Paper deep dive
AOP-Smart: A RAG-Enhanced Large Language Model Framework for Adverse Outcome Pathway Analysis
Qinjiang Niu, Lu Yan
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/14/2026, 2:26:22 AM
Summary
AOP-Smart is a Retrieval-Augmented Generation (RAG) framework designed to improve the reliability of Large Language Models (LLMs) in toxicological research by integrating structured data from AOP-Wiki. By utilizing Key Events (KEs), Key Event Relationships (KERs), and AOP pathways as external context, the framework significantly reduces hallucinations and improves accuracy in complex AOP-related reasoning tasks.
Entities (6)
Relation Signals (3)
AOP-Smart → utilizes → AOP-Wiki
confidence 100% · Based on the official XML data from AOP-Wiki, this method uses Key Events (KEs)...
AOP-Smart → enhances → Large Language Model
confidence 95% · AOP-Smart can significantly alleviate the hallucination problem of large language models
DeepSeek → achievedaccuracy → 100%
confidence 90% · The overall accuracy of DeepSeek increased to 100% after enhancement
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Adverse Outcome Pathways (AOPs) are an important knowledge framework in toxicological research and risk assessment. In recent years, large language models (LLMs) have gradually been applied to AOP-related question answering and mechanistic reasoning tasks. However, due to the existence of the hallucination problem, that is, the model may generate content that is inconsistent with facts or lacks evidence, their reliability is still limited. To address this issue, this study proposes an AOP-oriented Retrieval-Augmented Generation (RAG) framework, AOP-Smart. Based on the official XML data from AOP-Wiki, this method uses Key Events (KEs), Key Event Relationships (KERs), and specific AOP information to retrieve relevant knowledge for user questions, thereby improving the reliability of the generated results of large language models. To evaluate the effectiveness of the proposed method, this study constructed a test set containing 20 AOP-related question answering tasks, covering KE identification, upstream and downstream KE retrieval, and complex AOP retrieval tasks. Experiments were conducted on three mainstream large language models, Gemini, DeepSeek, and ChatGPT, and comparative tests were performed under two settings: without RAG and with RAG. The experimental results show that, without using RAG, the accuracies of GPT, DeepSeek, and Gemini were 15.0\%, 35.0\%, and 20.0\%, respectively; after using RAG, their accuracies increased to 95.0\%, 100.0\%, and 95.0\%, respectively. The results indicate that AOP-Smart can significantly alleviate the hallucination problem of large language models in AOP knowledge tasks, and greatly improve the accuracy and consistency of their answers.
Tags
Links
- Source: https://arxiv.org/abs/2604.10874v1
- Canonical: https://arxiv.org/abs/2604.10874v1
Trouble viewing inline? Open PDF directly →
Full Text
28,098 characters extracted from source content.
Expand or collapse full text
AOP-Smart: A RAG-Enhanced Large Language Model Framework for Adverse Outcome Pathway Analysis Qinjiang Niu 1 Lu Yan 1,∗ 1 Nanyang Normal University, Nanyang, China *Corresponding author: Lu Yan Abstract Adverse Outcome Pathways (AOPs) are an important knowledge framework in toxicological research and risk assessment. In recent years, large language models (LLMs) have gradually been applied to AOP-related question answering and mechanistic reasoning tasks. However, due to the existence of the hallucination problem, that is, the model may generate content that is inconsistent with facts or lacks evidence, their reliability is still limited. To address this issue, this study proposes an AOP-oriented Retrieval-Augmented Generation (RAG) framework, AOP-Smart. Based on the official XML data from AOP-Wiki, this method uses Key Events (KEs), Key Event Relationships (KERs), and specific AOP information to retrieve relevant knowledge for user questions, thereby improving the reliability of the generated results of large language models. To evaluate the effectiveness of the proposed method, this study constructed a test set containing 20 AOP-related question answering tasks, covering KE identification, upstream and downstream KE retrieval, and complex AOP retrieval tasks. Experiments were conducted on three mainstream large language models, Gemini, DeepSeek, and ChatGPT, and comparative tests were performed under two settings: without RAG and with RAG. The experimental results show that, without using RAG, the accuracies of GPT, DeepSeek, and Gemini were 15.0%, 35.0%, and 20.0%, respectively; after using RAG, their accuracies increased to 95.0%, 100.0%, and 95.0%, respectively. The results indicate that AOP-Smart can significantly alleviate the hallucination problem of large language models in AOP knowledge tasks, and greatly improve the accuracy and consistency of their answers. Keywords: Adverse Outcome Pathway; large language model; retrieval-augmented generation; knowledge