Large language models are getting better with every new development in the AI industry. With each modification and release, LLM is better able to meet different requirements in applications and scenarios. The recently released ChatGPT, developed by OpenAI, which runs on the GPT adapter architecture, is one of the most popular LLM platforms. With the latest GPT-4 architecture, ChatGPT now works well with multimedia data.
The goal of artificial intelligence has always been to develop models and technologies that help automate repetitive tasks and solve complex problems by imitating humans. Although LLMs are successful in processing text when performing computer tasks by taking keyboard and mouse actions, they face some challenges. These challenges include ensuring that the actions created are appropriate for the given task, feasible in the agent’s current state, and actionable. These three challenges are known as Mission Foundations, State Establishment, and Agent Grounding.
A new study has introduced an approach called iterative criticism and optimization (RCI), which uses a pre-trained LLM agent to perform natural-language-guided computer tasks. RCI uses a claim system that prompts the LLM to generate outputs. This is followed by identifying issues with the outputs and thus generating updated outputs.
🚀 Join the fastest ML Subreddit community
RCI improves on all three challenges of the previous methods, ie: task grounding, state grounding, and agent grounding, resulting in better performance in computer task execution. For computer tasks, the RCI claim is implemented in three stages. First, the LLM creates a high-level plan, then it creates an action based on the current plan and state, and finally, it coordinates the action into the correct keyboard or mouse action.
Mission foundations essentially involve producing a high-level plan based on the mission text to ensure that the actions taken by the agent are appropriate for the assigned mission. State grounding, on the other hand, links the high-level concepts derived from the task grounding step to the actual HTML elements present in the agent’s current state, thus ensuring that actions produced by the agent are possible in the current state. Finally, grounding the agent ensures that the actions generated by the agent are actionable and in the correct form.
This new approach to ChatGPT can be used to solve general computer tasks using the keyboard and mouse without the need for additional components. In an RCI prompt, the LLM first identifies the issues with the original answer and, based on those issues, improvises on the answer. The unique feature of this approach is that it requires only a few demos per task, in contrast to current approaches that require thousands of demos per task.
The RCI approach is superior to existing LLM methods for automating computer tasks and is superior to the reinforcement and supervised learning methods of the MiniWoB++ standard. When comparing RCI to prompting the Thought Chain (CoT), a method recognized for its effectiveness in reasoning tasks, the researchers discovered a significant synergistic effect between RCI stimulation and two CoT baselines. In conclusion, RCI appears promising for solving complex computer tasks and problem inference using LLM.
scan the paperAnd githubAnd And project. All credit for this research goes to the researchers on this project. Also, don’t forget to join 18k+ML Sub RedditAnd discord channelAnd Email newsletterwhere we share the latest AI research news, cool AI projects, and more.
Tania Malhotra is a final year from University of Petroleum and Energy Studies, Dehradun, pursuing a BTech in Computer Science Engineering with a specialization in Artificial Intelligence and Machine Learning.
She is passionate about data science and has good analytical and critical thinking, along with a keen interest in acquiring new skills, leading groups, and managing work in an organized manner.
🔥 MUST READ – What is an AI Hallucination? What’s going wrong with AI chatbots? How do you discover the presence of artificial intelligence hallucinations?