search

zjunlp / KnowAgent

KnowAgent: Knowledge-Augmented Planning for LLM-Based Agents
https://zjunlp.github.io/project/KnowAgent/
Apache License 2.0
171 stars 13 forks source link

how does the model predict only one step at a time #7

Open annyWangAn opened 2 hours ago

annyWangAn commented 2 hours ago

Thank you for posting such a great piece! I encountered some confusion. When constructing training data, I notice the output is the entire trajectory. But in the inference phase, the model only predicts action path or thought or action at a time. The training data is not divided into steps to output the different steps. So how does the model achieve step-by-step prediction? looking forward to your response! thanks

annyWangAn commented 2 hours ago

how does the model predict only one step at a time? not the entire trajectory