Projects

But what is your honest answer? Aiding LLM-judges with honest alternatives using steering vectors

We use steering vectors to obtain alternative, honest responses, helping external LLM-judges detect subtle instances of dishonest or manipulative behavior.

Short-circuiting Shortcuts: Mechanistic Investigation of Shortcuts in Text Classification

Mechanistic investigation of shortcuts in text classification with LLMs identifies attention heads that make premature predictions from shortcuts.calssification via reward scaling

SpotMicroAI:Learning-based gait modulation

Evaluationg sim-to-real transfer of simulation RL agents trained for gait modulations on a real world robot dog (Jake). Jake is build via 3D printing (personnally and with external help) and hobby electronics.

Balancing the Scales: Reinforcement Learning for Fair Classification

Evaluating Reinforcement Learning algorithms to mitigate bias in imbalanced text calssification via reward scaling (link to paper)

KIP:Knowledge Injection Through Prompting For Image Captioning

Master Thesis project:an image captioning model that leverages a knowledge graph for transformer promptig.

Paper Replication:Towards Visually Explaining Variational Autoencoders