Master thesis:Custom LLM Development

Join our Team

About this opportunity

AI coding tools like Claude Code, Kiro, or Codex have become indispensable to most developers. At Ericsson, many developers work with a C-like programming language customized to program Ericsson's own silicon – Ericsson's Many-Core Architecture (EMCA). Because this language is not represented in the training data of general-purpose LLMs, coding assistance for EMCA can be inefficient, both in terms of token consumption and output quality.

At Ericsson's System Comprehension Lab, we are training an in-house language model on EMCA source code to serve as a specialized tool for an orchestration agent. We are recruiting for two related master's thesis opportunities in this area: one on mitigating catastrophic forgetting during continued EMCA-specific training and one on tool-integrated reasoning. Both projects aim to improve the model's usefulness for EMCA development.

Please indicate in your application which project you are most interested in. The exact focus of each project will be refined together with the student based on their interests, background, and the state of the model training team's research when the project starts. The work will be carried out with the team in Kista. You will have access to training pipelines, GPU resources, evaluation suites, and coding assistants such as Claude Code.

What you will do

Direction 1: Mitigating Catastrophic Forgetting

The model has general capabilities, such as instruction following and broad coding knowledge, acquired during its original pre-training. At Ericsson, we adapt it to EMCA through continued pre-training and supervised fine-tuning on EMCA data. Additional EMCA-specific training may cause catastrophic forgetting: the model may improve on EMCA tasks while losing some of those general capabilities. This project investigates methods to improve EMCA-specific performance while preserving these capabilities.

• Establish a baseline using the current continued-pre-training and supervised-fine-tuning approach
• Implement and evaluate methods such as self-distillation, curriculum learning, model merging, or related techniques
• Measure both gains in EMCA-specific knowledge and retention of general capabilities, including instruction following and broad coding skills

Direction 2: Tool-Integrated Reasoning

This project explores how to extend the model with reasoning and tool interaction, potentially using reinforcement-learning-based training. Tool-integrated reasoning allows the model to invoke external tools and use their feedback during problem solving, with the aim of improving code accuracy and efficiency for complex or abstract EMCA tasks.

• Establish a baseline using the current model and standard reasoning methods
• Design, implement, and evaluate methods that enable the model to decide when to invoke tools – such as code execution, graph traversal, or vector search – interpret their feedback, and correct its reasoning when needed
• Assess improvements in EMCA task performance, code accuracy, tool-use effectiveness, token efficiency, and the ability to handle more abstract queries

The skills you bring

• You are pursuing a master's degree in computer science, engineering mathematics, physics, or a related field
• Good understanding of modern AI and machine learning, particularly generative AI and large language models
• Strong programming skills and an interest in experimental, research-oriented work
• Familiarity with deep-learning frameworks such as PyTorch
• Previous experience with LLM fine-tuning is a plus
• You are analytical, curious, and able to work independently while communicating your results clearly