Expedition Aya, Cohere For AI Evaluating Multilingual Reward Models Developing a multilingual evaluation framework for diverse reward models with appropriate metrics, datasets and baselines. Mislabeled Instances in Aya Datasets Automated identification label issues in Aya instruction and red-teaming datasets using language detection, perplexity analysis and LLM-as-a-Judge techniques. Global Exams Collecting exams in local languages from all over the world and structuring as a standardized collection. Maya: Multimodal Aya Curating a new multilingual image-text dataset and training Aya/SigLIP models with that data. Natural Language Processing Process-supervised Reward Models (PRMs) Curating high quality benchmark for PRMs and evaluate various LLMs and ORMs. Complex Named Entity Recognition in Bengali Identification and classification of named entities in Bengali language texts using transformer models with CRF and Random Forest baselines. Grammatical Error Detection Leveraging Transformer-based Token Classification Developing models for detecting grammar, spelling, and punctuation errors in Bengali text using transformer-based architectures, ensemble learning, and rule-based methods. Reinforcement Learning Q-TCP: Improving TCP Congestion Control with Machine Intelligence Implementing a deep Q-Learning agent for TCP congesting control with agent state and reward design. HummingBird Developing a Reinforcement Learning game using Unity ML-Agents and training PPO agents for competiting with human.