General-purpose large language models perform substantially worse on Bengali downstream tasks than on equivalent English tasks, with gaps consistent with data-volume explanations¶

Claim · All claims
Evidence · 1 study · review r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Benchmark evaluations including BenLLM-Eval and a 2025 evaluation of ten open-source LLMs show Bengali performance lower across model families than equivalent English performance. → Avijit Roy 2026

Evidence¶

Avijit Roy 2026¶

Avijit Roy, Proma Roy. (2026). Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages — A Case Study in Bengali, the Digital Divide, and the Hidden Cost of English-Centric Design. arXiv preprint. https://arxiv.org/abs/2608.12278 doi:10.48550/arxiv.2608.12278

q2 · i? · review · r2

Review of published benchmark evaluations (BenLLM-Eval, Kabir et al. 2024; Bhowmik et al. 2025, a ten-model Bengali/English comparison). The review reports consistent performance gaps relative to English equivalents; no effect sizes are printed in this article.

"A 2025 evaluation of ten open-source LLMs across Bengali and English benchmarks likewise showed that Bengali performance was lower across model families, particularly for smaller models like Mistral, and that the gaps correlated with tokenization efficiency and model scale."

Discussion¶

Related Claims¶

• Bengali's alphasyllabary script incurs a tokenization penalty: elevated token fertility fragments linguistic units and compounds the training data deficit — related
• Individual internet penetration in rural Bangladesh (36.5%) is about half the urban rate (71.4%), which one case-study preprint argues puts cloud-dependent AI tools out of reach for rural learners