What is seed?

SEED is a research project and codebase for training AI agents that need to complete tasks made up of many steps, like navigating a simulated house or shopping on a website. It comes from an academic paper and includes the code needed to reproduce its experiments, along with links to the paper and a sample trained model. The core idea is that the same AI model plays two roles during training. First it acts in an environment and produces a full attempt at a task. Then, using its own completed attempt, it writes a short reflection describing what worked, called a hindsight skill. That reflection is used to nudge the model's future behavior during training, without needing the reflection to be present when the model is actually being used later. As training continues, the model keeps re-analyzing its own newer attempts, so the reflections stay matched to how the model currently behaves rather than becoming outdated. The README reports that this approach was tested on three types of tasks: a simulated household environment, search based question answering, and a shopping website simulator, and that it performed better than a more standard training method across all three. It also reports that learning from these self generated reflections works better than simply showing the model example reflections as a prompt at answer time. Setting this up is meant for people already comfortable with machine learning research code. It requires creating multiple separate Python environments with specific package versions, downloading pretrained detection models and datasets, and for one of the three tasks, spinning up a local search and retrieval server with its own dependencies. This is not a beginner friendly tool, it is a research codebase meant to be run on machines with capable GPUs.