Large Language Model Reasoning Failure Case Library: A Systematic Review of LLM Reasoning Boundaries and Pitfalls

This article introduces the Awesome-LLM-Reasoning-Failures project, which systematically collects and categorizes failure cases of large language models (LLMs) in reasoning tasks, covering multiple reasoning scenarios such as mathematics, logic, and common sense. It analyzes the causes of failures and their application value, providing empirical data for understanding LLM capability boundaries and improving model reliability.

In the AI field, we often focus on model successes, but Donald Knuth noted: "We learn more from failure than from success."

Significance of studying LLM reasoning failures:

1. Reveal model capability boundaries and help set reasonable application expectations;
2. Identify systemic flaws in model architecture or training methods;
3. Provide specific directions for model improvement (e.g., designing solutions for specific error patterns).

The project classifies LLM reasoning failures into four categories:

Mathematical Reasoning Failures

• Arithmetic errors (incorrect operations on large numbers), symbol confusion (using "x" as both an unknown variable and a multiplication sign), skipped steps, unit confusion

Logical Reasoning Failures

• Affirming the consequent fallacy, denying the antecedent fallacy, transitivity errors, conditional sentence comprehension biases

Common Sense Reasoning Failures

• Physical common sense errors (stones floating on water), time sequence errors, causal confusion, lack of social common sense

Multi-step Reasoning Failures

• Loss of intermediate results, goal drift, premature convergence, circular reasoning

Reasoning failures can be attributed to:

1. Architectural Limitations: The unidirectional nature of the Transformer autoregressive mechanism and lack of explicit symbolic reasoning;
2. Training Data Biases: Insufficient examples of specific reasoning tasks or presence of errors/biases in data;
3. Attention Mechanism Limitations: Soft attention struggles to accurately track relationships between multiple entities;
4. Lack of Metacognition: Inability to self-monitor and correct errors, leading to persistent mistakes.

Value of the project for different groups:

• Model Developers: Provides test cases to evaluate new models and identify improvement directions;
• Application Developers: Helps design robust systems (e.g., integrating calculator APIs to compensate for mathematical flaws);
• Users: Enables setting reasonable expectations and avoiding blind trust in AI in critical scenarios;
• AI Security Researchers: Identifies vulnerabilities that could be exploited maliciously.

Awesome-LLM-Reasoning-Failures represents the AI community's attitude of facing problems directly. Systematic analysis of failures is a key path to improving AI. As LLMs evolve, new failure patterns may emerge, making continuous research necessary.

Readers are advised to visit the project's GitHub repository to view specific failure cases and gain a deeper understanding of the actual capabilities of LLMs.