Fine-Grained Failure Mode Decomposition for Language Models

1,023 concepts752 papers31 proposals

When a student fails a complex mathematics exam, a simple score of 50% tells us very little about their underlying struggles. Did they misread the text of the questions? Did they understand the concepts but make simple arithmetic slips? Or did they lack the foundational formulas entirely?

For years, evaluation in artificial intelligence has treated Large Language Models (LLMs) and autonomous agents like students receiving a single, aggregate grade. Benchmarks routinely report a single accuracy score, or rely on coarse linguistic metrics like BLEU or ROUGE. However, as models are deployed in high-stakes environments—from clinical settings to autonomous software engineering—these binary or aggregate metrics fail. They obscure why a model fails, masking critical vulnerabilities behind a facade of high fluency.

To solve this, a major research direction has emerged around fine-grained failure mode decomposition. By replacing monolithic error metrics with structured, multi-dimensional taxonomies, researchers can isolate the exact cognitive, perceptual, or procedural bottlenecks within a model's reasoning chain.

The Limits of Aggregate Metrics

In specialized domains, an incorrect output can range from a harmless stylistic repetition to a catastrophic safety violation. Treating these outcomes identically prevents developers from understanding systemic model risks.

In the medical domain, for example, Yu et al. demonstrate in Evaluating Progress in Automatic Chest X-Ray Radiology Report Generation that standard natural language processing metrics fail to reflect clinical reality. They argue for a diagnostic approach that categorizes failures into specific clinical error types—such as false predictions, omissions, or incorrect anatomical locations—to expose when highly fluent reports contain dangerous hallucinations. Similarly, when evaluating AI-generated clinical consultation notes, Mark Perera in Human Evaluation and Correlation with Automatic Metrics in Consultation Note Generation utilizes a multi-dimensional error taxonomy to separate critical omissions from minor linguistic repetitions, providing a granular view of clinical safety that aggregate scores miss.

This need for granular categorization extends to structural text analysis. In Named Entity Recognition - Is There a Glass Ceiling?, the authors employ a multilevel error taxonomy to separate data-level noise from document-level context requirements, allowing developers to distinguish between inherent model limitations and external dataset flaws.

Decomposing Perception from Reasoning

One of the most powerful applications of failure decomposition is separating a model's input comprehension from its subsequent logical processing. When a multimodal model fails to solve a geometry problem or interpret a chart, it is vital to know whether the vision encoder misread the visual tokens or the language backbone failed the logic.

[Input Image/Prompt] 
       │
       ▼
[Perception Layer] ──► Failure Mode A: Visual Misreading / Tokenization Error
       │
       ▼
[Reasoning Engine] ──► Failure Mode B: Logical Fallacy / Calculation Error
       │
       ▼
[Final Output]

Wang et al. formalize this boundary in Benchmarking Multimodal Mathematical Reasoning with Explicit Visual Dependency by systematically classifying model failures into distinct categories: perception, calculation, and logic. This framework isolates whether a model's bottleneck resides in its visual extraction layer or its cognitive architecture. Kazemi et al. apply a similar philosophy in ReMI: A Dataset for Reasoning with Multiple Images, using perception-reasoning error decomposition to pinpoint exactly why models struggle to synthesize information across multiple visual sources.

This structural decomposition is equally valuable in text-only quantitative tasks. In MedCalc-Bench: Evaluating Large Language Models for Medical Calculations, Nikhil Khandekar maps failures to a taxonomy of knowledge-based, extraction-based, and computation-based errors. This reveals whether a model failed because it could not retrieve the correct medical formula or because it executed the arithmetic incorrectly. Similarly, the diagnostic framework in DateLogicQA: Benchmarking Temporal Biases in Large Language Models splits failures into representation-level errors (tokenizer misinterpretations) and logical-level errors (reasoning failures), isolating the exact layer responsible for temporal biases.

Disentangling Knowledge Gaps from Logical Flaws

Even when a model perfectly perceives its input, its internal processing can break down in two distinct ways: it can suffer from a lack of factual information (a knowledge gap), or it can fail to synthesize the information it does possess (a reasoning flaw).

Cheng et al. explore this distinction in Evaluating Hallucinations in Chinese Large Language Models by separating "imitative falsehoods"—misconceptions learned directly from biased training data—from factual errors caused by a complete absence of information. When evaluating how models handle complex multi-step deductions, Qiao et al. in We-Math: Does Your Large Multimodal Model Achieve Human-like Mathematical Reasoning? categorize failures into insufficient knowledge versus inadequate generalization. This taxonomy successfully exposes instances of rote memorization, where a model manages to output a correct final answer despite failing the intermediate logical steps.

When the failure is purely logical, taxonomies help map the precise nature of the cognitive slip. The authors of A Closer Look at the Self-Verification Abilities of Large Language Models in Logical Reasoning deploy a hierarchical taxonomy that separates formal fallacies (structural logical breakdowns) from informal fallacies (semantic or contextual mistakes), revealing whether a model genuinely understands formal rules or is merely relying on surface-level linguistic patterns.

From Diagnostics to Course Correction

The ultimate value of decomposing failure modes is not merely to document errors, but to fix them. By transforming a vague failure into a specific taxonomic label, researchers can provide targeted feedback that allows models to self-correct or enables supervisor models to steer them in real time.

[Agent Trajectory] ──► [Supervisor Model] ──► [Taxonomy Lookup] ──► [Targeted Feedback]
                                                                            │
                                                                            ▼
                                                                   [Course Correction]

This mechanism is highly effective in long-horizon agent tasks. Gandhi et al. introduce this paradigm in When Agents go Astray: Course-Correcting SWE Agents with PRMs. They use a structured taxonomy of agent failures—spanning specification, reasoning, and coordination—to power a Process Reward Model (SWE-PRM). Instead of giving a generic penalty, the supervisor identifies specific errors, such as tool hallucinations or task constraint violations, and provides natural language feedback that steers the agent back toward its objective.

Similarly, Tong et al. demonstrate in Can LLMs Learn from Previous Mistakes? Investigating LLMs' Errors to Boost for Reasoning that explicitly categorizing intermediate reasoning failures into calculation, logical, or linguistic errors allows models to effectively learn "what not to do" through targeted self-rethinking.

By moving away from binary metrics and embracing fine-grained failure decomposition, the AI research community is shifting from black-box evaluation to precise, actionable diagnostics. This systematic mapping of model weaknesses provides the foundational transparency required to build the next generation of reliable, self-correcting intelligent systems.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hallucination_taxonomy_a4c5d134,
  title = {hallucination taxonomy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hallucination-taxonomy/rdr_a4c5d134e9124de9a9053387776095af},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_hallucination_taxonomy_a4c5d134_v1,
  title = {hallucination taxonomy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hallucination-taxonomy/rdr_a4c5d134e9124de9a9053387776095af/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake