Pre-computed memory or on-the-fly encoding? A hybrid approach to retrieval augmentation makes the most of your compute
Proposes a hybrid approach to retrieval augmentation that significantly outperforms pure memory on multiple question-answering tasks while being cheaper than FiD
Provides a cost-effective method to improve question-answering tasks using a hybrid approach to retrieval augmentation
Imitating Human Behaviour With Diffusion Models
Suggests using diffusion models as observation-to-action models for imitating human behaviour in sequential environments and introduces several innovations to make them suitable for sequential environments
Offers a potential solution for imitating human behaviour in sequential environments using diffusion models
Comparing process-based to outcome-based supervision for natural language tasks
Compares process- and outcome-based approaches for natural language tasks and finds that pure outcome-based supervision produces similar final-answer error rates with less label supervision
Offers insights into the effectiveness of process- and outcome-based approaches for natural language tasks