Do LLMs Have Habits in Self-Evaluation? Prototypical Self-Evaluation Enhanced Named Entity Recognition for LLMs

Named Entity Recognition (NER) is a funda mental task in information extraction, and Large Language Models (LLMs) have demonstrated strong zero-shot per formance on this task. However, LLM-based NER often suffers from type misidentification (TM), particularly in single target-type settings, where entity spans are correctly identified but assigned incorrect entity types. To mitigate this issue, we prompt LLMs to identify both target and non-target entities and to self-evaluate the relevance of each extracted entity to the target type. Experiments across multiple LLMs show that the self-evaluation rating distributions of correctly typed entities and TM entities form clearly separable clusters. Based on this finding, we propose PFILTER, a self-evaluation-based filtering method that constructs rating prototypes from historical genera tions and filters predictions according to the relationship between entity ratings and these prototypes. Depending on whether one or two additional dialogue turns are used, we introduce two variants: PFILTER-S and PFILTER D. Experiments on multiple datasets demonstrate that PFILTER reduces TM up to 32.1% for GPT-4.1 and 38.8% for GPT-3.5-Turbo, while remaining highly compatible with other prompt-engineering techniques. Our code is available at https://github.com/GTZN2/PFilter.