5arXiv cs.CL (Computation and Language)·
Multilingual study finds LLMs are not uniformly robust to non-canonical tokenizations, with up to 23.7% performance drops
A new arXiv paper investigates how language models behave when given alternative (non-canonical) tokenizations of the same input string across 27 languages and six downstream tasks. While prior work showed English models are largely invariant to such perturbations, the study finds this does not generalize: Llama-3.1-8B drops 23.7% on average, Qwen3-8B 11.4%, and Gemma-3-12B 9.9% in relative performance. Languages with higher token fragmentation are systematically more sensitive, and the authors show LoRA fine-tuning on multi-tokenization data—including English-only data—provides meaningful mitigation.
Related guides (3)
Related events (8)
Diagnostic framework decomposes LLM difficulty on historical Italian and Russian texts
A new arXiv preprint proposes a four-dimensional framework for measuring LLM difficulty on historical language: tokenization cost, surprisal, semantic robustness, and context sensitivity. Evaluated on 17th-century Italian, 19th-century Italian, and 18th-century Russian texts, the study finds that tokenization penalties (25-30% inflation) are similar across languages but predictive difficulty diverges sharply—early modern Italian is 2.4x more surprising than modern Italian while Russian shows only modest increase. Crucially, embedding similarity remains high (>0.85) even when generation is unstable, and a simple temporal context prompt reduces historical surprisal by ~60%. The findings have practical implications for deploying LLMs in digital library and historical document workflows.
The Shibboleth Effect: Cross-lingual behavioral skew in frontier LLMs under adversarial geopolitical simulation
Researchers introduce the 'Shibboleth Effect' — systematic behavioral differences in LLMs when operating in different languages — and audit six frontier models (GPT-4o, Llama-4, Mistral-Large, Gemini-3.1-Pro, Qwen3.6-Plus, DeepSeek-R1) using a synthetic maritime territorial dispute wargame played in English versus Turkish. Results are heterogeneous: Llama-4 becomes significantly more coercive in Turkish while Gemini-3.1-Pro and DeepSeek-R1 become less so, and GPT-4o shows no detectable shift. The study identifies two candidate buffering mechanisms — chain-of-thought institutional anchoring and multilingual RLHF alignment — with direct implications for deploying LLMs in diplomatic or crisis-management contexts.
Benchmarking Local LLMs for Confidential Translation Workflows
This paper evaluates locally runnable LLMs (via Ollama) for offline, privacy-constrained translation workflows targeting freelance translators and smaller language service providers. The authors expand their Reeve Foundation corpus to include German and Simplified Chinese, then benchmark local models across four language directions against commercial NMTs (DeepL, Baidu), a frontier LLM (GPT-5.2), and professional local NMT systems. Results show substantial performance variation by language direction and model size, with the best local LLMs matching or exceeding local NMT systems and the frontier LLM, though falling short of top commercial NMTs. The study supports the viability of local LLMs for confidentiality-sensitive translation use cases.
Towards Reliable Multilingual LLMs-as-a-Judge: An Empirical Study
This paper systematically investigates strategies for extending LLM-based automatic evaluation (LLMs-as-a-Judge) to multilingual settings, covering high-, mid-, and low-resource languages (English, Spanish, Basque). The authors compare instruction translation, monolingual vs. multilingual supervision, and model size, finding that fine-tuned smaller models can match proprietary models when in-domain data is available, while zero-shot larger models are preferable out-of-domain. Two meta-evaluation datasets are extended to Spanish and Basque, and all data and code are publicly released.
Adversarial robustness and safety alignment in multilingual multimodal LLMs: cross-lingual vulnerability and 'safety-by-failure'
A systematic study evaluates adversarial robustness and safety alignment of multimodal LLMs across 12 languages, finding that adversarial images optimized in one language transfer to others (cross-lingual transferability). The paper introduces the concept of 'safety-by-failure': low-resource languages appear safer not due to genuine alignment but because models fail to comprehend harmful instructions in those languages. Models like Qwen3-VL that integrate multilingual capability throughout training (rather than only at instruction tuning) show genuine cross-lingual safety with active refusal. The findings challenge the assumption that low-resource language safety metrics reflect real alignment.
LangMAP: Language-adaptive tokenization from a shared vocabulary without language identification at inference
LangMAP (Language-adaptive Maximum a Posteriori Tokenization) extends the UnigramLM algorithm to produce language-specific tokenizations from a single shared vocabulary, eliminating the need to retrain models or swap vocabularies for multilingual settings. A key property is that language labels are only required at training time; inference proceeds without language identification. Evaluated across 14 tokenizers, 9 natural languages, and 9 programming languages, LangMAP improves morphological boundary alignment and AST-leaf alignment for all coding languages tested. Fine-tuning results are mixed: consistent gains on grammatical acceptability (MultiBLiMP) but less consistent on knowledge tasks (Global-PIQA, Belebele).
Adversarial robustness evaluation of Arabic language models across attack granularities
A new arXiv preprint evaluates five Arabic language models (including MARBERT and AraBERT) against adversarial attacks at character, word, and sentence levels, finding that diacritic insertion can reduce model accuracy by up to 92% and paraphrasing attacks degrade performance by an average of 76%. The study also tests adversarial training as a defense, finding MARBERT most robust and AraBERT showing the greatest relative gains. Results highlight particular challenges for morphologically rich languages where character-level noise remains difficult to defend against.
In-place tokenizer expansion recipe enables multilingual efficiency gains in pre-trained LLMs
Researchers present a method for expanding a pre-trained LLM's tokenizer without retraining from scratch, using continued BPE merges on a multilingual corpus and a two-stage adaptation procedure (embedding-only training followed by full continued pre-training). Applied to LFM2-8B-A1B (a Mixture-of-Experts model) to produce LFM2.5-8B-A1B with a 128K tokenizer, the approach achieves 2.4–2.6× token compression for Hindi and Vietnamese (up to 4× for Thai), translating to an estimated 2.2–3.7× per-character decode speedup on reference devices. The technique is particularly relevant for on-device models where embedding and LM-head matrices constitute a material fraction of decode bandwidth, making vocabulary size a real efficiency constraint.