Neural Recovery of Historical Lexical Structure in Bantu Languages from Modern Data

The deeper provocation here is epistemological: the model was never trained on historical data, yet it surfaced structure that took comparative linguists generations to reconstruct manually. That raises a genuine question about what transformers are actually learning when they process morphological patterns, and whether the representations encode something closer to linguistic universals than task-specific features.

Recent Modelwire coverage has focused heavily on what transformer architectures can and cannot represent internally. The 'Stability and Generalization in Looped Transformers' paper from mid-April examined fixed-point behavior and the conditions under which transformer representations become meaningfully input-dependent. That theoretical framing is directly relevant here: the Bantu reconstruction result is essentially an empirical data point about what falls out of a well-trained transformer's latent space when the training signal is rich and structured. This work belongs to a quieter but important thread in NLP research, one concerned with low-resource and endangered language documentation, which has been largely absent from recent coverage on this site.

Watch whether the BantuMorph team applies the same pipeline to verb stem reconstruction, where Proto-Bantu evidence is thinner and validation against the existing database would be harder to game. If accuracy holds above 80% on that subset, the method has real generalization; if it drops sharply, the noun cognate result may reflect data density rather than genuine historical inference.