PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Language Model Reasoning Failures

Peiyang Song, Pengrui Han, Noah Goodman

arXiv:2602.06176v1cs.AIcs.CLcs.LG

TL;DR

LLM reasoning failures remain significant despite strong performance, yet existing evidence is fragmented across individual studies. This survey introduces a two-axis taxonomy, unifies research on failure definitions, causes, and mitigations, and provides a structured collection for studying more reliable reasoning. Its scope includes fundamental, application-specific, and robustness failures across embodied and non-embodied reasoning.

  • Problem

    LLM reasoning failures persist even in seemingly simple scenarios, while prior studies remain fragmented and underexplored as a unified research area.

  • Method

    The survey organizes reasoning as embodied or non-embodied and classifies failures as fundamental, application-specific, or robustness-related, while synthesizing causes and mitigation strategies.

  • Results

    The survey unifies fragmented studies into a structured perspective on systemic LLM reasoning weaknesses and releases a comprehensive research collection.

  • Takeaways & Limitations

    The taxonomy lays a foundation for anticipating, detecting, and mitigating reasoning failures and for building systems that fail more gracefully, transparently, and recoverably.

  • Takeaways & Limitations

    Text-based physical reasoning, formal physics and scientific reasoning, and real-world embodied-agent reasoning remain areas with notable limitations.

Abstract

from arXiv · show

Large Language Models (LLMs) have exhibited remarkable reasoning capabilities, achieving impressive results across a wide range of tasks. Despite these advances, significant reasoning failures persist, occurring even in seemingly simple scenarios. To systematically understand and address these shortcomings, we present the first comprehensive survey dedicated to reasoning failures in LLMs. We introduce a novel categorization framework that distinguishes reasoning into embodied and non-embodied types, with the latter further subdivided into informal (intuitive) and formal (logical) reasoning. In parallel, we classify reasoning failures along a complementary axis into three types: fundamental failures intrinsic to LLM architectures that broadly affect downstream tasks; application-specific limitations that manifest in particular domains; and robustness issues characterized by inconsistent performance across minor variations. For each reasoning failure, we provide a clear definition, analyze existing studies, explore root causes, and present mitigation strategies. By unifying fragmented research efforts, our survey provides a structured perspective on systemic weaknesses in LLM reasoning, offering valuable insights and guiding future research towards building stronger, more reliable, and robust reasoning capabilities. We additionally release a comprehensive collection of research works on LLM reasoning failures, as a GitHub repository at https://github.com/Peiyang-Song/Awesome-LLM-Reasoning-Failures, to provide an easy entry point to this area.

1 Introduction

LLMs have achieved strong results across diverse reasoning tasks, but important failures remain understudied as a unified area. This survey organizes fragmented work to identify patterns, causes, and mitigation strategies for more robust reasoning.

  • LLMs have achieved notable results across linguistic, mathematical, scientific, and reasoning tasks.
  • Whether LLMs use human-like reasoning procedures remains debated, but the survey focuses on their reasoning failures instead.
  • Psychological research motivates studying failures because identifying and learning from them can reveal fundamental elements relevant to improvement.
  • Existing studies examine LLM reasoning failures case by case, leaving the topic fragmented and underexplored as a unified research area.
  • The survey unifies studies, analyzes underlying causes, discusses mitigation strategies, and releases a continuously updated research collection.

2 Definition and Formulation

The survey defines reasoning through embodied and non-embodied categories, then classifies LLM failures by their scope and stability. Its two-axis framework organizes fragmented findings and common evaluation practice.

