Qwen Language Confusion Gate (arxiv.org)

Researchers introduced the Language Confusion Gate (LCG), a lightweight plug-in that reduces unintended mixing of languages in large language models (LLMs) by filtering token choices during decoding—without retraining the base model. LCG is language-aware: it predicts the appropriate language family for the next token and masks out off-language tokens only when needed, preserving legitimate code-switching. This makes it a practical, deployable fix for multilingual hallucinations and fidelity problems in chatbots, translation systems, and other multilingual AI services. Technically, LCG is trained via norm-adjusted self-distillation, exploiting three empirical findings: language confusion events are rare, correct-language tokens typically appear among the top predictions, and output token embedding norms are larger for high-resource languages (which biases sampling). The gate uses these signals to selectively mask tokens during beam/sampling decoding, mitigating the norm-driven bias. Evaluations on multiple models (Qwen3, GPT-OSS, Gemma3, Llama3.1) show LCG often reduces language confusion by an order of magnitude with no measurable hit to downstream task performance. The approach is model-agnostic, easy to integrate into existing inference pipelines, and the authors have released code for adoption and further research.