利用实体幻觉索引微调大型语言模型进行文本摘要
January 9th, 2026
我们提出了一种基于强化学习的微调方法，用于大型语言模型，利用恩蒂幻觉指数（EHI）作为奖励信号，以减少文本摘要中的实体级幻觉。会议记录的实验表明，这种方法提高了实体的忠实性和准确性。
本研究利用实体幻觉指数作为奖励信号，以减少AI总结文本中错误的实体生成。当前的指标有两种粗略指标，该协议采用细粒度实体奖励以提升事实性，无需人工标签。首先，使用两个主要语料库，包括多回合会议文字记录的对话转录数据集和XSum News摘要基准。
清理每个数据集，将数据平整，并划分为80%训练集、10%验证集和10%测试集。使用基于变换器的 NER 模型 D slim/基于 BERT 的 NER 进行标记化和实体提取，以确保稳健的接地。使用宽度为四的光束搜索生成零射点模式的初始摘要，以估算基线幻觉流行率。
计算实体幻觉指数，作为一个无参考的实体层级指标，将实体分类为五个因素以量化忠诚度。奖励抽取行为和积极幻觉，惩罚负面幻觉、过度专注的实体和注意力的迷失。定义奖励函数与实体忠实度成正比，并在生成的汇总中最大化预期EHI奖励。
应用强化样式更新机制，优化模型参数以最大化实体忠实性。使用低秩适配实现参数高效的微调，只更新LoRA适配器，同时保持基础模型权重冻结。然后应用策略梯度更新以鼓励更高的奖励输出，并每500次更新重新生成总结以刷新奖励估计。
最后，利用微调模型生成最终总结，并与基线结果进行比较。使用Rouge 1、Rouge 2和Rouge L评估模型的信息量，使用事实CC和CAGs评估流畅性和事实一致性。EHI微调减少了幻觉，提升了模型和数据集间的事实性，EHI右移，事实CC分布也证实了实体忠实度和事实一致性的提升，而CAGS分数基本保持不变。
经过微调后，Mistral中负面、过度聚焦和失去焦点等有害幻觉类型减少，而提取性因子几乎翻倍。相关分析显示，EHI和CAGS在Mistral和DistilBART中相关性较弱，但Flan-T5的正相关性更强。研究人员可以通过自动EHI奖励信号测量实体层面的忠实度和幻觉率。
最关键的挑战是NER的准确性，因为提取误差可能在培训过程中传递到错误的奖励计算中。未来研究可以整合共引用解析和信息性奖励，更好地平衡忠实性与摘要覆盖。
View the full transcript and gain access to thousands of scientific videos
Overview
This article presents a novel reward-driven fine-tuning framework for large language models (LLMs) aimed at reducing entity-level hallucinations in abstractive summarization. By leveraging the Entity Hallucination Index (EHI) as a guiding metric, the approach enhances the factual accuracy of generated summaries without sacrificing informativeness or generalizability.
Key Study Components
Area of Science
- Natural Language Processing
- Machine Learning
- Text Summarization
Background
- LLMs have significantly improved abstractive summarization quality.
- Entity-level hallucination, where incorrect or fabricated entities appear in summaries, remains a major challenge.
- Existing models often struggle to balance informativeness and factual accuracy.
- Metrics to directly penalize hallucinated entities are underexplored in model training.
Purpose of Study
- To develop a fine-tuning framework that reduces entity hallucination in summarization models.
- To introduce and utilize the Entity Hallucination Index (EHI) as a reward signal.
- To demonstrate improved factuality and robustness in LLM-generated summaries.
Methods Used
- Initial summaries generated using pre-trained models (e.g., Flan-T5, DistilBART, Mistral) on structured transcript datasets such as XSUM.
- Named entity extraction from both generated summaries and gold reference summaries.
- Calculation of EHI by evaluating entity precision and penalizing fabricated entities.
- Reinforcement learning fine-tuning using a REINFORCE-style update mechanism with EHI as the reward signal.
Main Results
- Fine-tuned models exhibited lower rates of entity hallucination.
- Informativeness of summaries was maintained despite reduced hallucination.
- EHI-guided models demonstrated better generalization on out-of-domain summarization tasks.
- Enhanced robustness and factual accuracy were observed in experimental evaluations.
Conclusions
- The EHI-based reward framework effectively reduces entity-level hallucinations in LLM summarization.
- Factuality and informativeness can be jointly optimized using this approach.
- The method offers a practical direction for improving the reliability of automated summarization systems.
Frequently Asked Questions
Executive Industry Relevance
Entity-level hallucination in large language model (LLM) summarization poses a significant risk to data integrity and decision-making in biopharma R&D. The Entity Hallucination Index (EHI)-guided fine-tuning framework directly addresses this challenge, enabling more reliable extraction of factual insights from scientific and operational text sources. Enhanced factuality in automated summarization supports higher predictive confidence and reduces downstream risk in knowledge-driven pipelines.
Strategic Applications in Biopharma R&D
Early Discovery & Target Validation
- Improves accuracy of literature and data mining for target identification by minimizing fabricated entities.
- Supports mechanistic de-risking by ensuring extracted entities reflect true biological relationships.
- Enables more confident hypothesis generation from automated text summarization outputs.
Screening & Assay Development
- Facilitates standardized extraction of assay targets and conditions from diverse documentation sources.
- Reduces risk of propagating incorrect entity information into screening workflows.
- Supports reproducibility by maintaining entity fidelity across automated summaries.
Translational & Preclinical Research
- Enhances mapping of preclinical findings to disease-relevant entities by reducing hallucination.
- Improves continuity of entity tracking from discovery through preclinical validation phases.
- Supports risk-adjusted advancement by providing more reliable entity-level evidence.
Pipeline & Workflow Integration
The EHI-guided fine-tuning framework integrates into the informatics layer of discovery, screening, and translational research pipelines, supporting robust entity extraction and summarization.
- Discovery Biology: Enables hypothesis testing and pathway clarification by ensuring entity accuracy in mined summaries.
- Screening: Provides reproducible, entity-faithful summaries for assay documentation and compound evaluation.
- Analytics: Delivers quantitative EHI scores to benchmark and compare summarization outputs across conditions.
- Translational Research: Maintains entity fidelity for biomarker and disease mapping when moving from discovery to preclinical stages.
- Enterprise Reuse: Establishes a scalable, reward-driven summarization capability for diverse biopharma text sources.
Operational & Enterprise Impact
- Scientific Value: Increases predictive confidence and reduces mechanistic ambiguity in text-derived insights.
- Operational Value: Standardizes entity extraction and summary generation across teams and projects.
- Strategic Value: Supports better go/no-go decisions by improving the reliability of automated knowledge extraction.
- Portfolio Impact: Enables risk-adjusted prioritization by reducing the propagation of incorrect entity data.
Implementation Considerations
- Requires expertise in natural language processing and reinforcement learning for model fine-tuning.
- Needs access to structured transcript datasets and gold-standard reference summaries for EHI computation.
- Demands cross-team agreement on entity definitions and extraction standards.
- May require adaptation of EHI thresholds for different biomedical domains or document types.
- Dependent on the quality of initial pre-trained models and reference data for optimal performance.
