TLDR: A new research paper investigates the intrinsic causes of hallucination in Gemma Large Language Models (LLMs), identifying five key “symbolic properties”—modifiers, named entities, numbers, negation, and exceptions—that consistently trigger factual inaccuracies. The study found that hallucination rates remain significantly high across different Gemma model sizes, with modifiers and named entities being the most problematic. It also revealed that open-ended question-answering tasks expose more vulnerabilities compared to constrained formats, and that hallucination rates vary non-linearly with input length, peaking in short-to-medium contexts. The findings suggest that these symbolic vulnerabilities are fundamental architectural issues, not merely a matter of model scale, highlighting a need for deeper representational stability in LLMs.
Large Language Models (LLMs) have revolutionized many areas of natural language processing, from answering complex questions to summarizing texts and engaging in dialogue. However, a significant challenge that continues to plague these advanced AI systems is ‘hallucination’ – the confident generation of factually incorrect or nonsensical information. While much research has focused on detecting and reducing these errors, a recent study delves deeper, aiming to uncover the fundamental, intrinsic causes of hallucination within LLMs. This research, titled Investigating Symbolic Triggers of Hallucination in Gemma Models Across HaluEval and TruthfulQA, was conducted by Naveen Lamba, Sanju Tiwari, and Manas Gaur.
The core of this study identifies and characterizes specific “symbolic properties” that consistently trigger hallucinations in LLMs. These aren’t just surface-level errors but point to deeper vulnerabilities in how models process information. The researchers focused on five key symbolic triggers: modifiers (like adjectives and adverbs), named entities (such as names of people or places), numbers, negation (words like “not” or “never”), and exceptions (edge cases or conditional rules).
To test their hypothesis, the team utilized two well-known datasets, HaluEval and TruthfulQA, which are designed to evaluate LLM hallucinations. They ingeniously converted the standard question-answering format of these datasets into various other formats, including multiple-choice questions (MCQ) and Odd-One-Out (OOO) tasks. This allowed them to isolate whether these symbolic properties were indeed the root cause of hallucinations, regardless of the task structure or domain.
Their findings were quite revealing. Across different sizes of Gemma models (Gemma-2-2B, Gemma-2-9B, and Gemma-2-27B), the hallucination percentages linked to symbolic properties remained remarkably high. For the smallest model, Gemma-2-2B, the average hallucination rate was 79.0%. While this rate did decrease with larger models – dropping to 73.6% for Gemma-2-9B and 63.9% for Gemma-2-27B – a substantial amount of hallucination persisted. This indicates that simply scaling up model size doesn’t fully resolve these fundamental weaknesses.
Specifically, modifiers and named entities proved to be the most problematic symbolic triggers. Hallucination rates for modifiers ranged from 84.76% to 94.98% across all Gemma models and datasets, while named entities showed similarly high rates, from 83.87% to 93.96%. This suggests that these linguistic elements consistently confuse the models, highlighting a core weakness in their processing of such inputs, irrespective of their scale.
The study also explored how task format influences hallucination. The traditional Question-Answering (QA) format consistently produced the highest hallucination rates compared to MCQ and OOO formats. This implies that when models are given more freedom to generate open-ended responses, their underlying symbolic processing vulnerabilities are more exposed. Constrained formats, like multiple-choice, seem to offer some protection by limiting the scope for fabrication.
Furthermore, the researchers investigated the relationship between hallucination rates and the length of the input question. They found that hallucinations induced by modifiers and named entities were particularly high in short-to-medium contexts (10-30 tokens). While rates tended to decline for longer queries, suggesting enhanced contextual grounding, the trend wasn’t uniform across all properties. This non-monotonic behavior indicates that context length interacts uniquely with each symbolic property, pointing to localized representational instabilities rather than just a general issue with input size.
Also Read:
- Understanding How AI Connects Words to Images
- Understanding Large Language Models in Legal AI: A Deep Dive into Current Trends and Future Paths
In conclusion, this research provides compelling evidence that symbolic linguistic properties are consistent triggers for hallucination in Gemma models. The persistence of these vulnerabilities, even in larger models, suggests that they are not merely a function of model capacity but stem from how LLMs internally encode and generalize over symbolic constructs. This deeper understanding is crucial for developing more robust and trustworthy AI systems in the future, moving beyond superficial fixes to address the fundamental architectural limitations of LLMs.
