Hybridizing the Incompatible: A Monolingual Model Meets NLLB – Failures, Insights, and Future Paths

This document is an independent research log documenting experimental attempts to hybridize a monolingual language model with NLLB (No Language Left Behind) from Meta. The goal was to create a third, emergent model by allowing internal "genes" (weights, embeddings, tokenization) to interact. The result was unstable, non‑homogeneous, and partially successful (≈50%). The hybrid showed contradictory behavior, collapsed on arithmetic tasks, and exhibited clear dominance of NLLB. Key insight: Hybridizing two ready‑made models is not merging them. It is like genetic engineering – the host may reject foreign genes unless made compatible. This work is not peer‑reviewed. It is a personal research log shared to invite discussion. All experiments used publicly available tools. No proprietary datasets were used. Correspondence for academic purposes only: [email protected]