[2E1-GS-5b] Agents

9:15 AM - 9:30 AM JST(12:15 AM - 12:30 AM UTC)

[2E1-GS-5b-02] Impact of Model Capability on MCP Tool-Call Optimization

〇Toshiaki Sota¹, Mitsuhiro Nakayama², Takehiko Yamaguchi², Hiromi Kawatsu², Michiaki Tatsubori² (1. IBM Japan Systems Engineering , 2. IBM Japan)

Keywords:

LLM Agent,Small LLM,Model Context Protocol (MCP),Tool Use

We evaluated a tool-context optimization method for improving tool calling by small LLMs under the Model Context Protocol (MCP) across six models. The effectiveness of synthetic-test-based automatic configuration selection (Proposed) was strongly dependent on model capability: it achieved up to +0.50 absolute improvement for weaker models (Gemma 3 4B, Granite 3.3 8B), while for mid-to-strong models (e.g., Phi-4 14B, Qwen 3 8B) it could underperform the baseline in some cases (Spearman ρ = −0.73), revealing a negative correlation with model strength. We also found that the impact of description variants (as-is, concise, rich) is domain-specific, indicating that a one-size-fits-all strategy is not optimal.