Research Projects
Characterisation of Goalkeeper Actions from Skeletal Data
Characterised goalkeeper actions from skeletal data by learning robust representations via contrastive learning and analysing their structure using clustering techniques. Click to read more!
Mechanistic Interpretability of VLM on Spatial Relational Reasoning
Studied LLaVA-1.5-7B using a controlled synthetic benchmark; revealed an existence bias causing yes-bias in binary relational tasks through logit lens, linear probing, and attention analyses, showing mid-layer representations are linearly decodable but misaligned with final decisions. Click to read more!
Evaluating and Finetuning LLMs for Multilingual Legal Summarisation
Adapted Apertus 8B/70B for Swiss legal summarisation, showing full fine-tuning outperforms GPT-4o and Claude 3.5 Sonnet on BERTScore and ROUGE, while highlighting trade-offs between LoRA and full fine-tuning. Click to read more!