Lightweight but Linguistically Fair? Quantisation Robustness across Bavarian, German, and English
Post-training quantisation reduces the memory cost of large language models, but it remains unclear whether compression preserves language skills of regional varieties as reliably as their standard-language counterparts.
ProofPaper ↗
Key points
- We study this question for Bavarian, Standard German, and English using BAVGROUND, a multiple-choice benchmark, where 206 source questions are aligned across all three languages.
- Across 14 open-weight 7B--10B instruction models, we compare BF16 with 8-bit GPTQ and two 4-bit settings, GPTQ and AWQ.
- Calibration further shapes these cross-language effects: the relative Bavarian--German effect changes with calibration language, while changing the calibration source has little average effect on GPTQ-4 but can affect AWQ-4 much more strongly.
- Overall, our results show that quantisation robustness for regional language varieties depends on the model, quantisation method, calibration data, and evaluation setting, and cannot be measured by aggregate multilingual accuracy alone.
Sources (1)
- [1]Lightweight but Linguistically Fair? Quantisation Robustness across Bavarian, German, and EnglisharXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 01:15 PM
Post-training quantisation reduces the memory cost of large language models, but it remains unclear whether compression preserves language skills of regional varieties as reliably as their standard-language counterparts.
We study this question for Bavarian, Standard German, and English using BAVGROUND, a multiple-choice benchmark, where 206 source questions are aligned across all three languages.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026BELT : BEhavior-Aligned Low-Rank KV Cache Compression with Learned Top-k Rank Allocation
- Oct 13, 2026Towards Green AI: Evaluating Energy Efficiency for Quantized LLM Inference
- Oct 13, 2026Agent-Reflex: Teaching Language Agents to Act Reflectively
- Oct 13, 2026Virtual Head Attention
- Oct 12, 2026Fast Quantized Johnson$-$Lindenstrauss Transforms
- Oct 11, 2026More Reasoning Is Not Always Better: Executing Clinical Guidelines with LLMs and Rules
