大型语言模型(LLM)中减轻幻觉的策略.pptx

1、Strategies To Strategies To Mitigate Mitigate Hallucinations In Hallucinations In Large Language Large Language ModelsModels(LLMs)(LLMs)LLM Hallucinations?Faithfulness:Generated text is not faithful to the input contextFactualness:Generated text is not factually correct with respect to world knowled

2、geExamples? https:/https:/ LLMs hallucinate?Source-Reference DivergenceExploitation through Jailbreak PromptsReliance on Incomplete or Contradictory DatasetsOverfitting and Lack of NoveltyGuesswork from Vague or Insufficiently Detailed PromptsRead more at:https:/ LLMs hallucinate?(Contd.)Missing Con

3、tent-Retrieval strategy didnt work Missing Top Ranked Answer is in DB(vector)but didnt rank high Not in Context Correct documents retrieved but not in LLM context window Mitigation StrategiesContextual Prompt Engineering/TuningContextInstructionsInput ExamplesOutput FormatContextual Prompt Engineeri

4、ng/Tuning(Contd.)Positive Prompt FramingReasoningPrompt Fine-TuningImage Source:Wei et al.(2022)Contextual Prompt Engineering/Tuning(Contd.)Self-Reflection PromptingReflection PromptInteractive QueryingRefinementRetrieval Augmented Generation(RAG)QuestionFull PromptResponseRetrieval QueryRetrieved T

5、extUserUserEmbeddingVector StoreIILLMRetrieval Augmented Generation(RAG)(Contd.)Retrieval(R):Retrieval(R):Searching for relevant information from a database or knowledge base.Augmented(A):Augmented(A):Enhances retrieved information by summarizing and connecting key points.Generation(G):Generation(G)

6、:Utilizes the augmented information to create new,original response.Combines retrieval and generation aspects in language processing Empowers LLM Model with domain-specific domain-specific knowledge.Allows for more accurate and context aware responses.Ensures access to most updated and reliable fact