Ivan Titov

Professor in Natural Language Processing & ML · Edinburgh & Amsterdam

• Sparsity and differentiable structure: sparse attention, binary variables, adaptive computation, neuron pruning
• Knowledge in LLMs: editing without retraining, unlearning, memorisation, knowledge offloading
• Interpretability: differentiable masking, representation analysis, information bottleneck
• Trustworthy foundation models: chain-of-thought faithfulness, controllability
• Post-training methods: routing, modularity, reasoning
• Decentralization: federated/privacy-preserving NLP, agent collaboration

Earlier interests include graph neural networks, structured prediction, latent variable models, multilinguality, machine translation, and semantic parsing.

I direct the CDT in NLP at Edinburgh and the Edinburgh ELLIS unit; ELLIS Fellow, former Turing Fellow, co-director of the ELLIS NLP program. Program Co-Chair at ICLR 2021 and CoNLL 2018; Action Editor for JMLR and TACL; paper awards at ACL and EMNLP.

Funded by ERC, NWO (Vici), UKRI, Google, SAP, Booking.com, Amazon.