| Meno: | Artsiom |
|---|---|
| Priezvisko: | Shcherba |
| Názov: | A Reward-Driven Framework for Behavioral Adaptation of Large Language Models |
| Vedúci: | prof. Ing. Igor Farkaą, Dr. |
| Rok: | 2026 |
| Kµúčové slová: | large language models, reward-driven learning, behavioral adaptation, cognitive distortions, rejection sampling |
| Abstrakt: | This thesis proposes a reward-driven framework for behavioral adaptation of large language models in dialogue environments. The initial motivation of the work was the simulation of cognitive and logical distortions in language; however, the proposed approach generalizes to arbitrary textual behaviors defined through reward functions. The framework is based on interaction between a fixed Doctor model and a trainable Patient model. Instead of relying on supervised target responses, model behavior is optimized using modular reward functions evaluating generated dialogue responses. The thesis analyzes the limitations of supervised fine-tuning for behavioral simulation and compares several optimization strategies, including reinforcement learning methods such as Proximal Policy Optimization (PPO). Based on their methodological requirements, implementation complexity, compatibility with modular reward functions, and suitability for the proposed framework, rejection sampling was selected as the primary training strategy. As part of the work, a configurable experimental system was implemented, including a dialogue environment, modular reward architecture, and support for user-defined behavioral criteria. Experimental results demonstrate that the proposed method can model both cognitive distortions and general lexical or structural text patterns. |
Súbory bakalárskej práce:
Autor nedal súhlas so zverejnením svojej bakalárskej práce.
Súbory prezentácie na obhajobe: