Subject

Large Language Models (LLMs) such as ChatGPT and LLaVA are increasingly used as intelligent agents capable of reasoning, planning and communicating in natural language. This thesis explores the integration of LLMs as central components in intelligent systems — ranging from robotics and drones to autonomous perception pipelines or domain-specific tools (e.g., agriculture, smart city monitoring, medical applications).

The goal is to design a system where the LLM actively enhances task performance by providing high-level plans, interpreting user intent, explaining outputs, or correcting lower-level modules. The student can also investigate how to distill this intelligence into a smaller LLM for efficient deployment.

Kind of work

The student will:

Design an LLM-driven module to interpret commands or generate high-level plans.

Integrate LLM outputs into a control or perception loop (e.g., feedback, decision-making, explanations).

Evaluate performance based on task success, explanation quality and adaptability.

Explore lightweight distillation or routing strategies to improve smaller models.

Example Application Areas (student can choose):

Natural language planning for robots or drones (e.g. in a simulated environment).

Interactive agents that provide reasoning during perception or control tasks.

LLM-enhanced classification/segmentation (e.g. via text explanations).

Compact LLM agents trained for specific tasks (e.g. in agriculture)

Framework of the Thesis

Multi-Agent LLMs and Collaborative Reasoning: https://arxiv.org/abs/2308.08155

Distillation for LLMs: https://arxiv.org/abs/2305.02301

Mixture of Experts: https://arxiv.org/abs/2407.06204

Number of Students

1

Expected Student Profile

Strong knowledge in Machine Learning, Deep Learning or Robotics.

Experience with Python and PyTorch.

Interest in combining language reasoning with perception or control systems.

Bonus: Familiarity with robotics simulation environments (e.g. Airsim, Gazeebo, CARLA) or LLM deployment/usage.