Entity disambiguation (ED) links ambiguous mentions in text to entries in a knowledge base and is a core task in entity linking systems. While pretrained decoder-only language models (DLMs) offer strong generalization capabilities, their effective use in ED has been restricted due to sensitivity to candidate order, susceptibility to hallucinated outputs, and potential dataset leakage. We introduce Agnus a zero-shot ED framework that addresses these challenges through three core innovations: (1) order-invariant candidate encoding via shared positional embeddings and modified autoregressive attention masking, which eliminates bias on input ordering; (2) constrained decoding that ensures outputs are restricted to valid candidates, effectively preventing hallucinations; and (3) synthetic dataset creation approach as a diagnostic tool for data contamination detection and mitigation. Agnus eliminates up to 15.2% of F1 variability caused by candidate permutations, delivering consistent and order-robust predictions previously unattainable with autoregressive architectures. In our experiments, Agnus achieves state-of-the-art performance on four standard ED benchmarks, surpassing prior zero-shot approaches by an average 3.7% using small language models. We release code, data including candidate sets, and a synthetic benchmark to support reproducibility and controlled evaluation.

A major domain of research in natural language processing is named entity recognition and disambiguation (NERD). One of the main ways of attempting to achieve this goal is through use of Semantic Web technologies and its structured data formats. Due to the nature of structured data, information can be extracted more easily, therewith allowing for the creation of knowledge graphs. In order to properly evaluate a NERD system, gold standard data sets are required. A plethora of different evaluation data sets exists, mostly relying on either Wikipedia or DBpedia. Therefore, we have extended a widely-used gold standard data set, KORE 50, to not only accommodate NERD tasks for DBpedia, but also for YAGO, Wikidata and Crunchbase. As such, our data set, KORE 50ˆDYWC, allows for a broader spectrum of evaluation. Among others, the knowledge graph agnosticity of NERD systems may be evaluated which, to the best of our knowledge, was not possible until now for this number of knowledge graphs.