en- hancement; hallucination problem 1. Introduction Adverse Outcome Pathways (AOPs) are a conceptual framework used to describe the causal relationship from Molecular Initiating Events (MIEs) to Adverse Out- comes (AOs), and are of great significance in toxico- logical research and chemical risk assessment [1]. By organizing Key Events (KEs) at the molecular, cellu- lar, tissue, organ, and individual levels, AOPs provide systematic support for the expression of toxicity mech- anisms, knowledge integration, and risk judgment [2]. With the continuous development of AOP-Wiki, the number of related knowledge entries has been increas- ing, and the content covers more species, biological lev- els, and complex Key Event Relationships (KERs). Al- though the expansion of knowledge scale improves the richness of AOP resources, it also increases the diffi- culty for researchers in retrieving, understanding, and integrating information. To improve the usability of AOP knowledge, previous studies have developed various supporting tools, such as AOP-HelpFinder and AOP-Explore, to support the mining, retrieval, and visualization analysis of AOP- related information [3, 4]. At the same time, large language models (LLMs) have shown strong capabil- ities in tasks such as natural language understanding, knowledge question answering, and complex reasoning, and have demonstrated broad application prospects in biomedical text mining and mechanistic knowledge analysis [5, 6, 7]. Existing studies have shown that LLMs can assist in the extraction, integration, and structured representation of scientific knowledge, and can help with the construction of AOP chains to a cer- tain extent [8]. However, the generated results of LLMs are limited by the coverage of training corpora and the generation mechanism itself, and they are prone to producing hallucinations when dealing with highly specialized domain knowledge, that is, outputs that are fluent in expression but factually incorrect, thereby af- fecting their reliability and interpretability in scientific research [9, 10]. Retrieval-Augmented Generation (RAG) provides an effective idea for alleviating the above problems. By introducing external knowledge before LLMs generate content, RAG provides knowledge relevant to the ques- tion as context to the model, thereby enhancing the factual basis of model responses and reducing the risk of hallucination [11]. A typical RAG framework usu- ally includes a retrieval module and a generation mod- ule: the former obtains relevant content from exter- nal knowledge sources, and the latter generates an- swers based on the retrieved results [12, 13]. Most 1 arXiv:2604.10874v1 [cs.CL] 13 Apr 2026 existing RAG methods are based on vector databases and semantic similarity retrieval, and have achieved good results in open-domain question answering and knowledge-intensive tasks. However, research on RAG in the AOP domain is still relatively limited. AOP knowledge is different from general unstructured text; it has significant hierarchical structure, causal relationships, and network character- istics, which place higher requirements on knowledge organization and retrieval methods. Existing studies still lack RAG designs oriented to the causal relation- ship characteristics of AOPs, especially in how to ef- fectively utilize KEs, KERs, and overall AOP struc- tural information to support complex AOP mechanistic knowledge modeling and reasoning. Based on this, this paper proposes an AOP analysis- oriented RAG framework, AOP-Smart. This method uses the XML data provided by AOP-Wiki to con- struct a knowledge base and index files, and designs an indexing and knowledge expansion mechanism based on question-related KEs to achieve joint retrieval and knowledge integration of KE, KER, and AOP informa- tion. On this basis, the KE, KER, and AOP knowledge related to the question are used as external context and input into LLMs, thereby improving the reliability of LLM results. The main contributions of this study are as follows: (1) A RAG framework for the AOP domain is pro- posed, which introduces AOP knowledge into the reasoning process of LLMs; (2) A knowledge expansion strategy based on question-related KEs is designed to achieve as- sociated retrieval and knowledge supplementation among KEs, KERs, and AOPs; (3) Comparative experiments verify the effectiveness of this method in alleviating the hallucination problem in the AOP domain and improving an- swer accuracy. 