BROWSE
Detail
Conference Paper
Tool-Catalog Scaling and Tool RAG in Local sLLM Agents
Cited - time in 
Share 
Share
- Authors
- Dong Hoon Son, Seung Hun Oh, Yun Ji Ban, Hong Yeon Yu, Jeongeun Kim, Seihyoung Lee
- Issue Date
- 2026-10
- Citation
- International Conference on Information and Communication Technology Convergence (ICTC) 2026, pp.1-3
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- This paper measures how the number of candidate tools affects the tool-selection performance of open-source small language models (sLLMs) deployed as local tool-using agents, and analyzes the effectiveness of Tool RAG as a mitigation. On a Korean function-calling benchmark (100 queries, 119 tools), four Qwen3.5 models (2B–27B) are evaluated while the tool catalog is scaled from 1 to 119 tools, and the catalog-size effect is verified on EXAONE 3.5 7.8B and Gemma 3 12B. Accuracy decreased as the catalog grew, and more sharply for smaller models (2B 97%→66.3%; 27B 98%→92.0%). The drop reproduced in two other families (EXAONE 3.5 7.8B 99%→70.0%; Gemma 3 12B 98%→83.7%), and its size was not ordered by parameter count across families. When Tool RAG was applied with a baseline retriever (bge-m3) and k=5 exposed tools, accuracy did not improve and dropped sharply for larger models (27B −18.0%p), as retriever recall caps the attainable accuracy. Replacing the retriever with one achieving 98% recall improved the 2B model by 16%p, whereas Qwen3.5 models of 4B and larger gained at most 0\~4%p even with larger k. Tool RAG thus serves as an accuracy lever for 2B-class models, while for larger models its value lies in a 91% prompt-token and 71% latency reduction.
- Keyword
- tool calling, sLLM, tool retrieval, local agent
- KSP Keywords
- Language Models, Latency reduction, Open source, size effect