Tag
This paper evaluates how reliably sentiment concept directions extracted from language model representations reproduce across splits, comparing English, Hausa, and Yoruba over four language models. It finds consistent language rank order in direction agreement (English > Hausa > Yoruba) and shows that high probe classification accuracy does not imply directional consistency, though it does not establish tokenizer fertility as the cause.
This paper measures tokenizer fertility across 25 European languages on parallel text, revealing a 2.5x spread from English to Greek/Maltese, with Ukrainian paying a 15-18% penalty. It demonstrates domain invariance of fertility rankings, analyzes subword fragmentation, and evaluates cross-lingual few-shot effects.
Benchmarks seven foundation models on Ukrainian legal text, finding tokenizer fertility varies 1.6×, few-shot prompting degrades performance, and cost-performance analysis shows NVIDIA Nemotron Super 3 outperforms larger models.