Reproduce “REGENT: A Retrieval-Augmented Generalist Agent That Can Act In-Context in New Environments” on OpenResearch

About this paper

Kaustubh Sridhar, Souradeep Dutta, Dinesh Jayaraman, Insup Lee

REGENT is a retrieval-augmented generalist AI agent that achieves rapid in-context adaptation to new, unseen environments without fine-tuning. It demonstrates superior zero-shot generalization performance compared to larger, purely parametric generalist agents, using significantly fewer parameters and pre-training data by leveraging a novel semi-parametric architecture that integrates retrieved expert demonstrations.

Key insights

• Direct retrieval (even simple nearest-neighbor search) can be a surprisingly effective baseline for rapid adaptation in new environments, challenging the necessity of large models for generalization.
• Pre-training a transformer to integrate retrieved expert demonstration context is crucial for robust performance, outperforming simple retrieval augmentation applied to pre-trained agents at inference.
• Semi-parametric architectures that combine a parametric model with a non-parametric retrieval component can achieve superior generalization and resource efficiency compared to purely parametric large models.

Summary by alphaXiv.