There have been a couple of studies showing that attempting to erase multilingual knowledge using only English data is insufficient for multilingual LLMs. However, their analyses remain highly performance-oriented. In this paper, we switch the point of view to evaluation, and address an additional blind spot which reveals itself when the multilingual LLM is fully finetuned with parallel multilingual dataset before unlearning. Here, language confusion occurs whereby a model responds in language different from that of the input prompt. Language confusion is a problematic phenomenon in unlearning, causing the standard reference-based metrics to fail. We tackle this phenomenon in three steps: (1) introduce N-gram-based Language-Mix (N-Mix) score to quantitatively show the language confusion is pervasive and consistent in multilingual LLMs, (2) demonstrate that reference-based metrics result in false negatives when N-Mix score is high, and(3) suggest the need of new type of unlearning evaluation that can directly assess the content of the generated sentences. We call this type of metrics as semantic-based metric.

Large language models (LLMs) often experience language confusion, which is the unintended mixing of languages during text generation. Current solutions to this problem either necessitate model retraining or cannot differentiate between harmful confusion and acceptable code-switching. This paper introduces the Language Confusion Gate (LCG), a lightweight, plug-in solution that filters tokens during decoding without altering the base LLM. The LCG is trained using norm-adjusted self-distillation to predict appropriate language families and apply masking only when needed. Our method is based on the findings that language confusion is infrequent, correct-language tokens are usually among the top predictions, and output token embedding norms are larger for high-resource languages, which biases sampling. When evaluated across various models, including Qwen3, GPT-OSS, Gemma3, Llama3.1, LCG decreases language confusion significantly, often by an order of magnitude, without negatively impacting task performance. Code is available at https://github.com/collinzrj/language_confusion_gate.

Language confusion -- where large language models (LLMs) generate unintended languages against the user's need -- remains a critical challenge, especially for English-centric models. We present the first mechanistic interpretability (MI) study of language confusion, combining behavioral benchmarking with neuron-level analysis. Using the Language Confusion Benchmark (LCB), we show that confusion points (CPs) -- specific positions where language switches occur -- are central to this phenomenon. Through layer-wise analysis with TunedLens and targeted neuron attribution, we reveal that transition failures in the final layers drive confusion. We further demonstrate that editing a small set of critical neurons, identified via comparative analysis with a multilingual-tuned counterpart, substantially mitigates confusion while largely preserving general competence and fluency. Our approach matches multilingual alignment in confusion reduction for many languages and yields cleaner, higher-quality outputs. These findings provide new insights into the internal dynamics of LLMs and highlight neuron-level interventions as a promising direction for robust, interpretable multilingual language modeling. Code and data are available at: https://github.com/ercong21/lang_confusion.

As they become increasingly multilingual, Large Language Models (LLMs) exhibit more language confusion, i.e., they tend to generate answers in a language different from the language of the prompt or the answer language explicitly requested by the user. In this work, we propose ReCoVeR (REducing language COnfusion in VEctor Representations), a novel lightweight approach for reducing language confusion based on language-specific steering vectors. We first isolate language vectors with the help of multi-parallel corpus and then effectively leverage those vectors for effective LLM steering via fixed (i.e., unsupervised) as well as trainable steering functions. Our extensive evaluation, encompassing three benchmarks and 18 languages, shows that ReCoVeR effectively mitigates language confusion in both monolingual and cross-lingual setups while at the same time -- and in contrast to prior language steering methods -- retaining task performance. Our data code is available at https://github.com/hSterz/recover.

Large language models often suffer from language confusion, a phenomenon in which responses are partially or entirely generated in unintended languages. This critically degrades the user experience, especially in low-resource settings. We hypothesize that this issue stems from limitations in conventional fine-tuning objectives, such as supervised learning, which optimize the likelihood of correct tokens without explicitly penalizing undesired outputs such as cross-lingual mixing. Analysis of loss trajectories during pretraining further reveals that models fail to distinguish between monolingual and language-mixed texts, highlighting the absence of inherent pressure to avoid such confusion. In this work, we apply ORPO, which adds penalties for unwanted output styles to standard SFT, effectively suppressing language-confused generations. ORPO maintains strong language consistency, even under high decoding temperatures, while preserving general QA performance. Our findings suggest that incorporating appropriate penalty terms can effectively mitigate language confusion in multilingual models, particularly in low-resource scenarios.

