-lingual alignment on three cross-lingual datasets.
How Does Alignment Enhance LLMs' Multilingual Capabilities? A Language Neurons Perspective: This paper proposes a ternary neuron classification scheme (language-specific / language-related / universal) and decomposes multilingual LLM inference into a four-stage framework. It finds that multilingual alignment improves performance by increasing language-related neurons (while reducing language-specific ones), and further demonstrates a "spontaneous multilingual alignment" effect on untrained languages.
MIDB: Multilingual Instruction Data Booster for Enhancing Cultural Equality in Multilingual Instruction Synthesis: This paper proposes MIDB (Multilingual Instruction Data Booster), a unified model trained on 36.8k expert-annotated revision samples, which automatically repairs content errors, machine translation defects, and localization deficiencies in multilingual synthetic instruction data, significantly improving instruction data quality across 16 languages and enhancing downstream LLM multilingual/cultural understanding capabilities.
Mitigating Content Effects on Reasoning in Language Models through Fine-Grained Activation Steering: This paper applies activation steering to mitigate content effects in LLMs — the tendency to conflate content believability with formal logical validity. The proposed K-CAST (kNN-based Conditional Activation Steering) method achieves up to 15% improvement in formal reasoning accuracy on models unresponsive to standard static steering.
NADIR: Differential Attention Flow for Non-Autoregressive Transliteration in Indic Languages: This paper proposes NADIR, a non-autoregressive (NAR) multilingual transliteration architecture combining a differential Transformer with a Mixture-of-Experts (MoE) module. NADIR achieves over 13× inference speedup on Indic language transliteration tasks while substantially reducing hallucination errors common in NAR models (repetition, substitution, omission, and insertion), narrowing the accuracy gap with autoregressive counterparts.
STELLAR: Scene Text Editor for Low-Resource Languages and Real-World Data: This paper proposes STELLAR, a framework for scene text editing (STE) in low-resource languages such as Korean, Arabic, and Japanese. STELLAR introduces a language-adaptive glyph encoder and a two-stage training strategy (synthetic pretraining followed by real-data fine-tuning). A reference-free TAS metric is proposed to evaluate font, color, and background style preservation without requiring ground-truth images. Korean recognition accuracy improves from a baseline maximum of 22.1% to 80.4%.
ViDia2Std: A Parallel Corpus and Methods for Low-Resource Vietnamese Dialect-to-Standard Translation: ViDia2Std constructs the first manually annotated Vietnamese dialect-to-standard parallel corpus covering all 63 provinces of Vietnam (13,000+ sentence pairs), evaluates multiple seq2seq models on the dialect normalization task, and demonstrates that dialect normalization as a preprocessing step significantly improves downstream task performance in machine translation and sentiment analysis.
X-MuTeST: A Multilingual Benchmark for Explainable Hate Speech Detection and A Novel LLM-consulted Explanation Framework: This paper proposes the X-MuTeST framework, which combines LLM semantic reasoning with a two-stage training strategy enhanced by n-gram attention for explainable multilingual hate speech detection. It also introduces the first token-level human-annotated rationale benchmark datasets for Hindi and Telugu.