Technical Summary: Hierarchical Memorization in Large Language Models

Problem Statement

Large Language Models (LLMs) exhibit a critical failure mode in academic citation generation: the fabrication of non-existent references. While LLMs demonstrate fluency across domains, they frequently hallucinate bibliographic records that appear structurally correct but are factually false. This study posits that hallucination and verbatim memorization are not distinct phenomena but opposite outcomes of the same probabilistic learning process. Specifically, the paper investigates how the redundancy of a bibliographic record in the pretraining corpus—proxied by its citation count—determines the model's ability to recall it accurately. The core problem is to characterize the internal structure of this memorization process: is it a binary state (memorized vs. hallucinated), or a graduated, hierarchical phenomenon where different components of a record are recalled at different levels of redundancy?

Methodology

The study employed a closed-system experimental design using GPT-4.1 (knowledge cutoff: June 2024) to generate bibliographic records without access to external retrieval tools.

• Experimental Design: The model was prompted to generate five recent academic papers across 20 distinct computer science domains (e.g., Vision Transformers, Diffusion Models, LoRA), resulting in 100 total samples.
• Data Verification: Generated records were manually verified against Google Scholar (as of October 2025) to determine existence and collect citation counts. Google Scholar was chosen for its broad coverage of preprints, conference proceedings, and web discussions, which better approximates the diversity of LLM pretraining corpora than traditional academic databases.
• Evaluation Metrics:
  • Manual Scoring: Records were classified as completely correct (Score 2), partially hallucinated (Score 1), or completely hallucinated (Score 0).
  • Computational Evaluation: For existing records, cosine similarity (using Sentence-BERT embeddings) was calculated between generated and authentic metadata to measure factual fidelity on a continuous scale.
  • Field-Level Analysis: For partially hallucinated records ($n=68$), the correctness of seven specific metadata fields (title, first author, co-authors, year, journal, volume, pages) was evaluated individually to map memorization trajectories.

Key Contributions

This study extends preliminary work by reframing the hallucination-memorization dichotomy as a graduated, hierarchically layered phenomenon. The primary contributions are:

1. Log-Linear Scaling: Establishing that factual accuracy scales log-linearly with citation count, serving as a proxy for training data redundancy.
2. Memorization Thresholds: Identifying two empirically distinct thresholds in the transition from probabilistic synthesis to verbatim recall.
3. Hierarchical Memorization: Demonstrating that within a single bibliographic record, different metadata fields cross the memorization threshold at vastly different citation levels.
4. Memory Interference: Characterizing a failure mode where semantically similar high-citation records are conflated, interpreted through the lens of spurious attractors in modern Hopfield networks.

Results

1. Citation Frequency and Factual Accuracy

A strong positive correlation ($r = 0.75, p < .001$) exists between log-transformed citation counts and cosine similarity. High-citation papers achieve significantly higher factual scores than low-citation ones (Cohen's $d = 1.02$). The relationship is not linear but exhibits threshold behavior:

• Inflection Point (~90 citations): The point where the model transitions from purely generative behavior (synthesizing tokens based on associations) to memory-based behavior (increasingly recalling specific records).
• Saturation Point (~1,248 citations): Beyond this threshold, bibliographic records are reproduced nearly verbatim (cosine similarity $\approx 1.0$) with minimal residual variance, indicating deterministic recall.

2. Hierarchical Memorization Patterns

Analysis of partially hallucinated records reveals that memorization is not uniform across metadata fields. The fields are recalled in a specific order based on their redundancy in the training corpus:

• Earliest Memorized: Titles and First Authors show steady accuracy increases starting from lower citation counts.
• Intermediate Memorization: Co-authors and Journal Names begin to be accurately reproduced only around the saturation threshold (~1,000 citations).
• Late/Unreliable Memorization: Volume and Page numbers remain noisy even at high citation counts.
• Non-Memorized: Publication Year shows essentially no correlation with citation count ($r = 0.038$). The model appears to generate years based on domain-level priors (estimating a plausible period based on when a field was active) rather than retrieving the specific year of the record.

3. Memory Interference

Even for highly cited papers, errors can occur due to the conflation of similar records. A notable case involved a generated citation for a GAN review paper that combined the author list of a seminal 2014 paper with the title phrasing of a 2020 survey. This suggests that when multiple high-frequency patterns overlap (similar titles/authors), their "attractors" in the model's energy landscape merge, creating a spurious attractor—a plausible but non-existent composite record.

Significance and Claims

The paper argues that memorization in LLMs is not a binary on/off state but a continuum shaped by the uneven distribution of knowledge in the pretraining corpus.

• Reframing Hallucination: Hallucination is not merely a lack of knowledge but a probabilistic reconstruction of low-redundancy data, while verbatim recall is a form of overfitting to high-redundancy data.
• Internal Structure: The study provides evidence that the "memory" of an LLM is internally structured hierarchically. A record can be "partially memorized," where the title and author are recalled verbatim, but the venue and year are hallucinated.
• Implications for Reliability: The findings suggest that in closed-system scenarios (without external retrieval), the reliability of LLM-generated structured data is fundamentally limited by the redundancy of the source material. Even highly cited records are susceptible to interference if they share semantic features with other high-frequency records.
• Theoretical Framework: The paper interprets these interference patterns through the framework of modern Hopfield networks, where the attention mechanism acts as an energy minimization process that can converge on spurious attractors formed by overlapping high-frequency patterns.

The study concludes that understanding these hierarchical thresholds and interference patterns is essential for characterizing the boundaries of LLM memorization and assessing the reliability of generated factual records in domains such as academic citation, legal precedents, and patent references.