# Large Language Model Reasoning Failures

**Authors:** Peiyang Song, Pengrui Han, Noah Goodman  
**arXiv:** [2602.06176](https://arxiv.org/abs/2602.06176)  
**Date:** 2026-02-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2602.06176v1/

## Key points

- Large Language Models (LLMs) have exceptional reasoning capabilities across various tasks
- A detailed survey on reasoning failures in LLMs introduces a novel categorization framework:
- Reasoning categorized into embodied and non-embodied types (informal and formal reasoning)
- Reasoning failures classified into fundamental, application-specific, and robustness issues
- The survey analyzes each reasoning failure with clear definitions, existing studies, root causes, and mitigation strategies
- A GitHub repository has been released with research works on LLM reasoning failures for easy access
- Complete root cause analyses for failures like compositional reasoning breakdowns and physical commonsense gaps
- Unified failure benchmarks to track persistence over time
- Injecting failure principles into general reasoning benchmarks for better evaluation comprehensiveness
- Expanding benchmark diversity to capture realistic interactive settings better
- Understanding and categorizing failure modes are crucial for building resilient systems as seen in early fault-tolerance research in computing and safety-critical industries
- Sustained attention to anticipating, detecting, and mitigating reasoning failures is essential for future LLMs to excel at tasks gracefully and transparently

## Summary

Large Language Models (LLMs) have demonstrated exceptional reasoning capabilities, achieving impressive results across various tasks. However, despite these advancements, significant reasoning failures persist. To address these shortcomings comprehensively, a detailed survey focusing on reasoning failures in LLMs has been presented. The survey introduces a novel categorization framework that distinguishes reasoning into embodied and non-embodied types, with the latter further divided into informal (intuitive) and formal (logical) reasoning. Additionally, reasoning failures are classified into three types: fundamental failures intrinsic to LLM architectures affecting downstream tasks broadly; application-specific limitations appearing in specific domains; and robustness issues characterized by inconsistent performance across minor variations. The survey delves into each reasoning failure by providing clear definitions, analyzing existing studies, exploring root causes, and presenting mitigation strategies. By consolidating fragmented research efforts, the survey offers a structured perspective on systemic weaknesses in LLM reasoning to guide future research towards developing stronger and more reliable reasoning capabilities. Furthermore,a comprehensive collection of research works on LLM reasoning failures has been released as a GitHub repository to facilitate easy access to this area of study. Moving forward,the survey highlights several gaps and opportunities for future directions in addressing reasoning failures in LLMs.It emphasizes the need for complete root cause analyses for various failures such as compositionalreasoning breakdowns and physical commonsense gaps.The field could benefit from unified failure benchmarks spanning all types of failures to enable longitudinal tracking of persistence over time.Moreover,injecting failure principles into generalreasoning benchmarks could enhance evaluation comprehensiveness and resistance to short-term overfitting. The survey also acknowledges potential biases in existing literature towards certain types ofreasoning or failure categories while underrepresenting others like multi-turn interactive contexts.Future work should expand benchmark diversity to capture realistic interactive settings better. Overall understandingand categorizingfailure modes are crucial for building resilient systems as evidenced by early computing fault-tolerance research and incident analysis in safety-critical industries. In conclusion,as reasoning-specialized models become more prevalent,sustained attention to anticipating,detecting,and mitigating reasoning failures will be essential for ensuring that future LLMs not only excel at performing tasks but also handle failures gracefully and transparently.