[2025F] Henry Gao (BS)

Instructing LLM to Learn Human Behavior through Dialogue

Henry Gao

Date: 2025-10-03 / 3:00 - 4:00 PM

Location: White Hall 100

Abstract

Large Language Models (LLMs) demonstrate strong linguistic and reasoning abilities but remain limited in executing structured, task-oriented dialogues that demand procedural understanding. This work explores how LLMs can autonomously infer task instructions from dialogue histories to better mimic user behavior. We propose a two-step framework in which the model first derives structured task instructions and related preconditions from prior dialogues and then uses these instructions to guide response generation. Using the MultiWOZ and SGD benchmarks with open-source models (Qwen 3, LLaMA 3), we expect our approach to achieve performance comparable to proprietary systems such as ChatGPT-5. This study introduces the first method for automatically deriving task instructions, advancing LLMs toward more consistent and human-like task execution in dialogue systems.

Link

Presentation