Lukkien, Thijs (2025) The effect of four-bit quantization on multi-agent LLM coding performance. Master's Thesis / Essay, Artificial Intelligence.
|
Text
Thesis-T-Lukkien.pdf Download (783kB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
Toestemming.pdf Restricted to Registered users only Download (201kB) |
Abstract
The aim of this research was to investigate the effect of post-training four-bit quantization on collaborative large language model-based multi-agent systems (LLM-MAs). Collaborative LLM- MAs have shown impressive performance in many domains such as gaming operation or coding. However, their computational costs are high, as their iterative collaboration requires multiple prompts. This can result in greater economic cost and environmental impact. Quantization is a technique that maps data, such as model parameters, onto lower-precision formats, thereby compressing the model. This can increase inference speed and reduce VRAM requirements, thereby lowering economical and ecological barriers. However, the effect of quantization on LLM-MAs has not yet been explored. In this thesis, four models are quantized and retrained using QLoRA to regain performance. The four base models and the four quantized versions were then evaluated using the Mostly Basic Python Problems (MBPP) dataset and pass@k metric. It was hypothesized that quantization would reduce collaborative performance, due to its lossy nature and cascading error through iterative collaboration. The findings show that quantization can be employed with a minimal effect on LLM-MA programming performance. However, more research is needed to investigate the significance and implications of these findings.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Tashu, T.M. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 02 Oct 2025 06:40 |
| Last Modified: | 02 Oct 2025 06:40 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37068 |
Actions (login required)
 |
View Item |