Cracking the Code: How LLMs Solve Cryptic Clues, Linguistic Puzzles, and Logic Problems
Kotova, Daria
Kotova, Daria
Author
Supervisor
Department
Natural Language Processing
Embargo End Date
30/05/2026
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This study evaluates LLM reasoning on challenging tasks: logic puzzles, Rosetta Stone puzzles, and cryptic crosswords. We compare 0shot and fewshot prompting across three models, exploring prompt detail, example selection, explanations, and taskspecific information. Key findings: Prompting effectiveness varies by task and model. Human generated prompts outperform AI-generated ones. Task-specific information improves performance. Rosetta Stone puzzles and cryptic crosswords remain difficult. We also introduce MiniCryptic, a labeled cryptic crossword dataset.
Citation
Daria Kotova, “Cracking the Code: How LLMs Solve Cryptic Clues, Linguistic Puzzles, and Logic Problems,” Master of Science thesis, Natural Language Processing, MBZUAI, 2025.
Source
Conference
Keywords
Large Language Model (LLM), reasoning, abilities, prompting