University of Twente Student Theses
As of Friday, 8 August 2025, the current Student Theses repository is no longer available for thesis uploads. A new Student Theses repository will be available starting Friday, 15 August 2025.
Selective Knowledge Transfer via communication-aware Model Alignment
Kezins, N. (2025) Selective Knowledge Transfer via communication-aware Model Alignment.
|
PDF
557kB |
| Abstract: | Knowledge transfer techniques, such as knowledge distillation, supervised fine-tuning, typically involve training a student model on a fixed dataset or soft targets derived from a teacher model. While these methods can be effective, they often require extensive training data and do not account for the student model’s specific learning weaknesses, which can lead to redundant training on material the student model has already mastered, or, conversely, to the omission of new and potentially valuable knowledge. In this paper, we explore Selective Knowledge Transfer via Communication-Aware Model Alignment, a novel approach where a teacher model iteratively identifies and addresses a student model’s deficiencies through dynamic interaction. The teacher generates targeted examples, evaluates student responses, and adjusts its strategy to focus on areas of weakness, using Direct Preference Optimization (DPO) for alignment. Our method introduces an adaptive interaction mechanism, but empirical evaluations show mixed results: while most experiments exhibit performance degradation compared to traditional distillation, one setting demonstrates a slight improvement. These findings underscore the challenges of aligning dynamic interaction with effective knowledge retention and suggest directions for refining communication-based distillation strategies. This work contributes insights into the complexities of adaptive knowledge transfer and pathways for future research. |
| Item Type: | Essay (Bachelor) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Computer Science BSc (56964) |
| Link to this item: | https://purl.utwente.nl/essays/107677 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page