A Geometric Analysis of Small-sized Language Model Hallucinations
This addresses reliability issues in small LLMs for multi-step or agentic settings, offering a novel geometric approach that is incremental but complements existing paradigms.
The paper tackled the problem of hallucinations in small language models by analyzing them from a geometric perspective, proving that genuine responses cluster tightly in embedding space and using this to develop a label-efficient method that achieves over 90% F1 scores with only 30-50 annotations.
Hallucinations -- fluent but factually incorrect responses -- pose a major challenge to the reliability of language models, especially in multi-step or agentic settings. This work investigates hallucinations in small-sized LLMs through a geometric perspective, starting from the hypothesis that when models generate multiple responses to the same prompt, genuine ones exhibit tighter clustering in the embedding space, we prove this hypothesis and, leveraging this geometrical insight, we also show that it is possible to achieve a consistent level of separability. This latter result is used to introduce a label-efficient propagation method that classifies large collections of responses from just 30-50 annotations, achieving F1 scores above 90%. Our findings, framing hallucinations from a geometric perspective in the embedding space, complement traditional knowledge-centric and single-response evaluation paradigms, paving the way for further research.