12 October 2023
Proposes model selection framework for multi-modal reasoning agents
Represents reasoning process as graph with subtask dependencies
Learns to predict execution status for input and model choices
Creates new benchmark dataset for model selection research
Enables dynamic model selection and improves reasoning robustness
Multi-modal reasoning with model selection
This paper proposes a framework to select optimal AI models for each step in a multi-modal reasoning process. An agent decomposes complex tasks into subtasks and sequences AI models collaboratively. Model selection is critical but overlooked. This method represents reasoning as a graph, jointly modeling input, selected models, and subtask dependencies to predict execution status. Experiments on a new benchmark dataset demonstrate the approach enhances reasoning robustness by dynamic model selection.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper