QuechuaTok found the most efficient tokenizer for Quechua had the worst morphology, at 6.67 percent boundary accuracy
AspireAI community · arXiv
Fertility rate - how many tokens a word splits into - is the standard way to score a tokenizer, and it misses morphological correctness entirely for agglutinative languages where words are built by stacking affixes. QuechuaTok is a systematic benchmark comparing four tokenization strategies (BPE, Unigram LM, WordPiece, and a morphology-aware PRPE tokenizer) for Southern Quechua, a low-resource agglutinative language spoken by 8 to 10 million people.
Using a 200k-sentence corpus and the SQUOIA finite-state morphological analyzer as silver standard, the results are stark. BPE achieves the lowest fertility rate at 1.636 with a 16k vocabulary - because it memorises surface word forms - while scoring only 6.67% morphological boundary accuracy. PRPE, the morphology-aware tokenizer, reaches 83.33%, the highest of all systems. So the most efficient tokenizer by the standard metric is the worst one morphologically.
The authors validate this extrinsically: on a Spanish-to-Quechua (Chanka) machine translation task, a model trained with PRPE beats an identical BPE-trained model by 68.7% BLEU and 17.0% chrF. The limitation is that it is one language and one translation task, and PRPE needs a morphological analyzer to work.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.