← Volver a los artículos

PROOF OR BLUFF? EVALUATING LLMS ON 2025 USA MATH OLYMPIAD

★ ★ ★ ☆ ☆

Resumen del artículo

Título de Paperzilla
Large Language Models Fail Math Olympiad: Proof or Bluff?

This study evaluated eight large language models (LLMs) on the 2025 USAMO, a challenging math competition requiring rigorous proofs. The models performed poorly, with the best achieving an average score of less than 25%, revealing limitations in logical reasoning and proof generation.

Explícamelo como si tuviera cinco años

Even the smartest computer programs struggle with really hard math problems that need creative solutions and detailed explanations. They're good at calculations but not so good at thinking outside the box.

Posibles conflictos de intereses

None identified

Limitaciones identificadas

Limited Problem Solving Ability
The best-performing LLM only achieved a 24.4% average score on the USAMO problems. This suggests current LLMs are not proficient in solving complex, proof-based mathematical problems.
Lack of Creativity and Reasoning
The LLMs often applied incorrect logic, made unjustified assumptions, and failed to explore alternative solution strategies, indicating a lack of genuine understanding and problem-solving creativity.
Over-Reliance on Pattern Recognition
The models exhibited a tendency to overgeneralize patterns from simple cases to more complex ones without formal proof, demonstrating a superficial approach to mathematical reasoning.
Issues with Solution Clarity and Structure
Many LLMs produced chaotic and difficult-to-interpret solutions, hindering effective evaluation and highlighting the need for improvement in solution presentation.
Artifacts from Optimization Strategies
Behaviors such as consistently "boxing" answers and generating non-existent citations were observed, revealing unintended consequences of training methodologies.

Explicación de la calificación

This paper presents a well-structured evaluation of LLM capabilities on a challenging mathematical benchmark. While the study highlights significant limitations in current LLMs, it provides valuable insights into areas for improvement. The methodology is sound, and the analysis of failure modes is informative. However, the study focuses solely on the USAMO, limiting the generalizability of findings to other mathematical reasoning tasks. It could benefit from evaluating LLMs on a wider range of mathematical problems.

Conviene saber

Este es el análisis de Starter. Paperzilla Pro verifica cada cita, investiga los antecedentes de los autores y las fuentes de financiación, y utiliza razonamiento avanzado con IA para ofrecer información más exhaustiva.

Explorar Pro →

Jerarquía temática

Información del archivo

Título original: PROOF OR BLUFF? EVALUATING LLMS ON 2025 USA MATH OLYMPIAD
Subido: 10 sept 2025, 18:08:41
Privacidad: Público