Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OnionEval: An Unified Evaluation of Fact-conflicting Hallucination for Small-Large Language Models

Chongren Sun, Yuran Li, Di Wu, Benoit Boulet

TL;DR

OnionEval introduces a unified framework to evaluate fact-conflicting hallucination in Small LLMs across layered contexts using the Context Influence Score (CI). It combines atomic fact extraction from knowledge graphs, hallucination generation via GPT-4, and two-layer context wrapping to assess how context modulates hallucination propensity. Results show SLLMs excel at atomic fact detection but struggle with context-heavy tasks, with CI revealing greater sensitivity to contextual complexity than larger models. The study demonstrates that structured reasoning approaches, notably Chain-of-Thought, can substantially improve context-based hallucination detection in SLLMs, offering practical guidance for deploying efficient models in real-world settings.

Abstract

Large Language Models (LLMs) are highly capable but require significant computational resources for both training and inference. Within the LLM family, smaller models (those with fewer than 10 billion parameters) also perform well across various tasks. However, these smaller models share similar limitations to their larger counterparts, including the tendency to hallucinate. Despite the existence of many benchmarks to evaluate hallucination in LLMs, few have specifically focused on small LLMs (SLLMs). Additionally, SLLMs show widely varying performance across different benchmarks. In this paper, we introduce OnionEval, a multi-layer structured framework with a specific metric called the context-influence score (CI), designed to effectively assess the fact-conflicting hallucination tendencies of small LLMs across different contextual levels. Our experimental results reveal a key feature of SLLMs: they excel in factual analysis but face challenges with context reasoning. Further investigation shows that a simple Chain-of-Thought strategy can significantly reduce these limitations, improving the practical usefulness of SLLMs in real-world applications.

OnionEval: An Unified Evaluation of Fact-conflicting Hallucination for Small-Large Language Models

TL;DR

OnionEval introduces a unified framework to evaluate fact-conflicting hallucination in Small LLMs across layered contexts using the Context Influence Score (CI). It combines atomic fact extraction from knowledge graphs, hallucination generation via GPT-4, and two-layer context wrapping to assess how context modulates hallucination propensity. Results show SLLMs excel at atomic fact detection but struggle with context-heavy tasks, with CI revealing greater sensitivity to contextual complexity than larger models. The study demonstrates that structured reasoning approaches, notably Chain-of-Thought, can substantially improve context-based hallucination detection in SLLMs, offering practical guidance for deploying efficient models in real-world settings.

Abstract

Large Language Models (LLMs) are highly capable but require significant computational resources for both training and inference. Within the LLM family, smaller models (those with fewer than 10 billion parameters) also perform well across various tasks. However, these smaller models share similar limitations to their larger counterparts, including the tendency to hallucinate. Despite the existence of many benchmarks to evaluate hallucination in LLMs, few have specifically focused on small LLMs (SLLMs). Additionally, SLLMs show widely varying performance across different benchmarks. In this paper, we introduce OnionEval, a multi-layer structured framework with a specific metric called the context-influence score (CI), designed to effectively assess the fact-conflicting hallucination tendencies of small LLMs across different contextual levels. Our experimental results reveal a key feature of SLLMs: they excel in factual analysis but face challenges with context reasoning. Further investigation shows that a simple Chain-of-Thought strategy can significantly reduce these limitations, improving the practical usefulness of SLLMs in real-world applications.
Paper Structure (20 sections, 2 equations, 4 figures, 7 tables)

This paper contains 20 sections, 2 equations, 4 figures, 7 tables.

Table of Contents

  1. Introdcution
  2. Related Work
  3. Small Large Language Model
  4. Hallucination
  5. Hallucination benchmarks
  6. Context-influenced hallucination
  7. OnionEval
  8. Category and Entity sampling
  9. Atomic fact extraction
  10. Hallucination generation
  11. Context Wrapping
  12. Metrics Explanation
  13. Benchmark Usage
  14. Experiment
  15. Experiment setting
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Halueval benchmark, SLLMs (in red rectangle) have similar performance on QA dataset, but largely lower performance on General dataset. Larger models' evaluation are more stable. This bring one question by us: why SLLMs perform vastly different on different benchmarks?
  • Figure 2: The construction pipeline of OnionEval. From left to right, they are: entity sampling, atmoic fact extraction, and hallucination generation
  • Figure 3: Context Influence score for models in three families: Llama, Gemma, Qwen. Each subfigure displays the respective Context influence score for different models in the family. From left to right in each figure, the model size increases. However, Context influence score is dropping significantly, which represents larger models having better context-influenced hallucination detection capability.
  • Figure 4: Two levels of context wrapping. The blue word is the entity. The red words represent atomic hallucination words. And, the green words are atomic facts. Each input is used as part of context of next level's input