Doctoral Candidate at the University of Manchester, working on chain-of-thought alignment between language and actions in Vision Language Action models. I use mechanistic interpretability to understand how VLAs ground language in the physical world.
Looking to bring that research foundation into industry as an ML Engineer or Applied Scientist.
- Vision-Language-Action Models (VLAs): alignment between chain-of-thought reasoning and generated actions
- Language grounding and semantic alignment across vision, language, and embodied action
- Mechanistic interpretability in VLAs and foundation models
-
GPLA
Preference learning framework to align chain-of-thought reasoning with action generation. Published at CVPR Findings 2026 -
Action Language Transparency
Joint generation of action and language tokens to improve interpretability of action generation Published at IJCNN2025 -
PPO Street Fighter 2
PPO-based RL agent trained to play Street Fighter 2 (older project)
-
Grounding Hierarchical Vision-Language-Action Models Through Explicit Language-Action Alignment
Project Page · PDF T. Wulff, R.S. Maharjan, F. Tavella, M. Adikari, and A. Cangelosi -
Joint Action Language Modelling for Transparent Policy Execution
GitHub · PDF T. Wulff, R.S. Maharjan, X. Chi, and A. Cangelosi
- Designing Machine Learning Systems, Chip Huyen
- The Road, Cormac McCarthy
- How Tyrants Fall: And How Nations Survive, Marcel Dricus
- Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist, Liz Pelly
- Play guitar. Strictly bedroom level, no audiences harmed.
- Avid music listener. Always looking for recommendations.
- Boardgame enthusiast. Absolutely love Spirit Island.
- BJJ purple belt & karate black belt. I used to coach kids' karate, which taught me more about explaining things clearly than most of my academic training did.