微提示LLMs和SLM：从复制品到代理工作量.pdf

1、AI Hardware&SystemsaiandsystemsMicro-Prompting LLMs and SLMsFrom Copilots to Agentic WorkloadsDonald ThompsonDistinguished EngineerMicrosoft/LinkedInAI Hardware&SystemsaiandsystemsAgenda Macro-prompting vs micro-prompting Micro-prompting example Automatic prompt optimization Automatic fine-tuning SL

2、Ms vs LLMsAI Hardware&SystemsaiandsystemsMacro-PromptingThe Current Paradigm in GenAI Dominant approach since late 2022 Crafting extensive,detailed prompts Provides comprehensive instructions to frontier LLMsJailbreak and RAI safeguardsIntent ClassificationTask Instructions and ExamplesContext and C

3、onversational MemoryFunction Calling/Formatting InstructionsAnatomy of a Macro-Prompt 8k tokensAI Hardware&SystemsaiandsystemsMacro-PromptingApplication Engineering Challenges Limited expertise in prompt engineering Limited time budget for iteration and evaluation Default reliance on expensive,scarc

4、e models(GPT-4)Reduced opportunity for fine-tuningAI Hardware&SystemsaiandsystemsMicro-PromptingA Paradigm Shift in GenAI Development Automated Goal-oriented,measurable Concise,modular,highly optimized Cost-effective on commodity h/w Scalable,durable processTask DescriptionInput DescriptionOutput Sc

5、hemaExamples(Input+Output)Anatomy of a Micro-Prompt 300 tokensoptimizeAI Hardware&SystemsaiandsystemsMicro-PromptingA New Workflow Decompose problems into discrete functional tasks Clearly-defined inputs and outputs Synthesize examples(input+output)Define measurement criteriaModuleSampleTaskPropose

6、the top 3 best seats for a traveler with the provided seat preferences,given the list of available seats.Input 1“Seat Preferences”:The seating preference for the traveler.Input 2“Available Seats”:An array of seats,each in the format class_seat_aisle|middle|windowOutputtype:object,properties:recommen