2. Methods 2.1 Overall Framework of AOP-Smart To ensure the consistency of data sources and the re- producibility of the experimental process, this study conducted experiments based on the stable XML snapshot of the AOP-Wiki released on 2026-01- 01.This dataset was obtained from the official AOP-Wiki quarterly “Permanent Downloads” archive (https://aopwiki.org/downloads ), which provides ver- sioned XML dumps for citation and reproducible re- search. The 2026-01-01 release was selected to guaran- tee data stability, and all contents were subsequently parsed and processed in a unified manner. It should be noted that the constructed system does not depend on a specific version of the data, but has good scalability and updating capability. By replacing the XML file of the corresponding version, the knowl- edge base can be updated. As shown in Figure 1(A), without introducing the RAG mechanism, the user question is directly input into the LLM, and the model answers only relying on its internal parameters, lacking the support of exter- nal knowledge, and thus is prone to problems such as missing information or factual errors. As shown in Figure 1(B), the AOP-Smart frame- work introduces a knowledge retrieval process prior to response generation. Specifically, the official XML data from AOP-Wiki is first parsed into two structured resources: Index.txt and AOP-Smart.json. Among them, Index.txt stores the IDs and titles of all key events (KEs) and is used for efficient preliminary candi- date retrieval. This index is extremely lightweight, con- taining only approximately 20,000 tokens, which is neg- ligible compared to the context window of modern large language models (which can reach nearly one million tokens). Therefore, it can be directly loaded into the prompt without introducing significant computational overhead, while also enabling easy extensibility for fu- ture AOP-Wiki updates. In contrast, AOP-Smart.json stores detailed structured information about KEs, key event relationships (KERs), and adverse outcome path- ways (AOPs), supporting downstream reasoning and generation tasks. During the inference stage, the system first inputs the user question together with the content in Index.txt into the LLM, and ranks the relevance of all KEs through prompt engineering, from which the Top-N most relevant KEs are selected (N is a tunable pa- rameter). Subsequently, based on the IDs of the se- lected KEs, the corresponding KE details are retrieved from AOP-Smart.json, and their related KER and AOP information are further expanded to construct a structured context representation (including KE Detail, KER Detail, and AOP Detail). Finally, this structured context and the user ques- tion are jointly input into the LLM for generative rea- soning, thus completing the closed-loop process from question-driven retrieval to knowledge-enhanced gener- ation. This mechanism effectively alleviates the knowl- edge deficiency problem of traditional LLMs in AOP tasks and significantly improves the accuracy and in- terpretability of the generated results. 2.2 Expansion Method Based on TOP-N KE IDs After the LLM completes the preliminary screening of KEs, the system performs expansion based on the se- lected KEs to enhance the coverage of question-related knowledge. The related process is shown in Figure 2. The specific process is as follows. For each selected KE, the system first extracts its di- rectly upstream and directly downstream KE IDs from AOP-Smart.json, and merges them with the originally selected KE to form an expanded KE set. The pur- pose of this step is to supplement the local mechanistic information causally adjacent to the events explicitly involved in the question. After obtaining the expanded KE set, the system further traverses all Key Event Relationships (KERs) 2 User Question LLM Index KE ID | Describe 195 | Inhibition, NMDARs 350 | Increase, Mortality ... Top-N Relevant KE IDs [57,489,64...] Question Relevant KE Detail KER Detail AOP Detail LLM LLM Answer User Question ALL KE Detail KER Detail AOP Detail AOP-Wiki Filter by KE IDs Index Python (B) With RAG (A) Without RAG User Question LLM LLM Answer N is user-defined Figure 1: Overview of the AOP-Smart framework. (A) In the absence of retrieval augmentation, the LLM directly generates responses based solely on its internal parameters, which may lead to incomplete or hallucinated outputs. (B) With the proposed AOP-Smart framework, user queries are first used to retrieve relevant KE candidates from an indexed KE list, followed by structured expansion to KE, KER, and AOP-level information from the knowledge base. The retrieved structured context is then combined with the user query and fed into the LLM to produce a knowledge-enhanced and more reliable response. LLM Top-N Relevant KE IDs ALL KE Detail KER Detail AOP Detail Upstream KE IDs of the Top-N Relevant KEs Top-N Relevant KE IDs KE DetailAOP DetailKER Detail Downstream KE IDs of the Top-N Relevant KEs KE id title level organ cell description applicability upstream_ke_id downstream_ke_id description applicability AOP id title abstract MIEs_IDKEs_ID AOs_ID Event Relationship applicability User Question Select AOPs with ≥2 KEs (In KE IDs) LLM Answer Python Python Filter by KE IDs N is user-defined Figure 2: Overview of the KE-based expansion process, in- cluding KE augmentation, KER reconstruction, and AOP retrieval, followed by integration with LLM-based reason- ing. in the knowledge base, and selects those relationships that connect any two KEs within the expanded KE set, forming an expanded KER set. Through this joint ex- pansion process, the