Researchers · Interviewed at NeurIPS 2025
Cohere

NeurIPS 2025 · December 22, 2025 · 9 min
How do we give language models explicit control over what they generate—without retraining from scratch or exploding data requirements?
In this #NeurIPS 2025 highlight, Julia Kreutzer (Senior Research Scientist at @CohereAI ) presents "Treasure Hunt", a simple yet powerful idea: • add lightweight markers during fine-tuning that models can later retrieve at inference time to unlock better performance, especially for underrepresented domains, tasks, and languages.