About
I’m a Visiting Scholar in language technology at the University of Helsinki, where I designed and teach the course Large Language Models and Generative AI for NLP. Alongside research, I lead large-model inference infrastructure and performance engineering at AMD Silo AI.
My training is in logic and formal grammar, at the London School of Economics and King’s College London. After over a decade in industry, I returned to research for my PhD in language technology at the University of Helsinki, completed with distinction in 2024. More under Industry experience.
Research
- Formally grounded evaluation of LLMs
- What benchmark scores really measure: shortcuts, memorisation or inference. My doctoral work showed that models keep much of their accuracy when the words a reader needs are removed, and recent work shows that the robustness of LLMs follows cues in the data, not prior exposure to the benchmark. I also work on uncertainty-aware inference, and co-organise the ELOQUENT lab at CLEF, including the 2027 Fragments task.
- Formal semantics, grammar and prosody
- Type-theoretic semantics and formal grammar. With Stergios Chatzikyriakidis I contribute to the Formal Atlas of formalised semantic theories (paper), and work on Path Grammars, a formalism for incremental syntax. I also study how prosodic prominence connects to meaning, most recently in speech and language models.
- Neuro-symbolic reasoning
- Methods that address what diagnosis reveals: inference that can be checked by a proof or refuted by a countermodel, grammar-constrained decoding, and finding which linguistic phenomena Transformers need symbolic support for.
Publications
Also on Google Scholar.
Preprints
- 2026
- 2026
- 2026
Journal and conference papers
- 2025
- 2023
- 2022
- 2021
-
2019
Sentence embeddings in NLI with iterative refinement encoders Natural Language Engineering 25(4) PDF (arXiv)DOI
- 2019
- 2019
- 2019
Doctoral dissertation
-
2024
Towards natural language understanding: developing and assessing approaches and benchmarks Doctoral dissertation, University of Helsinki Thesis repository
Shared task overviews
-
2024
Overview of ELOQUENT 2024: shared tasks for evaluating generative language model quality CLEF 2024, Lecture Notes in Computer Science 14959 DOI
- 2024
- 2024
- 2024
-
2024
ELOQUENT CLEF shared tasks for evaluation of generative language model quality ECIR 2024, Lecture Notes in Computer Science 14612 DOI
Technical writing
-
2025
Scaling AI inference performance with vLLM on AMD Instinct MI355X GPUs AMD ROCm Blogs, December 2025
Talks and media
Selected talks
I have given numerous talks at academic and industry events, and in recent years have spoken mainly at industry events. A selection:
-
2022
How does data corruption affect natural language understanding models? *SEM 2022, Seattle
-
2022
Multimodal and multilingual vector search with hardware acceleration Berlin Buzzwords 2022, Berlin, with Dmitry Kan
-
2022
Demo: multilingual and multimodal vector search ValleyML AI Expo 2022
-
2021
NLI data sanity check: assessing the effect of data corruption on model performance NoDaLiDa 2021, Reykjavik
-
2019
Neural network models of NLI fail to capture the general notion of inference Invited talk, CLASP, University of Gothenburg
-
2019
Predicting prosodic prominence from text with pre-trained contextualized word representations NoDaLiDa 2019, Turku
-
2018
Unlock the value of your data assets Gartner Symposium, Barcelona
-
2018
Business value of AI AI Monday, Helsinki
In the media
-
2024
Tekoälyä uhkaa luhistuminen – ratkaisuja etsitään kuumeisesti TiVi, October 2024
-
2024
AI training data nearing exhaustion – focus on quality needed TiVi, April 2024
-
2024
Tekoälybuumi edellyttää uusia luonnollisten kielten ymmärtämisen mittareita University of Helsinki via STT, February 2024
Teaching
- Large Language Models and Generative AI for NLP, Master’s course, University of Helsinki, every autumn since 2024, with Jussi Karlgren and Dmitry Kan. I designed the syllabus and teach four of the seven weeks; in 2026, language models and pretraining, post-training, inference, and the course recap. Course materials: 2026, 2025, 2024.
- Natural Language Understanding and Representation Learning, Master’s course, University of Helsinki, 2020, with Alessandro Raganato.
- Supervision of Master’s theses in language technology, including Jaakko Kuurne, Fine-tuning Open-weight Large Language Model for Expense Classification (2025, co-supervisor).
Industry experience
I have worked in industry for about twenty years, from software engineering to leading AI teams. At Silo AI I was responsible leader of a 6 million euro Business Finland programme on trustworthy and reliable large language models, led a 1.15 million euro work package in the European Commission’s DeployAI consortium, and was co-PI of a 12 million GPU-hour LUMI Extreme Scale allocation for language model pre-training.
-
2025–
Senior Manager, Software Development AMD Silo AI, Helsinki Leading large-model inference infrastructure and performance engineering on AMD GPUs.
-
2024–2025
Senior Manager & EMEA ML Lead Accenture, Helsinki Led the Machine Learning and Computational Science practice and the Model Customization capability at the EMEA Centre for Advanced AI.
-
2023–2024
Head of R&D SiloGen (Silo AI), Helsinki Responsible for research on open multilingual large language models, including Poro.
-
2021–2023
Senior AI Engineer, Lead AI Engineer and Lead AI Scientist Silo AI, London and Helsinki Natural language processing, search, speech recognition and natural language understanding.
-
2020–2021
UK CTO and Global ML Practice Lead Nordcloud, United Kingdom
-
2019–2020
Founder and CEO Basement AI, Finland and United Kingdom A research lab and consultancy in NLP and machine learning.
-
2012–2018
Management consultant Gartner and Accenture, Helsinki Technology and analytics strategy for high-tech and telecoms clients.
-
2006–2011
Software engineering and R&D management Nokia, Tieto and Valuatum
Contact
- aarne.talman@helsinki.fi
- CV
- Curriculum vitae (PDF)
- Google Scholar
- Profile
- ORCID
- 0000-0002-3573-5993
- GitHub
- aarnetalman
- talman
- Research portal
- University of Helsinki