Vincent Dörig does research and engineering
at the
intersection of
- 01LLMs,
- 02interactions,
- 03& alignment.
A benchmark exposing how fast LLM safety collapses when models self-distill from adversarial user feedback.
An AI-enhanced interactive reader transforms textbooks into engaging visual learning experiences.
A workflow-integrated, mixed-initiative LLM assistant designed to streamline documentation in manufacturing environments.
An interactive visual dashboard using SHAP to explain XGBoost predictions for Anti-Money Laundering compliance.
Continuous learning enables LLM personalization—but also permanent learning of harmful behaviors. We created a benchmark based on adversarial user feedback to evaluate safety degradation during continual post-training (like SDPO). Our results showed that model guardrails can collapse in under 100 user interactions, but that introducing a constitutional safety prompt can largely restore them.
We plan on releasing the dataset on Hugging Face soon.
IRead is an LLM-powered educational tool that transforms static textbooks into interactive learning environments. Instead of passively reading, the system engages students with adaptive questioning and dynamic concept maps to support deeper comprehension. In our empirical user studies, the tool demonstrated measurable improvements in both student engagement and overall learning outcomes.
Designed to streamline manufacturing shift documentation, this workflow-integrated LLM assistant empowers shopfloor operators with varying levels of digital literacy. Moving away from conventional RAG-based chatbots, we engineered a system that uses context-aware triggers, AI-assisted voice-to-text structuring, and automated handover summaries. It is actively deployed with full audit traceability.
AMLXplainer is an interactive visual analytics framework that transforms black-box fraud detection models into interpretable, regulator-ready insights. By coupling an XGBoost classifier with dual-level SHAP explanations, the system enables compliance officers to systematically investigate alerts and document triage decisions without compromising underlying predictive performance.
