RL for Bangla Mathematical Reasoning

Large language models reason well in English but fall sharply behind in Bangla — a language spoken by 230M+ people yet largely missing from the mathematical-reasoning literature. This is an ongoing research project that trains a compact model to reason mathematically in Bangla through a staged curriculum, where each stage adds a progressively finer-grained learning signal. The goal is competitive Bangla math reasoning without frontier-scale model capacity.

Bangla's agglutinative morphology makes correct reasoning traces structurally longer, which quietly confounds the reward signals standard reinforcement-learning methods rely on. The approach is designed to separate “was the answer correct” from “was it wordy,” so valid Bangla reasoning isn't penalised for length. (Method details are held back while the work is under submission.)