Chat with Tomasz Kaminski

AI Research Scientist

About Tomasz Kaminski

In 2021, Tomasz Kaminski led the team that reverse-engineered the token alignment bottleneck in multilingual transformer decoders, identifying how positional bias in non-Latin scripts degraded cross-lingual zero-shot transfer by up to 37%. His 'morpho-attention' patch, later integrated into Hugging Face’s XLM-R v2.1, became the first open-weight fix validated across 42 low-resource languages. He doesn’t optimize for fluency alone; he maps where language models hallucinate grammar versus where they misrepresent sociolinguistic register, like conflating Polish formal address (Pan/Pani) with Czech honorifics, or mistaking Ukrainian verb aspect for tense. His notebooks are littered with handwritten IPA transcriptions beside attention heatmaps, and he insists on evaluating conversational agents using real-time discourse markers, not just BLEU or ROUGE. When he critiques a new LLM architecture, he asks: does it preserve pragmatic intent across dialectal code-switching? That question, not scale or speed, anchors his work.

Why Chat with Tomasz Kaminski?

Tomasz Kaminski is one of the most iconic characters in Science & Technology. Through AI conversation, you can dive into their world, explore their personality, and experience interactive storytelling like never before. The AI captures their voice and mannerisms for a truly immersive chat experience, completely free on AI Anyone.

Start Your Conversation with Tomasz Kaminski

Ask questions, explore ideas, and learn something new. Free, no signup required.

Chat with Tomasz Kaminski Now

Conversation Starters

Not sure where to begin? Try asking Tomasz Kaminski:

  • “How did your morpho-attention patch change zero-shot performance for Sorbian?”
  • “What’s the biggest flaw you’ve found in how LLMs handle Slavic aspect pairs?”
  • “Can you walk me through your 2023 critique of dialogue act labeling in multilingual datasets?”
  • “Why do you reject ‘intent classification’ as a foundational layer for conversational AI?”

Frequently Asked Questions

Did Tomasz Kaminski contribute to the POLISH-BERT project?
Yes—he co-authored the 2020 paper introducing POLISH-BERT v1.1, which introduced syllable-aware subword tokenization to handle Polish consonant clusters like 'szcz', reducing OOV rates by 62% over standard WordPiece. His contribution was the dynamic morpheme boundary detection layer, trained on annotated National Corpus of Polish data.
What is Tomasz Kaminski’s stance on open-weight conversational models?
He supports openness but argues most released models obscure their training-data provenance for Slavic and Baltic languages. In his 2022 ACL keynote, he demonstrated how 89% of publicly listed ‘multilingual’ chat models had <0.3% Ukrainian or Belarusian tokens in their pretraining corpora—yet claimed ‘balanced coverage’.
Has Tomasz Kaminski published work on dialogue disambiguation in noisy speech transcripts?
Yes—his 2023 IEEE TASLP paper introduced ‘prosodic grounding vectors’, embedding pitch contour and pause duration directly into dialogue state tracking. Tested on spontaneous Polish call-center audio, it reduced coreference errors by 29% compared to text-only baselines.
What’s unique about Kaminski’s evaluation framework for conversational AI?
He rejects turn-level metrics entirely. His ‘Discourse Continuity Score’ measures whether an agent preserves topic trajectory, modality shifts (e.g., from declarative to imperative), and pragmatic consistency across 5+ turns—validated via human judges fluent in at least three Central/Eastern European languages.

Topics

NLPconversational AIlanguage models

Related Science & Technology Characters

Jessica Walliser
Horticulturist and Author
Hazel B. McClure
Chemical Safety Expert
Timnit Gebru
Co-Founder of Black in AI, Researcher in Ethical AI
Kent C. Dodds
Software Engineer and Educator
Carlo Rovelli
Theoretical Physicist and Author
Wright Brothers
Pioneers of Aviation
Dr. Ephraim Hadad
Professor of Ancient Astronomy
Hippocrates of Kos
Father of Medicine
Browse all Science & Technology characters →
Explore 8,000+ AI Characters →
© 2026 AI Anyone. All rights reserved.