A Geometric Taxonomy of Hallucinations in LLMs
#hallucination #large language models #embedding space #taxonomy #unfaithfulness #confabulation #factual error #asymmetry #benchmark evaluation
📌 Key Takeaways
- Three hallucination types defined: (1) unfaithfulness—failure to engage with supplied context; (2) confabulation—generation of semantically foreign content; (3) factual error—incorrect claims within correct conceptual frames.
- Hallucinations are characterized by distinct geometric signatures in embedding space.
- The authors observe a striking asymmetry of hallucination frequency on standard benchmarks.
- The taxonomy is proposed to improve clarity and facilitate targeted mitigation strategies.
📖 Full Retelling
Researchers have published a new study on arXiv (ID 2602.13224v1) that introduces a geometric taxonomy for hallucinations in large language models. The paper identifies three distinct types—unfaithfulness, confabulation, and factual error—based on their signatures in embedding space. By mapping these phenomena, the authors aim to clarify why standard benchmarks exhibit a striking asymmetry in hallucination prevalence and provide a framework for more precise evaluation of language model behavior.
🏷️ Themes
Hallucinations in large language models, Embedding space geometry, Taxonomy development, Benchmark analysis, Model evaluation
Entity Intersection Graph
No entity connections available yet for this article.
Original Source
arXiv:2602.13224v1 Announce Type: new
Abstract: The term "hallucination" in large language models conflates distinct phenomena with different geometric signatures in embedding space. We propose a taxonomy identifying three types: unfaithfulness (failure to engage with provided context), confabulation (invention of semantically foreign content), and factual error (incorrect claims within correct conceptual frames). We observe a striking asymmetry. On standard benchmarks where hallucinations are
Read full article at source