GLiNER-Relex: A Unified Framework for Joint Named Entity Recognition and Relation Extraction is one of the most product-facing papers in this group. It extends the GLiNER family into joint NER and relation extraction with arbitrary entity and relation labels specified at inference time.

The model uses a shared bidirectional encoder over text, entity labels, and relation labels. Entity spans are recognized, entity-pair representations are constructed, and relation labels are scored against those pairs.

Why we're excited

The appeal is usability: one model, arbitrary label text, and no separate relation extraction stage. That is a much friendlier interface for practitioners than training a bespoke extractor for every schema.

The paper evaluates on CoNLL04, DocRED, FewRel, and CrossRE, and positions the model as computationally efficient relative to large generative systems.

Our community read

This is best read as a generalist extractor. Its flexibility is the point, and that flexibility will often be more valuable than benchmark-specialized peak performance.

The unresolved question is whether zero-shot convenience can close the gap on dense relation taxonomies and long-document evidence chains without losing the model's lightweight deployment story.

Source

• arXiv: 2605.10108