Preeti
Researchers from New York University’s Tandon School of Engineering tested various natural language processing systems on solving The New York Times’ daily puzzle, Connections. This word game requires grouping 16 words into four categories based on their connections, ranging from simple to tricky associations. The study aimed to uncover limitations in how these systems process semantic information.
The experiment compared two AI approaches: OpenAI’s ChatGPT models and sentence embedding models like BERT and MiniLM. The AI struggled with the Connections puzzles, with GPT-4 solving only 29% of them, particularly struggling with tricky associations. Providing step-by-step prompts improved performance slightly to 39%, but the AI still fell short of mastering the game.
The study highlighted the benefits of prompting language models to think in more structured ways to improve their performance. This research helps benchmark AI capabilities and explores the potential for models like GPT-4 to be used in creating new word puzzles in the future. The results will be presented at the Institute of Electrical and Electronics Engineers’ 2024 Conference on Games in Milan.
Overall, the study revealed the challenges AI faces in solving language-based puzzles like Connections, shedding light on areas where these systems can improve in processing semantic information and reasoning tasks.
Source link
Source link: https://www.newsweek.com/ai-solve-connections-puzzle-machine-learning-1899440
GIPHY App Key not set. Please check settings