Jingyu Wang

Research profile

Research summary

Jingyu Wang is a natural language processing researcher at Beijing Institute of Technology whose cataloged work centers on named entity recognition (NER) in low-resource and cross-domain settings, using large language models. The Evolve catalog contains two 2026 arXiv preprints, both with Wang as first author (non-corresponding). The first, URA-NER, presents a training-free, unified retrieval-augmented framework with retrieval alignment and uncertainty reduction; Evolve's review notes it lifts Qwen2.5-14B to state-of-the-art low-resource NER on CrossNER, with good ablations but no significance tests or leakage controls. The second quantifies entity type confusion in cross-domain NER via a multidimensional measure (MCQM) and adds bidirectional LLM reasoning (PBRC), reporting strong CrossNER gains; reviewers flag SOTA-on-all-domains claims that exceed its own Table 2. Bibliometric signals reflect an early, concentrated profile: the field distribution contains only natural language processing (focus score 100), the co-author network is small (4 distinct coauthors; collaboration score 32) with repeated collaboration alongside Shijie Wu and Fusheng Jin, and mean reproducibility readiness across the two assessed papers is 38/100 (reproducibility score 38). Output, impact, and activity axes remain unscored for lack of data. Overall, this is an early-stage, method-focused profile whose published claims are promising but would benefit from stronger statistical evaluation and reproducibility practices.

Research keywords

Highlights

• URA-NER: training-free retrieval augmentation reaches low-resource NER SOTA with Qwen2.5-14B

  Evolve's review reports that this unified retrieval-augmented framework, combining retrieval alignment and uncertainty reduction, lifts Qwen2.5-14B to state-of-the-art low-resource NER on CrossNER without task-specific training. Ablations are strong; reviewers flag the absence of significance tests and leakage controls. Reproducibility readiness: medium.

• Quantifying entity type confusion and adding bidirectional LLM reasoning for cross-domain NER

  The paper introduces MCQM, a multidimensional quantification of entity type confusion, and PBRC, an LLM bidirectional reasoning enhancement, reporting strong CrossNER gains in cross-domain, low-resource transfer. Evolve's review cautions that SOTA-on-all-domains claims exceed what the paper's own Table 2 supports. Reproducibility readiness: low.

Timeline

1. 2026First-authored 'URA-NER: A Unified Retrieval-Augmented Framework with Retrieval Alignment and Uncertainty Reduction for Low-Resource NER' appears on arXiv and enters the Evolve review catalog (reproducibility readiness: medium).
2. 2026First-authored 'Mitigating Entity Type Confusion in Cross-Domain NER via Multidimensional Quantification and Reasoning Enhancement', introducing MCQM and PBRC, appears on arXiv and is reviewed in the Evolve catalog (reproducibility readiness: low).
3. 2026Both cataloged papers establish a research line squarely on retrieval-augmented, LLM-based entity recognition, giving the profile a field distribution concentrated entirely in natural language processing (focus score 100).

Papers in the Evolve catalog

• 2026-09-21first author

  URA-NER: A Unified Retrieval-Augmented Framework with Retrieval Alignment and Uncertainty Reduction for Low-Resource NER

• 2026-09-21first author

  Mitigating Entity Type Confusion in Cross-Domain NER via Multidimensional Quantification and Reasoning Enhancement

Frequent co-authors