Recently, a startup named MultiOn launched an intelligent agent called Agent Q, claiming an impressive 95.4% success rate in real-world tasks, which has garnered widespread attention.
What's more striking is that MultiOn's CEO, Div Garg, frequently uses a strawberry emoji on Twitter, which inevitably reminds people of OpenAI's enigmatic Q project.
Netizens are filled with curiosity about the technology behind Agent Q. Some speculate that it might be backed by OpenAI's Q* project. MultiOn has not only created a dedicated Twitter account for Agent Q but also themed its background image and basic information around strawberries, which undoubtedly fuels curiosity about its underlying technology.
Agent Q combines search, self-reflection, and reinforcement learning to plan and self-heal. By introducing a new learning and reasoning framework, it has overcome the limitations of previous LLM training techniques, enabling autonomous web navigation.
In tasks simulating online stores, Agent Q demonstrated robust search capabilities. During real booking tasks on Open Table, Agent Q boosted the zero-shot success rate of LLaMa-3 from 18.6% to 81.7%, a 340% increase, and this was achieved after just one day of autonomous data collection.
Although Agent Q performed excellently in evaluation experiments, there is still much room for discussion and improvement in the methods used. For instance, the design of the reasoning algorithm, the choice of search strategies, and aspects of online safety and interaction require further research and optimization.
The emergence of Agent Q is undoubtedly a significant advancement in the field of AI agents, but whether it will become the new darling of the AI world or merely a clever marketing stunt remains to be seen. Nevertheless, the release of Agent Q has brought new possibilities and insights to the development of AI.
References:
https://www.multion.ai/blog/introducing-agent-q-research-breakthrough-for-the-next-generation-of-ai-agents-with-planning-and-self-healing-capabilities