  • 2.1 Fundamentals of Reasoning: The taxonomy distinguishes embodied reasoning from non-embodied reasoning, which includes informal and formal reasoning.
  • 2.1 Fundamentals of Reasoning: Non-embodied reasoning covers intuitive judgments and explicit rule-based manipulation of symbols in logic, mathematics, and code.
  • 2.1 Fundamentals of Reasoning: Embodied reasoning depends on physical interaction, spatial intelligence, real-time feedback, and behavior constrained by physical laws.
  • 2.2 LLM Reasoning Failures & Common Research Practice: Because LLMs remain largely black-box systems, reasoning abilities are typically assessed behaviorally through outputs on designed prompts and tasks.
  • 2.2 LLM Reasoning Failures & Common Research Practice: LLM reasoning failures are responses that significantly diverge from expected logical coherence, contextual relevance, or factual correctness.
  • 2.2 LLM Reasoning Failures & Common Research Practice: Failures include straightforward poor performance and apparently adequate performance that is unstable across variations.
  • 2.2 LLM Reasoning Failures & Common Research Practice: Figure 1 represents reasoning categories as rows and failure categories as columns, providing a mutually consistent two-axis organization.
  • 2.2 LLM Reasoning Failures & Common Research Practice: Fundamental failures affect diverse downstream tasks, whereas application-specific limitations concern particular domains despite expected competence.

3 Reasoning Informally in Intuitive Applications

LLMs exhibit informal reasoning failures involving cognitive skills, biases, and social judgment. These failures can be fundamental, robustness-related, or application-specific, including brittle performance on social reasoning tasks.

  • Fundamental Cognitive Skills: LLMs show systematic failures in fundamental cognitive skills such as working memory, inhibitory control, and cognitive flexibility.
  • Fundamental Cognitive Skills: Limited working memory causes failures when task demands exceed the model’s capacity to hold and manipulate information.
  • Fundamental Cognitive Skills: LLMs exhibit proactive interference, weak inhibitory control, and difficulty adapting previously learned patterns when contexts or rules shift.
  • Fundamental Cognitive Skills: Even advanced LLMs struggle to infer abstract rules, understand implicit conceptual relationships, and handle symbolic or temporal abstractions reliably.
  • Cognitive Biases: Cognitive biases affect LLM reasoning across tasks, reflecting content effects, confirmation bias, order bias, anchoring, and framing effects.
  • Cognitive Biases: These biases are linked in the survey to training data, causal masking, and alignment processes that can amplify human biases.
  • Cognitive Biases: Proposed bias mitigations include data curation, adversarial training, prompt engineering, and output filtering.
  • Implicit Social Reasoning: LLMs struggle with Theory of Mind tasks, and newer models remain brittle under simple phrasing perturbations despite solving many standard tests.

4 Reasoning Formally in Logic

The survey examines formal reasoning failures across increasingly explicit logical and mathematical paradigms, from natural-language relations to benchmark-based evaluation and arithmetic. It distinguishes fundamental failures from targeted and robustness-related weaknesses, while reviewing mitigation strategies.

  • Formal reasoning spans implicit natural-language structures, explicit symbolic representations, and mathematical systems, motivating a progression from less to more formal paradigms.
  • 4.1 Logic in Natural Languages: Reversal curse occurs when models fail to infer “B is A” after seeing “A is B,” even though the relation is bidirectionally equivalent.The failure has been observed broadly across tasks on GPT-based models and attributed to directional training objectives and asymmetric representations.
  • 4.1 Logic in Natural Languages: Compositional reasoning fails when models cannot integrate individually known facts, with performance worsening as reasoning depth increases or distractors are added.Chain-of-thought and graph-structured reasoning data make intermediate reasoning explicit, while latent compositionality remains harder to achieve.
  • 4.1 Logic in Natural Languages: Purpose-built datasets expose targeted weaknesses in converse relations, syllogisms, causal inference, and other specific logical relations, while enabling direct fine-tuning as mitigation.
  • 4.2 Logic in Benchmarks: Logic-preserving transformations reorder options, premises, events, or irrelevant context while preserving task semantics, making robustness failures measurable in benchmarks.Performance drops under these transformations indicate that strong static benchmark scores may not guarantee consistent reasoning.
  • 4.2 Logic in Benchmarks: Math word-problem variants perturb numeric values, irrelevant entities, or known-versus-unknown components to reveal deeper robustness limitations.
  • 4.2 Logic in Benchmarks: Coding benchmarks test robustness through edits to docstrings, function and variable names, and control-flow logic such as swapped if-else cases.
  • 4.3 Arithmetic & Mathematics: Counting and arithmetic remain fundamental failures: models struggle with character-level operations, larger operands, multiplication, and middle digits, producing inconsistent downstream reasoning.Models can fail at identifying the last digit while succeeding at the first digit, and these inconsistencies affect practical temporal reasoning.