system can not only identify the event nodes most relevant to the question, but also re- cover the causal relationships among these events, pro- viding a basis for mechanism explanation and relation- ship judgment in subsequent responses. Subsequently, the system scans all AOPs. When at least two matches exist between the MIE, KE, or AO in an AOP and the expanded KE set, the system in- cludes that AOP into the expanded AOP set. This matching rule is intended to reduce the noise caused by accidental matching of a single event and improve the relevance between the retrieval results and the ques- tion. Through this mechanism, the system can fur- ther expand from local event relationships to complete AOP chains, thereby providing higher-level background knowledge support for subsequent question answering. Finally, the system integrates the expanded KE, KER, and AOP information, and inputs them together with the user question into the LLM for subsequent reasoning and answer generation. In this way, the final output of the model is established on the basis of ex- plicit external knowledge support, which helps reduce free generation detached from the facts in the knowl- edge base. 2.3 Engineering Implementation This study implemented a complete software system based on Python, and its interface is shown in Figure 3. The system mainly consists of three functional areas: the configuration area, the user input area, and the LLM output area. In the configuration area, the user needs to input the Service Endpoint and API key of the large model being used. After completing the configuration, clicking the Fetch Models button will obtain the list of currently available large language models, and system initializa- tion is completed by selecting a model. The system also provides adjustable interfaces for multiple key parameters to support flexible control un- der different task requirements. These include: Max Output, which is used to control the maximum number of tokens generated by the model; Temperature, which is used to adjust the randomness and creativity of the 3 Run Stop AOP-Smart API Status : NOT Connected Fetch Models Base URL API Key Model Max Output Temperature TOP N KEs XML Version 2026-01-01 User input LLM Output 4096 0 5 https://api.deepseek.com ************************ Service endpoint Authentication key Enter the Base URL and API Key, click “Fetch Models” to retrieve available models, then select the desired LLM. Max Output : The maximum output token of LLM Temperature : Controlling the stability of LLM output (lower values increase determinism) Top N KEs : Determines how many top-ranked KEs are selected before expanding KE, KER, and AOP information XML Version : The XML version of AOP-Wiki Figure 3: Overview of the Python-based system interface, including configuration, input, and output modules, with adjustable parameters for model selection and KE-based re- trieval control. model output, where a lower value leads to more stable output and a higher value leads to more diverse gen- erated results; and the Top-N KE parameter, which is the core control variable in this study, used to specify the number of Key Events (KEs) screened by the LLM, thereby controlling the expansion scale and information coverage of RAG retrieval. In the user input area, after the user enters the ques- tion to be processed and clicks the RUN button, the system will call the LLM based on the configuration pa- rameters and the expanded knowledge retrieval mech- anism, and output the final result to the output area. The software system developed in this study and the related experimental code have been made public as supplementary materials to support method reproduc- tion. The project code can be accessed via GitHub: https://github.com/qinjiang-lab/AOP-Smart 3. Experimental Design and Results 3.1 Experimental Design To evaluate the performance of large language models on AOP knowledge tasks and verify the effectiveness of the proposed method, this paper constructed an AOP question answering test set containing 20 questions, as shown in Figure 4.The full set of questions is provided in the Supplementary Material. KE Identification Downstream KE retrieval Upstream KE retrieval Complex task AOP QA Evaluation Set (4 tasks ×5 questions) What is the biological event corresponding to KE X, and in which species is it applicable? × 5 What is the KE ID of “X", and what are all its Downstream key events? × 5 What is the KE ID of “X", and what are all its Upstream key events? × 5 For the AOPs where “X" serves as the MIE and “X" serves as the AO, how many such AOPs exist? Briefly list all key events involved and specify the applicable species. × 5 ??? ? Title ? ??? ? ? ? ? ? ? MIE AO KE ID X KE ID X Figure 4: Overview of the 20-question AOP benchmark, grouped into four task categories: KE identification, down- stream KE retrieval, upstream KE retrieval, and complex AOP query. According to different task types, the test set is di- vided into four categories: (1) Key Event Identification (KE Identifica- tion): Given a KE ID, the model is required to identify the corresponding biological event and its applicable species; (2) Downstream Key Event Retrieval (Down- stream KE Retrieval): Given an event name, the model is required to return the corresponding KE ID and list all directly downstream KEs; (3) Upstream Key Event Retrieval (Upstream KE Retrieval): Given an event name, the model is required to return the corresponding KE ID and list all directly upstream Key Events; (4) Complex AOP Query Task (Complex AOP Query): This involves retrieving the complete AOP chain based on a given MIE and AO, and requires listing the related Key Events and their applicable species information. The above tasks cover different types in AOP knowl- edge, such as KE identification, KER retrieval, and complex AOP query, and can evaluate the hallucina- tion problem of LLMs from multiple perspectives. The evaluation metric adopts Accuracy, defined as: Accuracy = N correct N total whereN correct represents the number of completely cor- rect answered questions, andN total represents the total number of questions. In this study, the factual informa- tion in the corresponding entries of AOP-Wiki is used as the basis for judgment, and when the model answer is completely consistent with the standard knowledge, it is judged as correct. 3.2 Experimental Results In this experiment, the Top-N parameter was uniformly set to 5, and Temperature was set to 0 to ensure the retrieval scale and the determinism of the generated results. The specific models used were DeepSeek-V3.2, GPT-5.4, and Gemini-3.1 Pro, respectively. The exper- imental results are shown in Table 1. Without introducing RAG, the performance of the three large language models on AOP-related question answering tasks showed obvious differences. Among them, the overall accuracies of DeepSeek, Gemini, and GPT were 35.0%, 20.0%, and 15.0%, respectively, indi- cating that it is difficult to stably complete AOP tasks by relying only on the internal knowledge of the models. AfterintroducingtheAOP-Smartretrieval- augmented framework proposed in this paper, the performance of each model on all tasks was signif- icantly improved. The overall accuracy of DeepSeek increased to 100% after enhancement, while Gemini and GPT increased to 95.0% and 95.0%, respectively, verifying the effectiveness of RAG in improving the ability of models to utilize external knowledge. From the task-level analysis, RAG has a consistent improvement effect on different types of tasks. In the Key Event Identification (Task 1) and Downstream 4 Table 1: Experimental results of different large language models on AOP-related tasks with and without RAG enhance- ment. TaskDescriptionDeepSeekGeminiGPTOverall W/O RAG W/ RAG W/O RAG W/ RAG W/O RAG W/ RAG W/O RAG W/ RAG Task 1KE Identification80%100%60%100%60%100%66.67%100% Task 2Downstream KE retrieval20%100%20%100%0%100%13.33%100% Task 3Upstream KE retrieval20%100%0%80%0%100%6.67%93.33% Task 4Complex task20%100%0%100%0%80%6.67%93.33% OverallAll tasks combined35%100%20%95%15%95%23.33%96.67% Key Event Retrieval (Task 2) tasks, all three models achieved accuracies close to or reaching 100%; in the Upstream Key Event Retrieval (Task 3) and Complex AOP reasoning task (Task 4), RAG also significantly improved the performance of the models in multi-hop structural relationship queries, but compared with ba- sic retrieval tasks, there were still certain differences in difficulty. These results indicate that the RAG method can ef- fectively improve the stability and reliability of large language models in professional toxicological knowledge question answering, and significantly alleviate the hal- lucination problem of LLMs. 4. Limitations Although the AOP-Smart framework proposed in this study has achieved relatively significant experimental results in AOP-related tasks, there are still certain lim- itations. First, the current method has not yet introduced in- formation such as evidence level, evidence strength, and confidence in the AOP knowledge system. Therefore, in complex reasoning processes, it is not possible to perform weighted processing on knowledge from differ- ent sources, which may affect the reliability and inter- pretability of the final results. Second, when constructing the retrieval knowledge base, this paper performed necessary text filtering and truncation processing on the KE, KER, and AOP de- scriptive information in AOP-Wiki to adapt to the in- put length limitation of large language models. Al- though this strategy improves computational efficiency, it may also lead to the loss of some contextual seman- tic information, thereby affecting the completeness of answers in complex query tasks. Third, the evaluation set constructed in this study is relatively limited in scale (20 questions), and is mainly based on manually designed query tasks. Although it can cover typical KE identification and AOP relation- ship retrieval scenarios, it is still insufficient in cover- ing larger-scale, multi-hop reasoning, and open-domain AOP question answering tasks. Finally, different large language models are some- what sensitive to prompt format and context organi- zation methods. Therefore, the experimental results may be jointly affected by both model capability and prompt engineering design, and this factor has not yet been systematically analyzed through ablation in this study. 