THaMES: An End-to-End Tool for Hallucination Mitigation and Evaluation in Large Language Models

This paper investigates how large language models handle complex reasoning tasks and introduces new benchmarks to measure their true capabilities. The authors point out that while LLMs perform well on many problems, their reasoning often relies on surface-level patterns rather than deep understanding. To study this, they design challenging tests that require multi-step logic and careful problem solving. Their experiments show that even advanced models struggle with consistency and often fail on harder cases. The work emphasizes the gap between fluent language generation and reliable reasoning, highlighting the need for better methods to build models that can think more systematically and robustly.