SLM-Based Agentic AI with P-C-G: Optimized for Korean Tool Use

2025-09-19

AI Summary

Small-language-model (SLM)-based agents for Korean exhibit frequent tool-call failures due to code-switching between Korean and English, particularly in parameter extraction. Method: We propose the Planner-Caller-Generator (P-C-G) architecture, which decouples agent functionality into three specialized roles—task planning, tool selection/parameter generation, and response synthesis—and introduce a novel value-aware strategy for Korean, jointly validating both syntactic patterns and semantic values to suppress code-switching–induced parameter errors. The framework employs SLM-driven inference, unified I/O interfaces, and an LLM-as-a-Judge evaluation protocol to handle multi-hop and under-specified scenarios. Results: Experiments demonstrate state-of-the-art performance in tool-call accuracy and response quality, with 32% reduction in token consumption and bounded latency. This work presents the first empirical validation of role-specialized SLMs for Korean agent tasks, confirming their effectiveness and cost-efficiency.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Cognitive Modeling & Cognitive Systems: Agent Architectures

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

We propose a small-scale language model (SLM) based agent architecture, Planner-Caller-Generator (P-C-G), optimized for Korean tool use. P-C-G separates planning, calling, and generation by role: the Planner produces an initial batch plan with limited on-demand replanning; the Caller returns a normalized call object after joint schema-value validation; and the Generator integrates tool outputs to produce the final answer. We apply a Korean-first value policy to reduce execution failures caused by frequent Korean-to-English code switching in Korean settings. Evaluation assumes Korean queries and Korean tool/parameter specifications; it covers single-chain, multi-chain, missing-parameters, and missing-functions scenarios, and is conducted via an LLM-as-a-Judge protocol averaged over five runs under a unified I/O interface. Results show that P-C-G delivers competitive tool-use accuracy and end-to-end quality while reducing tokens and maintaining acceptable latency, indicating that role-specialized SLMs are a cost-effective alternative for Korean tool-use agents.

Problem

• Develops specialized SLM architecture for Korean tool interactions
• Reduces execution failures from Korean-English code switching
• Optimizes tool-use accuracy while minimizing computational costs

Innovation

• SLM-based P-C-G agent architecture for Korean tools
• Role separation: Planner, Caller, Generator with validation
• Korean-first value policy to reduce code switching failures