End-to-End Task-Oriented Dialogue Agent Construction Using Monte Carlo Simulation and LLM Fine-Tuning

Traditional Rasa‑like dialogue systems rely on tightly coupled modules such as NLU, dialogue management, and policy, which creates integration, error‑propagation, maintenance, and development challenges. An end‑to‑end (E2E) model that unifies all stages simplifies architecture, improves efficiency, and reduces error transmission, making the system easier to maintain and update.

The proposed solution uses Monte Carlo simulation to model user behavior and a directed graph to represent dialogue state transitions. Random walks on this graph generate diverse conversation paths, while LLMs generate realistic utterances for each state, creating a rich training dataset that captures intent, slot filling, and function‑calling scenarios.

Key challenges addressed include: (1) recognizing user intent and slots after fine‑tuning, (2) handling ambiguous or unrelated inputs with confirmation or fallback strategies, (3) triggering function calls at appropriate moments, and (4) extracting key information from context.

Data construction starts with defining dialogue nodes (Start, IntentAcquire, UserInquiry, IntentConfirm, UserConfirm, UserDeny, AskSlot, ProvideSlot, FunctionCalling, Chitchat, End) and connecting them with directed edges to form a state‑transition graph. Monte Carlo methods randomly decide which slots appear in the initial user query and simulate subsequent user actions based on predefined transition probabilities (e.g., [0.8, 0.1, 0.1]). This yields varied dialogue flows that reflect real‑world interactions.

To enhance contextual understanding, the method omits already‑provided slot information in later turns, forcing the model to rely on context and pronouns, thereby improving the agent’s ability to handle ellipsis.

LLMs are employed to generate both user queries and system responses. Example prompt for generating a restaurant‑recommendation query is wrapped below:

你是一个用户，你现在想要「根据自己的位置、兴趣和预算，让智能客服推荐当地的餐厅」，请向智能客服寻求帮助
你的问题需要满足以下几个条件
1. 在问题中需要提到具体的用户当前的区域或希望探索的区域。
2. 在问题中一定不要提到具体的用户感兴趣的餐厅类型，中餐，日料，西餐等。
3. 在问题中一定不要提到具体的用户的最大预算。
请生成一句满足当前的场景和设定的问题

A YAML configuration file defines intents and slots, enabling easy extension without modifying the state‑graph:

name: recommend_restaurant 
description: 根据自己的位置、兴趣和预算，让智能客服推荐当地的餐厅
parameters:
  - name: destination
    description: 用户当前的区域或希望探索的区域。
    type: text 
    required: True
  - name: cuisine_type
    description: 用户感兴趣的餐厅类型，中餐，日料，西餐等 
    type: text 
    required: True
  - name: budget
    description: 用户的最大预算
    type: float 
    required: False

Fine‑tuning is performed with LLaMA Factory, an open‑source framework supporting LoRA. Experiments on an A6000 GPU (fp16, LoRA rank 4, 3 epochs) compare several open‑source models (ChatGLM3, Qwen 1.5, Yi) before and after fine‑tuning. Results show the Qwen 1.5 Chat model achieves the best overall performance, with notable gains in intent recall and slot‑filling accuracy.

Qualitative examples illustrate that the fine‑tuned agent can correctly resolve omitted references (e.g., linking “当地” to a previously mentioned city) and ask follow‑up slot questions, surpassing traditional Rasa‑like agents.

Conclusion: Combining Monte Carlo‑based data generation with LLM‑driven content creation and efficient fine‑tuning via LLaMA Factory yields an effective E2E task‑oriented dialogue agent. The approach retains LLM capabilities while improving intent detection, slot filling, and contextual reasoning, though future work should address dataset generation scalability, model interpretability, and controllability.