Editorial illustration for Tiny GPT Memorizes Facts But Fails to Connect Them
Small GPT Models Fail at Connecting Facts: Two-Hop Curse
Tiny GPT Memorizes Facts But Fails to Connect Them
A small GPT model trained on invented facts got every individual fact right, 100 percent, and then failed almost every question that required linking two of them. That gap is the subject of a recent paper by Balesni, Korbak and Evans, which calls the failure the Two-Hop Curse. The setup is simple enough to rebuild on a laptop: feed a model a fact like "X performed song Y," feed it another like "Y's performer married Z," then ask it to answer "Who did the performer of Y marry?" in one step, no scratchpad, no chain of thought.
When both facts lived in the same training document, the model could chain them. Split across two separate documents, accuracy dropped to chance, as if the facts existed in sealed rooms with no door between them. The author decided to test the paper's proposed fix firsthand, rebuilding the experiment at a scale small enough to inspect every weight and every training example.
What follows is a walkthrough of that setup: 200 invented subjects, 200 bridge entities connecting them, and a question nobody in the training data ever actually asked.
The model knows every fact it needs, and it can answer every one-hop question perfectly. It still can't put two of them together without being shown how.
Why this matters
This small experiment cuts against a comfortable assumption many of us make when evaluating model progress: that memorization is a step toward reasoning. Here, it isn't. A model that hits 100% training accuracy by step 100 and still scores 0 out of 200 on two-hop questions, across three separate runs, is telling us something specific about how these systems store knowledge versus how they use it.
The one-line fix from the referenced paper only partially closing that gap matters too. It suggests the "curse" isn't a simple bug with a known patch, it's a structural property of single-step inference that researchers haven't fully solved. For developers shipping products that assume a model can chain facts without visible reasoning, this is a warning to test that assumption directly rather than infer it from benchmark scores.
For researchers, it's a cleaner, smaller version of a problem usually buried in messier, larger models, worth building on rather than waving away.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Challenge-Aware In-Between Retrieval Outperforms Upfront on MEDAGENTSBENCH Hard
From Archives ChatGPT Now Natively Generates Animated GIFs and Edits Individual Frames
Common Questions Answered
What is the Two-Hop Curse that the paper by Balesni, Korbak and Evans describes?
The Two-Hop Curse is a phenomenon where a small GPT model can memorize individual facts with 100% accuracy but fails to answer questions requiring the connection of two related facts. For example, the model knows both "X performed song Y" and "Y's performer married Z" but cannot answer "Who did the performer of Y marry?" when both facts must be linked together.
Why does the tiny GPT model fail at two-hop reasoning despite achieving perfect one-hop accuracy?
The model demonstrates that memorization of individual facts does not automatically translate into the ability to reason by combining those facts. Although the model can answer every single one-hop question perfectly, it cannot synthesize two pieces of information together without explicit training on how to do so, revealing a fundamental gap between knowledge storage and knowledge application.
What does the Two-Hop Curse experiment reveal about the relationship between memorization and reasoning in language models?
The experiment challenges the assumption that memorization is a stepping stone toward reasoning capability. A model achieving 100% training accuracy while scoring 0 out of 200 on two-hop questions demonstrates that memorization alone does not lead to reasoning, suggesting these systems store knowledge differently than they use it for complex tasks.
How significant is the one-line fix mentioned in the paper for closing the gap between fact memorization and two-hop reasoning?
According to the article, the one-line fix from the referenced paper only partially closes the gap between memorization and two-hop reasoning, indicating that addressing this limitation requires more substantial changes than a simple code adjustment. This suggests the Two-Hop Curse represents a deeper architectural or training issue in how models process multi-step reasoning.
Further Reading
- Memorization, Emergence, and Explaining Reversal Failures: A Controlled Study of Relational Semantics in LLMs - arXiv
- Reason to Rote: Rethinking Memorization in Reasoning - arXiv
- Rote Learning Considered Useful: Generalizing over Memorized Data in LLMs - arXiv
- On the Structural Memory of LLM Agents - arXiv
- Rethinking How to Remember: Beyond Atomic Facts in Lifelong Learning - arXiv
