Shanda and the Chen Institute’s OMNE Framework for Long-term AI Memory Claims Top Spot on GAIA Leaderboard

Shanda and the Chen Institute’s OMNE Framework for Long-term AI Memory Claims Top Spot on GAIA Leaderboard

Redwood City, October 23, 2024. Drawing on their deep knowledge of the brain and memory, the Tianqiao & Chrissy Chen Institute and Shanda Group’s internal AI team achieved a major breakthrough in artificial intelligence, with their self-developed OMNE Multiagent Framework which took the top position on the GAIA (General AI Assistants) benchmark leaderboard (https://huggingface.co/spaces/gaia-benchmark/leaderboard), published by Hugging Face. OMNE outperformed frameworks from some of the world’s leading institutions, including Microsoft Research. This achievement builds on years of brain research supported by TCCI, equipping agents with Long-Term Memory (LTM) capabilities, which enable the framework to engage in deeper, slower thinking and enhance the decision-making capabilities of Large Language Models (LLMs) in complex problem-solving.
 

This milestone is a major accomplishment for the integrated AI team since the institute’s founder, former Chinese tech giant Tianqiao Chen, announced the “All-In AI Strategy” last year.

OMNE currently boasts an overall success rate of 40.53%, surpassing submissions from well-known institutions such as Meta, Microsoft, Hugging Face, Princeton University, the University of Hong Kong, the British AI Safety Research Institute, and Baichuan, among others. In comparison, GPT-4 equipped with plugins achieved a success rate of only 15%.

GAIA, co-launched by Meta AI, Hugging Face, and AutoGPT, is a benchmarking system designed to rigorously test AI assistants on real-world challenges. It evaluates core competencies such as Reasoning, Multi-Agent Coordination, Web Browsing, and Tool Usage. As one of the most demanding datasets for multi-agent intelligence, topping the GAIA leaderboard showcases the depth of Shanda’s AI expertise and their ability to push the boundaries of innovation.

OMNE is a multi-agent collaboration framework based on long-term memory (LTM). Each agent has the same and independent system structure and can autonomously learn and understand the complete world model, thereby independently understanding the environment. The multi-agent collaboration system based on LTM enables the AI system to adapt to individual behavior changes in real time, optimize task planning and execution, and promote personalized and efficient self-evolution.

This major breakthrough is the integration of long-term memory mechanism, which greatly reduces the search space of MCTS and improves the decision-making ability on complex problems. By introducing more efficient logical reasoning, OMNE not only improves the intelligence level of a single agent, but also significantly enhances the overall capabilities of the multi-agent system by optimizing the collaboration mechanism. This enhancement mechanism is inspired by the study of the columnar structure of the human cerebral cortex. As the basic unit of the brain’s cognitive and behavioral functions, the cortical column realizes information processing through a complex collaboration mechanism. By strengthening the collaboration between single intelligence and agents, the AI model may gradually produce the emergence of cognitive abilities, build an internal representation model, and then promote a leap in the overall intelligence of the system.

“We are incredibly proud to see OMNE top the GAIA leaderboard,” said the head of the AI team. “This achievement demonstrates the vast potential of using long-term memory to drive AI self-evolution and solve real-world problems. We believe that advancing research in Long-Term Memory and AI self-evolution is crucial for the ongoing development and practical application of AI technologies.”

The AI team’s paper on AI Long-Term Memory, “Long Term Memory: The Foundation of AI Self-Evolution”, has been published on arXiv: https://arxiv.org/abs/2410.15665