Training AI Agents to Accumulate Skills and Habits

Thumbnail Image

Files

master_Ananov_Georgy_2025.pdf (343.77 KB)

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-02-24

Department

Major/Subject

Machine Learning, Data Science and Artificial Intelligence

Mcode

Degree programme

Master's Programme in Computer, Communication and Information Sciences

Language

en

Pages

33

Series

Abstract

Recent advances in development of AI agents have demonstrated that pairing interactive frameworks with Large Language Models (LLMs) can enable autonomous systems to perform a wide range of real-world tasks. However, while agents can handle many objectives without additional training, more complex problems demand that the agent's skills be adapted to specific domains. The need for domain-specific knowledge highlights a broader challenge: developing training strategies that not only enable an agent to master individual skills but also facilitate the effective composition of these skills to solve complex, multi-step problems. In this thesis, we investigate prompt distillation — a technique that transfers textual guidance directly into the parameters of a student model — to understand how training protocols can be structured to promote compositional learning. By training the agent on a series of tasks that represent sequential steps of a larger workflow, and then evaluating the agent on a unified composite task, we examine the relationship between individual skill mastery and the ability to combine these skills effectively. Our experimental results demonstrate that while individual tasks can be mastered with relatively few training epochs, achieving robust compositionality requires more extensive training. Nonetheless, the findings confirm that prompt distillation can successfully instill a flexible, integrative skill set, paving the way for scalable and adaptable AI agents.

Description

Supervisor

Marttinen, Pekka

Thesis advisor

Ilin, Alexander
Valpola, Harri

Keywords

LLM, AI, AI agent, prompt distillation, continual learning, deep learning

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202503182911

Collections

[dipl] Perustieteiden korkeakoulu / SCI