5 Reasoning in Embodied Environments

Embodied reasoning requires physical grounding, spatial understanding, and real-time interaction, yet LLMs and VLMs exhibit failures from text-based commonsense tasks through dynamic environments. The survey reviews these failures and mitigation strategies across text, vision, and action.

  • LLMs lack true physical grounding, causing systematic errors and unrealistic predictions even in basic physical reasoning.Research on physical reasoning failures remains sparse compared with the broader study of reasoning.
  • 5.1 1D – Text-Based Physical Reasoning Failures: Text-based physical commonsense failures involve object attributes, spatial relations, physical laws, and object affordances that humans acquire through embodied experience.Text-only models rely on linguistic data without direct perceptual or embodied experience, limiting real-world understanding.
  • 5.1 1D – Text-Based Physical Reasoning Failures: Formal physics and scientific reasoning require multi-step deduction, quantitative reasoning, and correct law application, areas where state-of-the-art models retain notable deficits.These failures constrain LLM application in scientific domains, including complex problems and real-world scientific discovery.
  • 5.1 1D – Text-Based Physical Reasoning Failures: Text-based mitigation combines structured physical-knowledge training, prompting, and integration with external tools to address limitations of the text modality.
  • 5.2 2D – Perception-Based Physical Reasoning Failures: VLMs fail at anomaly detection, object counting, overlap identification, and spatial-relation understanding from static images.
  • 5.2 2D – Perception-Based Physical Reasoning Failures: Adding visual inputs does not resolve physical commonsense or advanced-physics failures, and dynamic spatial tasks expose gaps in composing and applying spatial knowledge.These tasks include motion prediction, object manipulation, spatial communication, and embodied planning.
  • 5.2 2D – Perception-Based Physical Reasoning Failures: Perception failures arise from reliance on textual priors, difficulty binding multiple objects, and insufficient spatial semantics in visual inputs alone.
  • 5.3 3D – Real-World Physical Reasoning Failures: In 3D embodied settings, models struggle with spatial estimation, object localization, multi-step manipulation, prompt-sensitive plans, adversarial manipulation, and harmful actions.The survey calls for robust, self-correcting, and safety-aware embodied systems before real-world deployment.

6 Discussions & Conclusion

The survey identifies cross-cutting failure patterns and argues that structured failure analysis is necessary for more resilient, reliable LLM reasoning. It also highlights incomplete root-cause analyses, benchmark gaps, and underrepresentation of interactive settings.

  • Failure Axis: Fundamental failures arise from architectural or training constraints and can appear across informal, formal, and embodied reasoning.Examples include the reversal curse, confirmation bias, and working-memory limitations causing proactive interference.
  • Failure Axis: Perturbation-based testing detects robustness issues by measuring output changes after semantically preserving variations across domains.Examples include reordering answer options, renaming code variables, and paraphrasing moral dilemmas.
  • Future Directions: Root-cause analyses remain incomplete for compositional reasoning, higher-order theory of mind, physical commonsense, and multi-agent planning failures.The survey calls for linking behavioral errors to mechanisms such as faulty attention-head coordination or insufficient representation alignment.
  • Future Directions: Unified, persistent, and dynamically evolving benchmarks could preserve historical failure cases while tracking whether failures persist in newer models.Proposed strategies include private benchmarks, evolving suites, event-driven tests, adversarial sections, multi-level difficulty, and cross-domain compositions.
  • Scope: The literature may over-represent some failure types, while multi-turn, interactive contexts and multi-agent coordination remain underrepresented.The survey identifies these settings as closer to real-world deployment conditions and calls for broader benchmark diversity.
  • Conclusion: By unifying fragmented observations into a two-axis taxonomy, the survey lays groundwork for anticipating, detecting, and mitigating reasoning failures.It frames sustained attention to failure modes as important for models that fail more gracefully, transparently, and recoverably.

A Taxonomy

The survey organizes LLM reasoning failures along two primary reasoning axes: embodied versus non-embodied, with non-embodied reasoning divided into informal and formal categories.

