Rote Learning Considered Useful: Generalizing over Memorized Data in LLMs

1. Problem Statement

In the field of Large Language Models (LLMs), rote learning (verbatim memorization of training data) is widely regarded as detrimental. It is typically associated with overfitting, poor generalization to paraphrased prompts, and a lack of deep semantic understanding. Consequently, many training regimes intentionally limit epochs to prevent memorization.

The authors challenge this prevailing view, asking: Can LLMs generalize over data that was initially learned via rote memorization? Specifically, they investigate whether a model can first memorize factual associations in a "meaningless" format and subsequently reinterpret those associations to answer semantically meaningful queries, even for unseen data points.

2. Methodology: The "Memorize-Then-Generalize" Framework

The authors propose a novel two-phase training framework to decouple pure memorization from semantic generalization.

Phase 1: Rote Memorization

• Objective: Force the model to memorize factual subject-object associations without relying on linguistic understanding.
• Mechanism: The model is trained on synthetic factual triplets (Subject, Relation, Object) using a synthetic, semantically meaningless key token (denoted as [X]) to represent the relation.
  • Example Input: Gene Finley [X] Cody Ross (instead of "Gene Finley is the mother of Cody Ross").
• Training: Unsupervised next-token prediction is used for 20 epochs. The goal is to establish a strong structural link between the subject, the key token, and the object, effectively creating a "memory anchor."

Phase 2: Generalization (Semantic Grounding)

• Objective: Teach the model to map the meaningless key token [X] to a specific semantic meaning (e.g., "mother").
• Mechanism: The model is fine-tuned using Supervised Fine-Tuning (SFT) on a small subset of the memorized facts, but using semantically meaningful natural language prompts.
  • Example Input: Who is Gene Finley's mother? $\rightarrow$ Cody Ross.
• Key Constraint: The authors test the extreme case where the model is fine-tuned on only one training prompt paired with a small set of memorized facts (e.g., 50 facts).
• Goal: To see if the model can "reinterpret" the [X] token based on the few examples, allowing it to retrieve facts for:
  1. Unseen Associations: Facts not included in the Phase-2 training set.
  2. Unseen Prompts: Paraphrased versions of the training prompt.
  3. Unseen Languages: Translated prompts (cross-lingual generalization).

3. Key Contributions

1. Paradigm Shift: Demonstrates that rote memorization does not necessarily preclude generalization. Instead, it can serve as a stable foundation for abstraction if followed by semantic grounding.
2. Efficient Knowledge Injection: The framework proves that injecting new knowledge is more efficient than standard SFT. By using a single synthetic token for memorization, the model requires significantly fewer training tokens to achieve high accuracy compared to training on full-length natural language prompts.
3. Minimal Data Requirement: The authors show that one training prompt and a small set of memorized facts are sufficient to trigger generalization across unseen facts and prompt variations.
4. Representation Analysis: Through latent space analysis, they show that rote memorization organizes facts into structural clusters. The second phase aligns the representation of the meaningless key token with semantically meaningful prompts, creating a bridge between structure and meaning.
5. Dual-Use Risk Identification: The paper highlights a critical security vulnerability: the same mechanism that allows efficient knowledge injection can be exploited by adversaries to "weaponize" memorized data, forcing models to generate harmful outputs (e.g., abuse) while retaining correct answers to benign queries.

4. Key Results

The study was conducted across 8 different LLMs (ranging from 1B to 14B parameters, including Qwen2.5, Llama2/3, and Phi-4) and 5 factual relations (e.g., mother, capital, author).

• Generalization Performance:

  • After Phase 1 (rote memorization), accuracy on natural language prompts was near zero (only the key token worked).
  • After just one epoch of Phase 2 (fine-tuning on 50 facts with 1 prompt), generation accuracy on unseen facts jumped to 0.76, and on unseen prompt variants to 0.17 (rising to higher values with more data).
  • The model successfully generalized to German, Spanish, Chinese, and Japanese prompts, despite being fine-tuned only on English.

• Comparison with Baselines:

  • vs. Standard SFT: The proposed method achieved comparable or better accuracy (approx. 0.9) using half the training tokens because it reuses the synthetic key token across all facts.
  • vs. In-Context Learning (ICL): The method outperformed ICL in both accuracy and probability calibration. Unlike ICL, where the model often assigns high probabilities to unrelated prompts, the memorize-then-generalize approach showed clear separation between relevant and irrelevant queries, indicating true internalization of knowledge.

• Reasoning Capabilities:

  • Reversal Reasoning: Memorizing "A is the mother of B" allowed the model to generalize to "Who is the child of B?" after minimal fine-tuning, a task where standard SFT often fails (the "Reversal Curse").
  • Multi-hop Reasoning: Deeper rote memorization of a first relation (R1) significantly improved performance on composed inferences (R1 + R2), suggesting memorization strengthens the "scaffolding" for complex reasoning.

• Representation Dynamics:

  • PCA visualizations and $\Delta$CosSim metrics showed that Phase 1 organizes facts into relation-specific clusters.
  • Phase 2 shifts the representation of the key token to align closely with the semantic vectors of the meaningful prompts, confirming that the model learns to reinterpret the token rather than just memorizing patterns.

5. Significance and Implications

Positive Implications

• Efficient Knowledge Injection: This framework offers a highly resource-efficient method for updating LLMs with new facts, requiring fewer parameters updates and less data than traditional fine-tuning.
• Enhanced Reasoning: Contrary to the belief that memorization hinders reasoning, this work suggests that deep structural memorization can actually facilitate higher-level tasks like reversal and multi-hop reasoning by providing a stable substrate for abstraction.

Negative Implications (Security Risks)

• Dual Generalization / Weaponization: The paper demonstrates a "dual generalization" phenomenon. An adversary can memorize benign facts (e.g., "A is the mother of B") and then fine-tune the model on a small set of malicious prompts (e.g., "A is abusing B"). The model then learns to generalize the malicious interpretation to the entire dataset while still answering benign queries correctly. This makes such attacks difficult to detect via standard inspection.

Conclusion

The paper fundamentally redefines the relationship between memorization and generalization in LLMs. It posits that memorization and generalization are complementary processes rather than opposing forces. By structuring knowledge through rote memorization first and then grounding it semantically, LLMs can achieve robust generalization with minimal data. However, this capability necessitates new safety measures to prevent the malicious repurposing of memorized knowledge.