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

Chongren Sun, Yuran Li, Di Wu, Benoit BouletarXiv 2025 · DOI · arXiv

Executive Summary

OnionEval is a novel evaluation framework designed to investigate why small language models (SLLMs) hallucinate even when they possess the underlying factual knowledge. By wrapping atomic facts in layers of increasingly complex context—a process akin to peeling an onion in reverse—researchers discovered that while SLLMs rival massive models in raw factual knowledge, their performance collapses when those facts are embedded within a narrative.

The Challenge: Knowledge vs. Context

As AI deployment shifts toward edge devices and smartphones, Small Language Models (SLLMs, typically <10B parameters) have become indispensable. However, a significant discrepancy exists: an SLLM might excel in factual Q&A benchmarks but fail in general reasoning tasks.

The authors posit that existing benchmarks fail to distinguish between a knowledge gap (the model lacks the fact) and a contextual reasoning failure (the model knows the fact but is confused by surrounding text). OnionEval isolates this "context tax," specifically focusing on fact-conflicting hallucinations where the model must identify a false statement hidden in a story.

Methodology: The OnionEval Pipeline

The researchers developed a multi-layer evaluation strategy that follows a four-step pipeline:

  1. Entity Sampling: 515 entities were selected across 19 categories (e.g., Vehicles, Plants).
  2. Atomic Fact Extraction: Knowledge graphs were used to extract simple factual statements (e.g., "Tulips are generally red").
  3. Hallucination Generation: GPT-4 created "fact-conflicting" versions (e.g., "Tulips are generally black").
  4. Context Wrapping: The core of the framework involves embedding the hallucinated statement into layers of increasing complexity.
The OnionEval layered evaluation strategy, showing the progression from an atomic fact to complex narrative layers.

The layers are defined as:

  • Atomic Layer: The statement in isolation.
  • Layer 1: The statement embedded in a short story (e.g., a character reading a book containing the false claim).
  • Layer 2: The story further embedded in a specific setting (e.g., the character in a coffee shop).

To quantify this effect, the researchers introduced the Context Influence (CI) Score, which measures the accuracy drop as context complexity increases.

Key Insights

  • Fact-Rich but Context-Poor: SLLMs like Llama 3.2-3B are nearly as accurate as 70B+ models at identifying hallucinations in isolated "atomic" sentences. They possess the knowledge but lose access to it when "flavor text" is added.
  • Context Collapse: For many SLLMs, accuracy plummeted from ~90% at the atomic level to near 0% at Layer 1. This suggests that extracting a fact from a story is a significantly higher hurdle for small models than memorization.
  • RAG Limitations: Counter-intuitively, Retrieval-Augmented Generation (RAG) decreased performance in these tests. For SLLMs, the additional retrieved context acted as noise rather than assistance.
  • CoT as a Solution: Chain-of-Thought (CoT) prompting dramatically recovered performance, indicating that SLLMs have the latent capability to handle context if guided step-by-step.

Key Results

The evaluation revealed a stark divide between model scales. While large models (70B+) maintained stable performance across layers, SLLMs showed extreme sensitivity to context.

Context Influence (CI) scores across Llama, Gemma, and Qwen model families, illustrating the dramatic 'context tax' paid by smaller models.
  • Performance Parity: Llama 3.2-3B achieved 90.71% accuracy on atomic facts, nearly matching Llama 3.1-70B (91.46%).
  • The CI Score Gap: Large models maintained CI scores around 10, while SLLMs saw scores skyrocket to between 70 and 99, indicating extreme sensitivity to context.
  • CoT Success: With CoT, Llama 3.2-3B’s Layer 1 accuracy jumped from a dismal 23.92% to 74.94%.

Conclusion and Future Directions

OnionEval highlights that the primary weakness of SLLMs is not a lack of knowledge, but an inability to maintain that knowledge under contextual pressure. Future work will explore knowledge cutoffs, longer context windows (e.g., full books), and more effective mitigation strategies beyond simple CoT.

Summary updated

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.
BibTeX
@misc{sun2025onioneval,
  author = {Chongren Sun and Yuran Li and Di Wu and Benoit Boulet},
  title = {{OnionEval: An Unified Evaluation of Fact-conflicting Hallucination for Small-Large Language Models}},
  year = {2025},
  eprint = {2501.12975},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.12975},
}
Figure