Multilingual large language models (LLMs) often exhibit language confusion, a tendency to generate responses in a dominant language irrespective of the prompt's language. To address this, we propose Smoothie-Qwen, a lightweight, post-hoc method that mitigates language bias without retraining. This technique selectively adjusts token-level output probabilities to effectively suppress undesired language generation. Applied to the Qwen model, our method reduces unintended Chinese output by over 95% while preserving task accuracy on multilingual benchmarks. This work provides a practical and efficient solution for enhancing the language controllability of LLMs, making them more reliable for global applications.

Recently, the Mixture of Expert (MoE) architecture, such as LR-MoE, is often used to alleviate the impact of language confusion on the multilingual ASR (MASR) task. However, it still faces language confusion issues, especially in mismatched domain scenarios. In this paper, we decouple language confusion in LR-MoE into confusion in self-attention and router. To alleviate the language confusion in self-attention, based on LR-MoE, we propose to apply attention-MoE architecture for MASR. In our new architecture, MoE is utilized not only on feed-forward network (FFN) but also on self-attention. In addition, to improve the robustness of the LID-based router on language confusion, we propose expert pruning and router augmentation methods. Combining the above, we get the boosted language-routing MoE (BLR-MoE) architecture. We verify the effectiveness of the proposed BLR-MoE in a 10,000-hour MASR dataset.

Language Confusion is a phenomenon where Large Language Models (LLMs) generate text that is neither in the desired language, nor in a contextually appropriate language. This phenomenon presents a critical challenge in text generation by LLMs, often appearing as erratic and unpredictable behavior. We hypothesize that there are linguistic regularities to this inherent vulnerability in LLMs and shed light on patterns of language confusion across LLMs. We introduce a novel metric, Language Confusion Entropy, designed to directly measure and quantify this confusion, based on language distributions informed by linguistic typology and lexical variation. Comprehensive comparisons with the Language Confusion Benchmark (Marchisio et al., 2024) confirm the effectiveness of our metric, revealing patterns of language confusion across LLMs. We further link language confusion to LLM security, and find patterns in the case of multilingual embedding inversion attacks. Our analysis demonstrates that linguistic typology offers theoretically grounded interpretation, and valuable insights into leveraging language similarities as a prior for LLM alignment and security.

Large Language Models (LLMs) are susceptible to malicious influence by cyber attackers through intrusions such as adversarial, backdoor, and embedding inversion attacks. In response, the burgeoning field of LLM Security aims to study and defend against such threats. Thus far, the majority of works in this area have focused on monolingual English models, however, emerging research suggests that multilingual LLMs may be more vulnerable to various attacks than their monolingual counterparts. While previous work has investigated embedding inversion over a small subset of European languages, it is challenging to extrapolate these findings to languages from different linguistic families and with differing scripts. To this end, we explore the security of multilingual LLMs in the context of embedding inversion attacks and investigate cross-lingual and cross-script inversion across 20 languages, spanning over 8 language families and 12 scripts. Our findings indicate that languages written in Arabic script and Cyrillic script are particularly vulnerable to embedding inversion, as are languages within the Indo-Aryan language family. We further observe that inversion models tend to suffer from language confusion, sometimes greatly reducing the efficacy of an attack. Accordingly, we systematically explore this bottleneck for inversion models, uncovering predictable patterns which could be leveraged by attackers. Ultimately, this study aims to further the field's understanding of the outstanding security vulnerabilities facing multilingual LLMs and raise awareness for the languages most at risk of negative impact from these attacks.

We investigate a surprising limitation of LLMs: their inability to consistently generate text in a user's desired language. We create the Language Confusion Benchmark (LCB) to evaluate such failures, covering 15 typologically diverse languages with existing and newly-created English and multilingual prompts. We evaluate a range of LLMs on monolingual and cross-lingual generation reflecting practical use cases, finding that Llama Instruct and Mistral models exhibit high degrees of language confusion and even the strongest models fail to consistently respond in the correct language. We observe that base and English-centric instruct models are more prone to language confusion, which is aggravated by complex prompts and high sampling temperatures. We find that language confusion can be partially mitigated via few-shot prompting, multilingual SFT and preference tuning. We release our language confusion benchmark, which serves as a first layer of efficient, scalable multilingual evaluation at https://github.com/for-ai/language-confusion.