VEPO: Variable Entropy Policy Optimization for Low-Resource Language Foundation Models

arXiv 2603.19152·medium signal

Proposes VEPO, an RL training framework using verifiable rewards with deterministic structural constraints (sequence length, format consistency, linguistic well-formedness) to improve LLM performance on low-resource languages. A variable entropy mechanism dynamically calibrates between literal fidelity and semantic naturalness via entropy-tempered advantage estimation with asymmetric clipping to prevent policy collapse. Empirical evaluations across 90 FLORES-200, COMET-22, and chrF directions show substantial improvements in tokenization efficiency and translation quality for underrepresented languages.