Léon Àmi Wagner

Humboldt University of Berlin, Germany · Oct 2025 - Jun 2026 · submitted, defense 30 Jul 2026

This thesis tackles a fundamental problem in deploying LLMs in enterprise settings: how to integrate evolving, domain-specific knowledge without catastrophic forgetting. Standard fine-tuning methods overwrite general knowledge with new information, while existing parameter-efficient approaches like LoRA still cause interference between old and new facts. The proposed 'Patch-and-Route' framework draws on cognitive science — treating knowledge updates not as destructive overwrites, but as discrete Knowledge Patches combined with a dynamic routing mechanism that inhibits outdated knowledge paths rather than erasing them, mirroring how humans re-route rather than delete disproven associations. The architecture consists of three components: a frozen base LLM, an Expert Pool of LoRA adapters (large Base Adapters for domain corpora and small Knowledge Patches for specific updates), and a two-level routing system with manual domain selection and an Intelligent Dispatcher. Two routing strategies are evaluated: Time-Aware Centroid Routing (embedding-based, with RAG-based source replay for conflict resolution) and a Parallel-Orchestrator Architecture (ensemble inference with an LLM synthesis agent for ambiguous queries). The framework is benchmarked against monolithic LoRA, LoRA+RAG, X-LoRA, and RECIPE on SituatedQA, CounterFact, and a proprietary enterprise QM corpus, measuring conflict resolution accuracy, catastrophic forgetting rate, and computational efficiency. Submitted 18 June 2026; defense on 30 July 2026.

Humboldt University of Berlin, Germany · Nov 2024 - Jun 2025

This study project investigates the effectiveness of reasoning-enhancement techniques (e.g., advanced prompting, multi-stage finetuning, and the hybrid STaR method) on a 1-billion parameter language model, comparing its performance against a larger 3B model. The research demonstrates that with the right approach—specifically by finetuning on a mixed dataset and applying Plan-and-Solve prompting—a 1B model can significantly outperform its larger counterpart on complex reasoning tasks like GSM8K. The project concludes that the success of these techniques is highly dependent on the model's pre-training and the specific task, highlighting a path to creating smaller, yet highly capable language models.

Humboldt University of Berlin, Germany · Spring 2024

Achieved second place in a Kaggle competition focused on image classification. The project involved designing, training, and fine-tuning a Convolutional Neural Network (CNN) to achieve a high accuracy score (Public: 0.937, Private: 0.928).

Humboldt University of Berlin, Germany · Fall 2023

This project simulates a 2D environment where robotic arm tools interact with various objects. The simulation includes elements like grip arms, sticks, and magnets, designed to explore intrinsic motivated learning and tool use in a controlled setting for a Machine Learning in Robotics seminar.

HTW Berlin, Germany · Apr 2022 - Aug 2022

In a team of four, we divided the tasks of construction, programming, and testing to create a customizable, self-sufficient vacuum cleaner. I was primarily responsible for the software development, ensuring the vacuum cleaner's efficient navigation and obstacle avoidance.

HTW Berlin, Germany · Oct 2021 - Feb 2022

Our team developed a speech control system for controlling a model factory as part of a collaborative project with PI-Informatik. I was instrumental in the development of designing the GUI, implementing the speech recognition, and the integration with the Raspberry Pi.

HTW Berlin, Germany · Apr 2021 - Aug 2021

My first software project: Collaborated with two fellow students to create a recipe generator that transforms leftovers into new meals. The repository link is unavailable as I've lost my old university credentials.