LCD Recovers Multilingual Performance in Quantized Models

Language-Conditional Dequantization (LCD) uses per-language rank-2 LoRA corrections to mitigate the 2-4x higher perplexity degradation seen in non-English languages during INT3 quantization. For 3B-parameter models, LCD recovers up to 83% of the perplexity gap for non-Latin scripts with minimal parameter overhead.