About me

I’m a first-year PhD student in Computer Science at the University of Cambridge and a MATS 12.0 Scholar in Neel Nanda’s stream. Previously, I worked on AI safety research at LASR in London. I am a member of LISA.

I find neural networks really cool and work on making LLMs safe. Driven both by curiosity about their internal mechanisms and by concern for their safe deployment, I currently explore how white-box interpretability methods can aid alignment and monitoring. I also appreciate the awesome, research-driven engineering1 behind highly performant deep learning systems!

Please, don’t hesitate to reach out! I’m very happy to hear about your research, tell you a bit about mine, or just chat about anything interesting!

Research updates

See previous updates
  • Accepted to Neel Nanda’s MATS 12.0 Exploration Phase.
  • Serving as a reviewer for ICLR 2027.
  • Secured ~£14k funding with Marek Masiak for further research on the interpretable-by-design TopKLoRA fine-tuning method.
  • Presented The Model Organism Lottery 🎲 poster at the ICML 2026 Mechanistic Interpretability Workshop in Seoul, South Korea! 🎉🇰🇷 (poster)
  • “The Model Organism Lottery 🎲” on unrealistically easily interpretable model organisms for AI safety is here!
  • LangMAP, our collaborative (Cambridge-ETH-EPFL) work towards an alternative tokenisation paradigm for multilingual LLMs, has shipped! 🇬🇧🇨🇭
  • Served as a reviewer for the ICML 2026 Mechanistic Interpretability Workshop.
  • Participated in the Research Accelerator Week. Cambridge, UK! 🇬🇧
  • Joined LASR Labs to work on technical AI safety research. London, UK! 🇬🇧
  • Presented Activation Transport Operators as a spotlight at the Mechanistic Interpretability Workshop at NeurIPS 2025 in San Diego, USA! 🎉🇺🇸 (poster)
  • Our recent work “Activation Transport Operators” on transporting SAE features across transformer layers is available!
  • Presented my MPhil dissertation at MobiUK 2025 in Edinburgh, UK! 🇬🇧 (poster)
  • Presented preliminary results for Data-Efficient Task Unlearning at EEML 2024 in Novi Sad, Serbia! 🇷🇸 (poster)

Selected publications

  1. Interpretable Sparse Fine-Tuning with TopKLoRA Enables Faithful Circuit Discovery Masiak M.*, Szablewski A.*, Cancedda N. and Venhoff C. Preprint, under review, 2026 arXiv code
    Key figure from Interpretable Sparse Fine-Tuning with TopKLoRA Enables Faithful Circuit Discovery
  2. The Model Organism Lottery: Model Organism Interpretability Strongly Depends on Training Methodology Szablewski A.*, Konar-Steenberg G.*, Fornasiere R.*, Menon N.* and Heimersheim S. ICML 2026 Mechanistic Interpretability Workshop arXiv code poster blog
    Key figure from The Model Organism Lottery: Model Organism Interpretability Strongly Depends on Training Methodology
  3. LangMAP: A Language-Adaptive Approach to Tokenization Meister C., Salhan S., Szablewski A., Lesci P., Buttery P. and Pimentel T. EMNLP 2026 (main conference) Oral Also presented at the COLM 2026 Tokenization Workshop. arXiv code
    Key figure from LangMAP: A Language-Adaptive Approach to Tokenization
  4. Activation Transport Operators Szablewski A.* and Masiak M.* NeurIPS 2025 Mechanistic Interpretability Workshop Spotlight arXiv code poster
    Key figure from Activation Transport Operators

Blog posts are on the Publications page.

More about me

In early 2026, I joined the LASR fellowship, which I highly recommend2. Under the supervision of Stefan Heimersheim, my team and I showed that model organism training methodology significantly affects white-box interpretability. We presented it as The Model Organism Lottery at an ICML 2026 Workshop in Seoul. At the end of the fellowship in April, my team was awarded a ~£200k research grant for further research work until September. Simultaneously, I worked on LangMAP – our collaborative (Cambridge-ETH-EPFL) work towards an alternative tokenisation paradigm for multilingual LLMs, which was accepted as an oral at EMNLP 2026. Towards the end of summer, I focused on TopKLoRA, a fine-tuning method providing interpretable-by-design and composable feature latents, which we released as a preprint in October. Following my mini-project on cross-model thought anchors, I was accepted into the exploration phase of Neel Nanda’s MATS 12.0 stream, after which I was selected as a MATS 12.0 Scholar.

Throughout my year-long MPhil degree, I was supervised by Prof. Nic Lane and worked with the CaMLSys group, where I focused on Federated Learning. My dissertation explored how several institutions with limited computational resources can collaborate on training a joint foundational language model. During my studies, I also benchmarked the inner workings of torch.compile(), and explored the KV-caching strategies in LLM inference. On the more theoretical front, I looked into the phenomenon of attention sinks3 in transformers and studied the concept of dynamic tokenisation. What a year it was!

Previously, I have been working with researchers from UCL NLP on BritLLM – a joint effort towards producing freely available Large4 Language Models for UK languages5. In my undergraduate dissertation supervised by Prof. Pontus Stenetorp, I focused on the problem of the poor availability of LLMs for low-resource languages and worked on language model adaptation methods for African languages. Furthermore, we explored Data-Efficient Task Unlearning in LMs, a method for increasing the safety of language models and removing their undesired capabilities.

Notes

  1. Awesome Engineering! ↩

  2. Kaushik, another LASR fellow, wrote a good retrospective on his experience at LASR. ↩

  3. Write-up in progress. Update: It’s arrived! ↩

  4. As of the time of writing (2024-08-27), 3 billion parameters make the model be considered large. Update: As of 2025-07-28, a 3B model is still pretty big. Reflecting on this during ICML 2026, I still think that 3B is quite substantial. In September 2026, I worked with a 1.7B model and even that was quite sizeable! ↩

  5. Isn’t this just… English?! Explore our work to see what other languages are spoken in the UK! ↩