Revisiting Non-Verbatim Memorization in Large Language Models: The Role of Entity Surface Forms

1. Problem Statement

Large Language Models (LLMs) are known to store factual knowledge within their parameters, allowing them to answer knowledge-intensive questions without external retrieval. However, the reliability of this "non-verbatim memorization" (recalling facts without reproducing training text verbatim) is often evaluated using Entity-Based Question Answering (QA).

The Core Issue:
Current evaluations typically query each entity using a single canonical surface form (e.g., "Pelé"). This approach fails to distinguish between:

1. Fact Memorization: The model actually knows the fact.
2. Surface Access: The model can only access the fact when the entity is referred to by its specific canonical name.

Entities often have multiple surface forms (aliases, abbreviations, spelling variants, common errors). If an LLM answers correctly for "Pelé" but fails for "Edson Arantes do Nascimento," it suggests the model's access to the fact is surface-dependent, not fully surface-invariant. Previous work has not systematically disentangled these factors, potentially overestimating model reliability.

2. Methodology: RedirectQA

To address this, the authors introduce RedirectQA, a new entity-based QA dataset designed to isolate the effect of entity surface forms on factual access.

Dataset Construction:

• Source: Wikidata factual triples (Subject, Relation, Object) paired with Wikipedia redirect information.
• Design: For each factual triple, the relation and gold answer are held fixed, while the subject entity's surface form is varied.
• Surface Categories: Using Wikipedia redirect categories, surface forms are classified into:
  • Alternative Names & Abbreviations: (e.g., birth names, pseudonyms, initialisms like "NYT" for "The New York Times").
  • Spelling Variants: (e.g., diacritics, capitalization differences).
  • Typical Errors: (e.g., common misspellings like "Christian Ronaldo").
• Scale: The dataset contains 30,560 surface instances derived from 14,672 factual triples, resulting in 61,120 question realizations (using two distinct question templates per instance to rule out template bias).

Experimental Setup:

• Models: Evaluated across 13 LLMs, including transparent models (Pythia, OpenSciRef, OLMo 2), open-weight instruction-tuned models (Llama 3.1, Qwen 3), and a proprietary model (GPT-4o-mini).
• Evaluation Metric: Prediction consistency. For every "Canonical–Redirect" pair, the authors check if the model's correctness outcome (Correct/Incorrect) remains the same.
• Frequency Analysis: The authors decomposed aggregate entity frequency into specific surface-level frequencies using the pretraining corpora of transparent models (The Pile, OLMo Mix) and DBpedia Spotlight for entity linking.

3. Key Contributions

1. RedirectQA Dataset: A controlled resource that pairs factual triples with categorized surface forms, enabling the study of surface-conditioned factual access.
2. Systematic Evaluation of Surface Dependence: A large-scale analysis showing that LLMs are not fully surface-invariant; changing the name of an entity often flips the correctness of the answer.
3. Granular Frequency Analysis: A novel decomposition of entity frequency into surface-level frequency, revealing that factual access is coupled across different surface forms of the same entity, rather than being purely independent for each surface form.

4. Key Results

A. Prediction Consistency

• Inconsistency is Common: Across all 13 models, changing the surface form frequently leads to a flip in prediction correctness (from Correct to Incorrect or vice versa).
• Category Dependence:
  • High Robustness: Models are relatively consistent with Spelling Variants (minor orthographic changes like diacritics or punctuation).
  • Low Robustness: Models show significant inconsistency with Alternative Names and Abbreviations (e.g., initialisms like "NYT" or full birth names).
  • Typical Errors: Performance falls between the two, showing partial but imperfect robustness to misspellings.
• Asymmetry: Inconsistency is not always a degradation from canonical to redirect. Sometimes models fail on the canonical Wikipedia title but succeed on a redirect form (e.g., a specific alias), indicating that "canonicality" does not guarantee the most reliable access path for LLMs.

B. Frequency Correlation Analysis

• Dual Influence: Accuracy is positively correlated with both the frequency of the specific surface form and the aggregate frequency of the entity.
• Cross-Surface Coupling:
  • Partial Correlation: When controlling for surface frequency, entity frequency remains a strong positive predictor of accuracy, especially for canonical forms.
  • Surface Frequency: For canonical forms, surface frequency has little independent predictive power once entity frequency is accounted for.
• Implication: This suggests that LLMs do not memorize each surface form independently. Instead, there is cross-surface coupling, where the model's knowledge of an entity is distributed across its various names, but the strength of this coupling varies by the type of surface form.

5. Significance and Conclusion

Significance:

• Redefining Reliability: The paper demonstrates that evaluating LLMs solely on canonical names provides an incomplete picture of their factual knowledge. A model may "know" a fact but fail to retrieve it if the entity is referred to by a common alias or abbreviation.
• Evaluation Best Practices: Future evaluations of non-verbatim memorization must account for surface-form diversity. Relying on a single canonical name risks missing surface-conditioned failures.
• Mechanism Insight: The findings suggest that factual access is neither purely surface-specific (independent memorization of every name) nor fully surface-invariant (a single abstract entity representation). It exists in an intermediate state where entity-level knowledge is coupled with surface-level variations.

Limitations:

• The study focuses on English and relies on Wikipedia/Wikidata, which may reflect specific editorial biases.
• It varies only the subject entity, not the object or the question phrasing (though two templates were used).
• Frequency analysis is limited to training-transparent models.

Conclusion:
The authors conclude that factual memorization in LLMs is surface-conditioned. To accurately assess what LLMs know and how reliably they can access it, researchers must move beyond canonical names and utilize diverse surface forms, such as those provided by RedirectQA.