  • Taxonomy: The taxonomy distinguishes reasoning that involves physical interaction from reasoning that does not.Non-embodied reasoning is further divided according to whether reasoning primarily requires instinct or logic.
  • Taxonomy: Informal, formal, and embodied reasoning each receive a dedicated survey section and category-specific taxonomy.The visual taxonomies are presented in Figures 3, 4, and 5.

B Artifacts

The authors release a continuously updated GitHub collection of research on LLM reasoning failures to make the area easier to enter and support future work.

  • Artifacts: The GitHub repository collects research works on LLM reasoning failures as an accessible entry point to the field.The collection is intended to facilitate future research and will be updated as the area advances.

C Other Emerging Areas of Reasoning

The survey identifies emerging reasoning areas spanning diverse media, general-purpose frameworks, and verifiable formal systems. It encourages systematic failure analysis as these areas mature and evaluations expand beyond traditional LLMs.

  • Emerging Areas: Emerging reasoning domains remain difficult to study systematically because generalizable failure investigations are limited by their novelty.The survey argues that its methodology for identifying and analyzing generalizable failures will become more valuable as these fields mature.
  • Taxonomy: The survey provides separate taxonomies for informal, formal, and embodied reasoning failures.Figures 3, 4, and 5 visualize these category-specific breakdowns.
  • Diverse Media: Language-vision advances enable visual and spatial reasoning, creating opportunities for systematic analyses of failure cases.The survey characterizes these areas as moving toward stable growth guided by incremental improvements and failure analysis.
  • General Frameworks: Analogical reasoning and inference-time scaling are emerging general-purpose frameworks whose effectiveness against established reasoning failures remains to be examined.Such studies could clarify the causes of reasoning errors and inform more reliable real-world deployments.
  • Verifiable Reasoning: Formal math and science validation offers a route toward grounding LLM reasoning in verifiable systems.The survey discusses proof search, agentic tools, and automated conjecturing, including applications requiring rigorous safety guarantees.

D Other Important LLM (Non-Reasoning) Failures

The survey also discusses important LLM failures outside reasoning, including hallucinations, social bias, and security or privacy risks. Existing mitigations help but remain limited across settings.

  • Trustworthiness: Hallucinations produce fluent, confident, factually incorrect or fabricated text, motivating retrieval augmentation and calibration methods.Retrieval grounds generation in external sources, while calibration aligns expressed confidence with accuracy.
  • Trustworthiness: Retrieved content can still be misinterpreted or misused, and reliable calibration remains difficult across diverse domains and prompt types.
  • Fairness: LLMs inherit harmful social and ethical biases from human-generated training data, creating risks in recruitment, healthcare, and law enforcement.
  • Fairness: Benchmarking, mitigation, and regulation have not eliminated biased outputs, especially under adversarial prompts, while implicit internal representations complicate debiasing.
  • Safety: AI safety includes protecting systems against malicious exploits and unauthorized sensitive-information exposure, alongside watermarking for detecting generated content.

E Examples

The examples section presents representative cases organized by the survey’s reasoning categories, showing failures in mathematical, coding, visual, spatial, and embodied tasks. These examples make failure manifestations more concrete.

  • Overview: The survey organizes representative examples in tables by reasoning sections and subsections to clarify how each failure manifests.
  • Formal Reasoning: A symbolic Math Word Problem template generates variants by changing numeric values, revealing success on one instance but failure on another.
  • Formal Reasoning: Adding irrelevant context to a Math Word Problem causes a model to fail despite an unchanged underlying question.
  • Formal Reasoning: Subtle changes to a code prompt, such as modifying a docstring or function name, can cause failures on otherwise equivalent problems.
  • Formal Reasoning: When two builtin functions are swapped, models prefer an incorrect statistically common continuation over a correct unusual one.
  • Formal Reasoning: Examples include unreasonable answers or endless thinking when Math Word Problems contain faulty assumptions.
  • Embodied Reasoning: Advanced models fail basic visual and physical-scene tests, including anomalous scenes that require simple physical understanding.
  • Embodied Reasoning: Visual grounding can favor an upright cylinder even when the prompt specifies the cylinder resting on its round side.
Loading 2602.06176v1…