ACL-Verbatim: hallucination-free question answering for research

Two 150M-parameter ModernBERT extractors have been released for question answering over retrieved text.

They aim to return exact supporting spans from source documents, reducing hallucinations versus generated answers.

The models reportedly outperform extractive baselines and even larger LLM-based extractors on multiple benchmarks.

This offers a smaller, faster option for grounded QA and RAG systems that need verifiable evidence.