Thesis information page

---

Information

Author	Bc. Alex Haščík
Supervisor	Mgr. Marek Šuppa
Study programme	Applied Informatics
Field of study	Computer Science
Department	Department of Applied Informatics
University	Comenius University in Bratislava

---

Thesis annotation

This thesis explores methods for improving the tool-routing capabilities of AI agents built on small, resource-constrained language models. Current agentic systems often rely on large models to select from a wide range of available tools - a capability that degrades significantly when model size is reduced. The work investigates lightweight approaches to tool routing, including candidate narrowing and confidence-based fallback mechanisms, evaluated on established benchmarks for agentic performance.

---

Thesis links

Summer semester goals

Goal	Status
Study relevant literature on tool-routing and agentic LLM systems	Done
Set up local inference environment (Ollama, Llama.cpp) and test open-source SLMs on local hardware	Done
Integrate the BFCL (Berkeley Function Calling Leaderboard) evaluation framework	Done
Register a custom local model endpoint in BFCL via OpenAI-compatible Ollama server	Done
Run first empirical measurement	Done

---

Next steps

• Explore other benchmarks, such as SkillsBench, ToolBench, ... (June)
• Evaluate multiple models, with multiple parameter sizes. (July)
• Analyse failure patterns such as wrong tool name vs. wrong parameter vs. wrong value type. (August and winter semester)
• Test lightweight interventions for improving tool calling accuracy: (Winter semester)
  • few-shot prompting,
  • chain-of-thought reasoning,
  • tool pre-filtering using embeddings.
• Study relevant literature. (Continuously)

---

Literature

1. Zehong Wang, Fang Wu, Hongru Wang, Xiangru Tang, Bolian Li, Zhenfei Yin, Yijun Ma, Yiyang Li, Weixiang Sun, Xiusi Chen, and Yanfang Ye. Why reasoning fails to plan: A planning-centric analysis of long-horizon decision making in LLM agents, 2026.
2. Jianyu Wen, Yang Wei, Xiongxi Yu, Changxuan Xiao, and Ke Zeng. Attention-MoA: Enhancing mixture-of-agents via inter-agent semantic attention and deep residual synthesis, 2026.
3. X. Li et al. SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks., 2026.