大语言模型(LLMS)F中减轻幻觉的策略.pptx

1、,Strategies To Mitigate Hallucinations In Large Language Models(LLMs),LLM Hallucinations?,Examples?,https:/,Why LLMs hallucinate?,Source-Reference DivergenceExploitation through Jailbreak PromptsReliance on Incomplete or Contradictory DatasetsOverfitting and Lack of NoveltyGuesswork from Vague or In

2、sufficiently Detailed PromptsRead more at:https:/,Why LLMs hallucinate?(Contd.),Missing Content-Retrieval strategy didnt workMissing Top Ranked Answer is in DB(vector)but didnt rank highNot in Context Correct documents retrieved but not in LLM context window,Mitigation Strategies,Contextual Prompt E

3、ngineering/Tuning,Context,Instructions,Input Examples,Output Format,Contextual Prompt Engineering/Tuning(Contd.),Positive Prompt Framing,Reasoning,Prompt Fine-Tuning,Image Source:Wei et al.(2022),Contextual Prompt Engineering/Tuning(Contd.),Self-Reflection Prompting,Reflection Prompt,Interactive Que

4、ryingRefinement,Retrieval Augmented Generation(RAG),Question,Full Prompt,Response,Retrieval Query,Retrieved Text,User,User,Embedding,Vector Store,I,LLM,Retrieval Augmented Generation(RAG)(Contd.),Retrieval(R):Searching for relevant information from a database or knowledge base.Augmented(A):Enhances

5、retrieved information by summarizing and connecting key points.Generation(G):Utilizes the augmented information to create new,original response.Combines retrieval and generation aspects in language processingEmpowers LLM Model with domain-specific knowledge.Allows for more accurate and context aware

6、 responses.Ensures access to most updated and reliable facts.Lowers computational costs associated with continuous retraining of models.,Utilization of Knowledge Graphs,Structured representations of knowledge.Provides explicit and standardized vocabulary of concepts and offers ri