2857 - IBM 的下一代代码助手 - 工作原理和演示.pdf

1、Orlando,FLOctober 69IBM TechXchange 202528572857IBMs next generation code assistant IBMs next generation code assistant-How it works and demosHow it works and demosJulio SnchezIBM,AI EngineerBob for IBM iAgenda0102030405Evolution of LLMsAgentsBob for iDemosQ&AIBM TechXchange|2025 IBM CorporationMy R

2、PG/GenAI Journey4IBM TechXchange|2025 IBM CorporationMaster thesis IBM Internship:Can we teach a trained model RPG?Fixed to Free format conversion Code Explanation Code Generation1st RPG-focused model ever.Only 15b parameters.Fine-tuned from open-source model,available on Hugging Face Result?IBM Tec

3、hXchange|2025 IBM Corporation5Popular techniques:1.Masked Language Modeling(MLM)2.Casual Language Modeling(CLM)How does LLMs learn?IBM TechXchange|2025 IBM Corporation61.MLM2.CLMAutoregressive behavior.1st generation code assistants:auto-complete.How does LLMs predict?IBM TechXchange|2025 IBM Corpor

4、ation7Supervised fine-tuning learning.2nd generation code assistants:instruction-following.How does LLMs follow instructions?IBM TechXchange|2025 IBM Corporation8This RPG procedure named Function takes three integer parameters(Term1,Term2,Term3)and performs a weighted mathematical calculation.It squ

5、ares the first term and multiplies by 17,multiplies the second term by 7,adds the third term unchanged,then multiplies the entire result by 45 and adds 23 to produce the final return value.Convert to free-formatAgenda0102030405Evolution of LLMsAgentsBob for iDemosQ&AIBM TechXchange|2025 IBM Corporat

6、ionFail with multi-step tasksThey guess,not reasonStatistical nature.Fluent answers,but prone to hallucination for complex task.Example:Example:“Whats 17 24?”Traditional LLM:418 Expected answer:408Why?Example never seen before.Do not reason about the problemFails when problems require multi-step pla