5. Conclusion and Future Work This study proposes a retrieval-augmented generation framework for Adverse Outcome Pathway (AOP) anal- ysis, AOP-Smart. This method constructs a struc- tured knowledge base based on the official XML data of AOP-Wiki, and through an expansion mechanism based on Key Events (KEs) and Key Event Relation- ships (KERs), introduces AOP hierarchical structural information into the reasoning process of large language models, thereby enhancing their knowledge acquisition capability in professional toxicological question answer- ing tasks. The experimental results show that, on 20 structured AOP question answering tasks, AOP-Smart signifi- cantly improves the overall performance of DeepSeek, GPT, and Gemini under the no-RAG condition, and achieves near-optimal or optimal accuracy in most task types. Among them, in the complex AOP query task, the RAG mechanism has a clear enhancement effect on multi-hop structural relationship reasoning. The above results indicate that the retrieval- augmented method based on structured AOP knowl- edge can effectively improve the stability and accuracy of large language models in professional-domain ques- tion answering tasks, and alleviate the model halluci- nation problem to a certain extent. Future work will be carried out from the following three aspects. First, expand the scale of evaluation data and construct a larger-scale and multi-type AOP question answering benchmark to improve the compre- hensiveness and robustness of evaluation. Second, in- troduce structured information such as evidence level and confidence, and perform weighted modeling on re- trieval results to enhance the interpretability of the rea- soning process. Finally, explore the application poten- tial of AOP-Smart in automatic information extraction from scientific literature and AOP knowledge graph completion, so as to provide support for the automated construction and expansion of the AOP system. References [1] Gerald T. Ankley, Richard S. Bennett, Rus- sell J. Erickson, Dale J. Hoff, Michael W. Hor- nung, Rodney D. Johnson, David R. Mount, John W. Nichols, Christine L. Russom, Patricia K. 5 Schmieder, Jose A. Serrrano, Joseph E. Tietge, and Daniel L. Villeneuve. Adverse outcome path- ways: A conceptual framework to support eco- toxicology research and risk assessment. Environ- mental Toxicology and Chemistry, 29(3):730–741, November 2009. [2] Anna Bal-Price and M.E. (Bette) Meek. Adverse outcome pathways: Application to enhance mech- anistic understanding of neurotoxicity. Pharmacol- ogy & Therapeutics, 179:84–95, November 2017. [3] Florence Jornod, Thomas Jaylet, Ludek Blaha, Denis Sarigiannis, Luc Tamisier, and Karine Au- douze. Aop-helpfinder webserver: a tool for com- prehensive analysis of the literature to support ad- verse outcome pathways development. Bioinfor- matics, 38(4):1173–1175, October 2021. [4] Saurav Kumar, Deepika Deepika, Karin Slater, and Vikas Kumar.Aopwiki-explorer: An in- teractive graph-based query engine leveraging large language models. Computational Toxicology, 30:100308, June 2024. [5] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information pro- cessing systems, 30, 2017. [6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural infor- mation processing systems, 33:1877–1901, 2020. [7] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234–1240, Septem- ber 2019. [8] Haochun Shi and Yanbin Zhao. Integration of ad- vanced large language models into the construction of adverse outcome pathways: Opportunities and challenges. Environmental Science & Technology, 58(35):15355–15358, August 2024. [9] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, An- drea Madotto, and Pascale Fung. Survey of hal- lucination in natural language generation. ACM Computing Surveys, 55(12):1–38, March 2023. [10] Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th annual meet- ing of the association for computational linguistics (volume 1: long papers), pages 3214–3252, 2022. [11] Patrick Lewis, Ethan Perez, Aleksandra Pik- tus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K ̈uttler, Mike Lewis, Wen-tau Yih, Tim Rockt ̈aschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Ad- vances in neural information processing systems, 33:9459–9474, 2020. [12] Danqi Chen, Adam Fisch, Jason Weston, and An- toine Bordes. Reading wikipedia to answer open- domain questions.In Proceedings of the 55th Annual Meeting of the Association forComputa- tional Linguistics (Volume 1: Long Papers), page 1870–1879. Association for Computational Lin- guistics, 2017. [13] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In Proceed- ings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 6769–6781, 2020. 6