Small LLMs struggle with multi-step tool use in agentic applications

A developer building an agentic assistant for battery engineering discovered limitations when using smaller language models, specifically the qwen2.5:7b model. While single-tool calls for simulations were reliable, multi-step tool orchestrations, such as comparing discharge rates, caused the model to drop steps or produce incomplete answers. This issue was exacerbated by the model's tendency to speculate on simulation results rather than strictly reporting them. The solution involved collapsing complex multi-tool plans into single, more comprehensive tools, thereby reducing the planning burden on the smaller language model. AI

IMPACT Highlights limitations of smaller LLMs in complex agentic tasks, suggesting a need for architectural changes or larger models for advanced orchestration.

RANK_REASON Developer's personal experience and findings on LLM limitations, not a formal release or research paper.

Read on dev.to — MCP tag →

• Chen2020
• compare_discharge_rates
• FastMCP
• LangGraph
• LG M50 21700
• llama3.1:8b
• Ollama
• PyBaMM
• qwen2.5:7b
• simulate_discharge

• product
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →