Our FARABI system ranked among the top three teams in the WMT 2026 Low-Resource Arabic–Asian shared task for translating into Arabic from all five languages (English, Hindi, Bengali, Indonesian, and Urdu), with NTK-Mirror, a lightweight post-hoc controller, improving BLEU without further updating the backbone's weights.
On Deutschlandfunk Kultur's pop culture show Kompressor, I talked about Jev, a lightning-fast "System 1" AI model, and what it could mean for gaming, live streaming, and events.
Player behavior in open-ended games has no ground truth, so we let a Transformer with deep clustering discover it directly from raw game logs, yielding interpretable behavioral patterns that are shared across players.
Quantized large language models can maintain benchmark accuracy while their predictions drift substantially from the original model. We show how measuring next-token distribution divergence alongside accuracy reveals hidden fidelity loss and helps distinguish quantization schemes with similar memory costs.
At IJCAI-ECAI 2026, I spoke with Gesellschaft für Informatik about reliable AI, rule-based systems, uncertainty, healthcare, model efficiency, and the role of human creativity in research.