Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
110 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
44 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Puzzle Solving using Reasoning of Large Language Models: A Survey (2402.11291v3)

Published 17 Feb 2024 in cs.CL and cs.AI

Abstract: Exploring the capabilities of LLMs in puzzle solving unveils critical insights into their potential and challenges in AI, marking a significant step towards understanding their applicability in complex reasoning tasks. This survey leverages a unique taxonomy -- dividing puzzles into rule-based and rule-less categories -- to critically assess LLMs through various methodologies, including prompting techniques, neuro-symbolic approaches, and fine-tuning. Through a critical review of relevant datasets and benchmarks, we assess LLMs' performance, identifying significant challenges in complex puzzle scenarios. Our findings highlight the disparity between LLM capabilities and human-like reasoning, particularly in those requiring advanced logical inference. The survey underscores the necessity for novel strategies and richer datasets to advance LLMs' puzzle-solving proficiency and contribute to AI's logical reasoning and creative problem-solving advancements.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Panagiotis Giadikiaroglou (1 paper)
  2. Maria Lymperaiou (32 papers)
  3. Giorgos Filandrianos (26 papers)
  4. Giorgos Stamou (55 papers)
Citations (14)

Summary

Exploring Puzzle Solving through the Lens of LLMs: A Comprehensive Survey

Introduction to Puzzle Solving with LLMs

The field of AI and ML has increasingly turned its attention toward understanding and enhancing the puzzle-solving capabilities of LLMs. This surge of interest is rooted in the belief that puzzles, with their inherent complexity and diversity, serve as an ideal proving ground for testing and refining the logical reasoning, strategic planning, and creative problem-solving abilities of LLMs. Notably, this survey presents a thorough examination of methodologies leveraged by LLMs in approaching puzzles, categorized under rule-based (both deterministic and stochastic) and rule-less challenges, encompassing riddles, programming puzzles, and commonsense reasoning tasks.

Methodological Approaches in LLMs for Puzzle Solving

Fundamental and Advanced Prompting Techniques

The survey identifies a spectrum of prompting techniques from fundamental strategies such as few-shot learning, chain-of-thought (CoT), and self-consistency, to more sophisticated approaches like the Plan-and-Solve method, the use of hints, and self-questioning techniques. These strategies underline the need for LLMs to exhibit a deep understanding of puzzles alongside the ability to dissect and approach them in a systematic, step-wise manner.

Puzzle Translation and Fine-tuning

A significant portion of the paper explores the puzzle translation method, highlighting how LLMs convert textual puzzles into forms more amenable to computational solving strategies, including translating puzzles to code or symbolic representations. Fine-tuning emerges as a pivotal tactic across different categories of puzzles, indicating its effectiveness in tailoring LLMs to specific reasoning tasks or puzzle types.

Insights from Datasets, Benchmarks, and Tasks Analysis

The survey meticulously compiles and analyzes an extensive array of datasets corresponding to various puzzle types. Its findings highlight a disproportionate focus on rule-based deterministic puzzles and a notable gap in datasets for rule-based stochastic puzzles and rule-less programming challenges. This imbalance underscores a potential area for future research and dataset development, aiming to broaden the scope of puzzles LLMs can tackle effectively.

Discussions and Directions for Future Research

The paper prompts a critical discussion on the current capabilities and limitations of LLMs in puzzle solving, drawing attention to the nuanced differences in performance across puzzle types. It stresses the importance of developing diverse and comprehensive datasets to challenge and refine the problem-solving abilities of LLMs further. Additionally, the survey speculates on the future of puzzle-solving LLMs, hinting at the potential for innovations in automatic puzzle generation and the broader implications of advanced puzzle-solving capabilities for AI's logical reasoning and creative problem-solving advancements.

Conclusion

This survey serves as a cornerstone for understanding the intersection of LLMs and puzzle solving, providing a detailed overview of the strategies, methodologies, and benchmarks that define this research area. By pinpointing existing gaps and highlighting potential avenues for future exploration, the paper shapes the pathway for advancing the capabilities of LLMs, pushing the boundaries of what these models can achieve in the context of complex reasoning and problem-solving tasks.

Youtube Logo Streamline Icon: https://streamlinehq.com