NYT-Connections

This repository contains the NYT-Connections dataset proposed in the work NYT-Connections: A Deceptively Simple Text Classification Task that Stumps System-1 Thinkers. This work will be published in the 31st International Conference on Computational Linguistics in January of 2025.

Authors: Angel Yahir Loredo Lopez, Tyler McDonald, Ali Emami

Paper Abstract

Large Language Models (LLMs) have shown impressive performance on various benchmarks, yet their ability to engage in deliberate reasoning remains questionable. We present NYT-Connections, a collection of 358 simple word classification puzzles derived from the New York Times Connections game. This benchmark is designed to penalize quick, intuitive ``System 1'' thinking, isolating fundamental reasoning skills. We evaluated six recent LLMs, a simple machine learning heuristic, and humans across three configurations: single-attempt, multiple attempts without hints, and multiple attempts with contextual hints. Our findings reveal a significant performance gap: even top-performing LLMs like GPT-4 fall short of human performance by nearly 30%. Notably, advanced prompting techniques such as Chain-of-Thought and Self-Consistency show diminishing returns as task difficulty increases. NYT-Connections uniquely combines linguistic isolation, resistance to intuitive shortcuts, and regular updates to mitigate data leakage, offering a novel tool for assessing LLM reasoning capabilities.

Puzzle Description

NYT-Connections puzzles are a subset of the New York Times' daily Connections contests. Each puzzle consists of 16 words, with the goal being to sort these words into 4 correct groupings of varying difficulty. The base game allows for hints when a solution is one word away from being a correct group, and allows up to 4 mistakes. Thus, the goal is to correctly identify all 4 groups without committing 4 mistakes.

Data Description

date - the original date the contest was offered.

contest - the title string for the contest.

words - the collection of 16 words available for use in puzzle solving.

answers - an array of objects, where each object is a correct group and contains:

• answerDescription - the group name
• words - the 4 words that classify into this group

difficulty - the difficulty of the puzzle as rated by community contributors (should such a rating be obtained, otherwise null).

Citation

@misc{lopez2024nytconnectionsdeceptivelysimpletext,
      title={NYT-Connections: A Deceptively Simple Text Classification Task that Stumps System-1 Thinkers}, 
      author={Angel Yahir Loredo Lopez and Tyler McDonald and Ali Emami},
      year={2024},
      eprint={2412.01621},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.01621}, 
}