Northeastern University and MIT researchers have created Reflexion, a self-reflective natural language agent that they claim can improve GPT-4 performance by an astounding 30%.
The agent can generate new prompts autonomously until it reaches the correct answer, by analyzing its mistakes and taking them into account. Reflexion was tested on two benchmarks, AlfWorld and HotPotQA, and achieved 97% success and a 17% improvement over the base agent, respectively.
Reflexion appears to be an important step toward natural language agents that can learn through trial and error, without constant human intervention. This approach enables agents, based on their experience history, to generate new ideas, explore previously unseen state spaces, and develop more precise action plans. It also allows agents to solve tasks and environments that were previously thought to be impossible.
Self-reflective agents such as Reflexion may play an important role in improving the performance of large-scale language models such as GPT-4, paving the way for more intelligent and autonomous AI systems.
Sources: MarkTechPost and arXiv
Images: MidJourney, Prompts by Lewis Farrell
Interesting story? Please click on the ? button below!