News
- 10-2026: CulturalRiddles is live! Read the paper and blog, explore the dataset, and try the quiz. The benchmark spans 61 cultures and 51 languages. Interested in contributing or researching? Start with the dataset and quiz. #culturalriddles
- 10-2026: We study how individual attention heads shape response language in One Head Away from Language Confusion, identifying heads whose ablation triggers language switches in ten models.
- 09-2026: In Fallacy Benchmarks, we show that existing benchmarks often measure argument-scheme recognition rather than whether arguments are actually fallacious.
- 09-2026: Our paper How Perturbations Propagate was accepted to the NeurIPS 2026 Workshop InterpScience.
- 09-2026: Our paper First-Token Broadcasters was accepted as a poster at the NeurIPS 2026 Workshop LP4FM.
- 09-2026: We introduce FTB Graph, an end-to-end structural circuit analysis of first-token language broadcasting across six multilingual language models.
- 06-2026: We study why multilingual language models generate in the wrong language in First-Token Broadcasters, using causal attention-head ablation to identify language-identity circuits.
- 09-2026: Our paper CulturalRiddles was accepted to EMNLP 2026 MRL. It introduces a multicultural benchmark of riddles spanning 61 cultures and 51 languages.
- 09-2026: We are building the Last Translation Benchmark and looking for contributors! Contribute samples and become a contributor with coauthorship.
- 09-2026: We are developing Meddies-PII, a multilingual framework for clinical personally identifiable information extraction.
- 09-2026: We are studying how perturbations propagate through large language models across behavior, representations, and attention.
- 11-2025: Our paper SilVar was accepted to the EMNLP 2025 Main Conference. It presents speech-driven multimodal reasoning for visual question answering and object localization.
Contact
Thanks to Vilém Zouhar for inspiring the format of this website.