OMEGA: A Structured Math Benchmark to Probe the Reasoning Limits of LLMs

The article introduces OMEGA, a structured math benchmark designed to evaluate the reasoning capabilities of large language models (LLMs) in mathematical problem-solving, particularly focusing on their ability to generalize beyond learned patterns. While models like DeepSeek-R1 demonstrate proficiency in Olympiad-level mathematics through chain-of-thought reasoning, their reliance on limited techniques such as rule repetition and pattern recognition hampers their performance on more complex, creative tasks that require genuine mathematical insight. Current evaluation methods and datasets are insufficient for thoroughly assessing the true reasoning skills of LLMs, as they often conflate surface-level pattern matching with