Core Concepts
LUMOS introduces a unified and modular framework for training open-source language agents, outperforming closed-source models on various complex interactive tasks.
Abstract
Agent LUMOS is a novel framework designed to address the limitations of closed-source language agents by offering affordability, transparency, and reproducibility. It features a unified and modular architecture with distinct modules for planning, grounding, and execution. The framework allows for easy upgrades and generalization across diverse interactive tasks. By collecting high-quality training annotations from various sources, LUMOS demonstrates superior performance compared to GPT-based agents on QA, web tasks, math problems, and multimodal reasoning. The framework showcases competitive performance across different task types and exhibits strong generalization capabilities to unseen tasks.
Stats
LUMOS excels multiple larger open-source agents on held-out datasets for each task type.
LUMOS surpasses GPT agents on QA and web tasks.
LUMOS outperforms open-source agents produced by chain-of-thoughts training.
LUMOS effectively generalizes to unseen tasks, surpassing 33B-scale agents.
Quotes
"LUMOS provides improved or comparable performance with GPT-based or larger open-source agents across various complex interactive tasks."
"LUMOS even surpasses GPT-based agents in web and QA tasks."
"LUMOS outperforms many baseline open-source agents across all held-out datasets."