Research map / Large language models / Reasoning and chain of thought

A Reasoning Paradigm for Named Entity Recognition

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ReasoningNER, which builds a CoT-annotated NER corpus (NER-CoT) from Pile-NER, fine-tunes Qwen3-8B to output reasoning before entities, and refines it with GRPO using F1 and schema rewards.

Problem

Generative LLM-based NER relies on direct semantic pattern matching ("cognitive shortcutting") without explicit, verifiable reasoning, which hurts zero-shot and low-resource generalization.

Approach

Shift NER from implicit label generation to explicit reasoning: distill LLM-generated, validated chain-of-thought rationales into the model via SFT, then optimize the reasoning-plus-answer output with GRPO under a composite F1 and schema-compliance reward.

Key results

• ReasoningNER 8B reaches 72.4 average F1 on seven CrossNER/MIT datasets, versus 60.1 for GPT-4 and 65.4 for Qwen3-8B.
• In zero-shot evaluation on 20 NER benchmarks, ReasoningNER 8B averages 56.8 F1, versus 47.8 for GLiNER-L and 45.7 for UniNER.
• Ablation on Qwen3-1.7B-Base rises from 31.9 to 65.1 average F1, with GRPO adding 2.1 points over SFT with CoT.

Surprising findings

• Training on NER-CoT annotations without CoT already yields a 20.3-point gain over Pile-NER SFT, suggesting annotation quality matters more than the reasoning trace alone.
• The reasoning-based 1.7B model outperforms DeepSeek-R1 32B on the CrossNER/MIT average.

Limitations

• CoT outputs are verbose and redundant, increasing inference latency and token cost.
• Scope limited to NER; relation and event extraction are not addressed.
• Supervised in-domain CoT data was synthesized by a policy model and filtered only for correct predictions, yielding a small CoT fraction.

Experimental setup

Datasets: Pile-NER InstructUIE collection (20 NER datasets) CrossNER MIT-Movie MIT-Restaurant CoNLL03 MultiCoNER22

Benchmarks: CrossNER MIT-Movie MIT-Restaurant InstructUIE 20-dataset suite (supervised and zero-shot) CoNLL03 low-resource MultiCoNER22 cross-lingual

Metrics: span-based micro-F1

Baselines: ChatGPT-3.5 UniNER InstructUIE GoLLIE KnowCoder GPT-4 GLiNER-L B2NER DeepSeek-R1 Qwen3

Builds on: GRPO Pile-NER InstructUIE KnowCoder B2NER

Closest papers by meaning

• UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition (ICLR 2024 · similarity 0.87)
• GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition (AAAI 2025 · similarity 0.86)
• ConsistNER: Towards Instructive NER Demonstrations for LLMs with the Consistency of Ontology and Context (AAAI 2024 · similarity 0.86)
• PaDeLLM-NER: Parallel Decoding in Large Language Models for Named Entity Recognition (NeurIPS 2024 · similarity 0.85)
• Small Language Model Makes an Effective Long Text Extractor (AAAI 2025 · similarity 0.85)
• MCL-NER: Cross-Lingual Named Entity Recognition via Multi-View Contrastive Learning (AAAI 2024 · similarity 0.84)