BenHalluEval: A Multi-Task Hallucination Evaluation Framework for Large Language Models on Bengali

Researchers introduced BenHalluEval, the first dedicated benchmark for hallucination detection in Bengali large language models.

The framework spans question answering, code-mixed QA, summarization, and reasoning, with 12,000 GPT-5.4-generated hallucinated candidates across 12 hallucination types.

It evaluates seven LLMs and adds BenHalluScore, a metric designed to balance hallucination detection with false-positive penalties.

The benchmark fills a major gap for Bengali, a low-resource language that had not been systematically assessed for hallucinations before.