Tsinghua University's "DeepSeek and AI Illusions" (PDF file) - AI Tutorial Materials

DeepSeek andAI"Illusion" is a project of the New Media Research Center of the School of Journalism and Communication, Tsinghua University, andartificialintelligentWritten by Dr. Zhang Jiacheng, a postdoctoral researcher at the college, the document discusses in detail...AIThe definition, causes, assessment methods, and coping strategies of hallucinations are analyzed.AIThe potential risks and creative value of hallucinations in specific scenarios.AIIllusions arise from model-generated content that contradicts reality or exhibits logical inconsistencies, and are related to factors such as data bias, generalization difficulties, knowledge rigidity, and misunderstanding of intent. A proposed approach utilizes networked search and dual...AIverify,Prompt wordsEngineering and other technical means to mitigateAIHallucination. In-depth analysis.AIThe multifaceted nature of illusion emphasizes the importance of finding a balance between technological limitations and the potential for innovation.

Station MasterrecommendHow to resolve the DeepSeek server being busy? (16 questions)freeR1 Full-Power Alternative

GetDeepSeek andAIOriginal PDF file of "Illusions"Scan the QR code to follow and reply: 20250220

What isAIHallucination

• AIIllusion refers to content generated by a model that is inconsistent with facts, logically broken, or out of context; essentially, it is a "reasonable guess" driven by statistical probability. It is divided into factual illusion (generated content inconsistent with real-world facts) and faithful illusion (generated content inconsistent with user instructions or context). Examples include incorrectly answering "diabetics can eat honey instead of sugar" or generating content irrelevant to the question.

Why does DeepSeek produce hallucinations?

• Data biasErrors or biases in the training data are amplified by the model.
• Generalization dilemmaThe model struggles to handle complex scenarios outside the training set.
• Knowledge solidificationThe model relies too heavily on parameterized memory and lacks the ability to be dynamically updated.
• MisunderstandingWhen user questions are vague, the model tends to "play it by itself".

AIHallucination assessment

• Generality testingThe hallucination rate is manually assessed by randomly generating generic prompts.
• Factual testBy extracting test questions from multiple fields, the correct answers are labeled with the type of hallucination.
• Evaluation ResultsA comparison of the illusion rate of models such as DeepSeekV3, DeepSeekR1, Qianwen2.5-Max, and Doubao.

How to slow downAIHallucination

• Technical solution:
  • Use online searches to reduce the rate of hallucinations.
  • pairAIVerification orLarge ModelCollaborate on cross-validation.
  • Prompt wordsEngineering, such as knowledge boundary constraints, adversarial hints, etc.
  • Triangulation method, cross-checking multipleAIAnswers or authoritative sources.
  • Be wary of overly reasonable responses.
  • Understand illusions and enjoy the creative inspiration they bring.

  AIThe Creative Value of Illusions

  • Scientific discoveries:AIHallucinations inspire the design of novel protein structures, contributing to scientific breakthroughs.
  • Literature and DesignAs a "surreal engine," it breaks through human mindset and provides creative inspiration.
  • Entertainment and GamesGenerate virtual environments, character designs, stories, dialogues, and poetry to enhance immersion.
  • Technological innovationThe transformation from "defects" to methodology, such asAIHallucination EnhancementautomaticDriving system recognition accuracy.

  GetDeepSeek andAIOriginal PDF file of "Illusions"Scan the QR code to follow and reply: 20250220

  Xiamen UniversityLarge ModelConcepts, Technologies, and Application Practices (PPT file) - AITutorial materials

  AI AgentMicrosoft launches software for beginners.AI Agentfreecourse