Model Fusion with Multi-LoRA Inference for Tool-Enhanced Game Dialogue Agents

2025-09-28

AI Summary

This paper addresses three core challenges in in-game dialogue AI: character consistency, world-model alignment, and robust tool invocation support. To this end, we propose a multi-LoRA fusion architecture that separately models (1) tool-call decision making, (2) response generation conditioned on tool execution results, and (3) open-domain dialogue generation without tool reliance. Leveraging vLLM for efficient MultiLoRA inference, our approach builds upon the Qwen3-14B foundation model and incorporates synthetic data augmentation alongside multi-task fine-tuning—enabling strong performance and generalization under resource constraints. Evaluated on the CPDC 2025 GPU Track, our method achieves first place in Task 1 (character-consistent dialogue) and Task 3 (tool-augmented reasoning), and second place in Task 2 (world-aligned response generation). These results demonstrate the effectiveness and state-of-the-art capability of our framework for building tool-enhanced, immersive game dialogue agents.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Game Design — Virtual Humans, NPCs and Autonomous Characters
• Natural Language Processing: Conversational AI/Dialog Systems

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

This paper presents the opdainlp team's solution for the GPU track of the CPDC 2025 challenge. The challenge consists of three tasks, aiming to build an in-game conversational AI that adheres to character personas, aligns with the game's worldview, and supports function calling. Considering both effectiveness and resource/time constraints during inference, we synthesized data for some of the tasks based on the datasets provided by the competition organizers. We employed Qwen3-14B with LoRA fine-tuning and model fusion, and utilized a base model integrated with multiple LoRA adapters during inference. Specifically, in the competition, we used three distinct LoRA adapters to handle tool calling, response generation with tool call results, and response generation without tool call results, respectively. MultiLoRA inference was implemented using vLLM. Our solution achieved the first place in Task 1 and Task 3, and the second place in Task 2 of the GPU track.

Problem

• Building game dialogue agents with character persona consistency
• Implementing tool-enhanced conversational AI with function calling
• Optimizing model fusion for multi-task game dialogue systems

Innovation

• Used Qwen3-14B with LoRA fine-tuning
• Integrated multiple LoRA adapters for inference
• Implemented MultiLoRA inference using vLLM