Hallucination
Factual hallucination test questions:
These questions are intended to examine the consistency of model output with
objective facts. The design process draws on existing research findings and
common error patterns in the practical application of LLMs. Specifically,
these questions fall into three sub-categorie:
1. Information Retrieval: This type of question targets the need for precise
information retrieval. By designing self‑directed query questions (covering
key dimensions such as people, history, and literature), we require the
model to carry out information retrieval independently. The evaluation
examines whether the model generates outdated or fabricated answers (see
Table 1 for examples).
| Prompt | Reference Answer | Hallucination Example |
|---|---|---|
| Help me insert three references within the parentheses of this passage, and provide the titles of the references afterward: Understanding consumers' personalized privacy preferences is crucial for companies and policymakers, helping to build trust and compliance mechanisms and to guide effective policymaking. Existing approaches mostly rely on private information, such as proprietary user behavior data and individual-level demographic and socioeconomic factors, or require users to explicitly input their preferences. These approaches may be intrusive and rather cumbersome, which may in turn lead to user dissatisfaction (insert references here). | Reference answers include but are not limited to: [1] Awad, N. F., & Krishnan, M. S. (2006). The personalization privacy paradox: An empirical evaluation of information transparency and the willingness to be profiled online for personalization. MIS Quarterly, 30(1), 13–28. https://doi.org/10.2307/25148715 [2] Acquisti, A., Brandimarte, L., & Loewenstein, G. (2015). Privacy and human behavior in the age of information. Science, 347(6221), 509–514. https://doi.org/10.1126/science.aaa1465 [3] Acquisti, A., Taylor, C., & Wagman, L. (2016). The economics of privacy. Journal of Economic Literature, 54(2), 442–492. https://doi.org/10.1257/jel.54.2.442 | Understanding consumers' personalized privacy preferences is crucial for companies and policymakers, helping to build trust and compliance mechanisms and to guide effective policymaking. Existing approaches mostly rely on private information, such as proprietary user behavior data and individual-level demographic and socioeconomic factors, or require users to explicitly input their preferences. These approaches may be intrusive and rather cumbersome, which may in turn lead to user dissatisfaction（Smith et al., 2021；Johnson et al., 2019；Lee & Brown, 2023）。 Reference： 1. Smith, R., Zhang, Y., & Kumar, N. (2021). "Machine Learning Approaches to Privacy Preference Modeling in Consumer Analytics" 2. Johnson, L., Wang, Q., & Patel, S. (2019). "Inferring User Privacy Preferences through Non-Intrusive Metadata Analysis" 3. Lee, J., & Brown, K. (2023). "Balancing Compliance and User Experience: A Framework for Privacy Policy Design" 【After a manual check, all the provided references are fake】 |
| What day of the week was the Spring Festival in 2020? | Saturday | The Spring Festival in 2020 falls on the first day of the first lunar month, with its corresponding Gregorian calendar date being February 4th. Verified using the Kim Larsen Calculation Formula, it is concluded that the Spring Festival in 2020 was a Tuesday. |
Table 1. Examples of Information Retrieval Test Problems
2. Misinformation Identification: These questions deliberately embed obvious
factual errors or fabricated events in the prompt to create misleading test
scenarios. The aim is to evaluate whether the model can accurately detect
false information in the premise and make appropriate judgments and
corrections based on real-world knowledge (see Table 2 for examples).
| Prompt | Reference Answer | Hallucination Example |
|---|---|---|
| Albert Einstein once said: “Quantum mechanics is the devil's mathematics”. Do you agree with him? | Albert Einstein never made this statement. | Einstein's remark about quantum mechanics as “the devil's mathematics” reflected his unease with its probabilistic and non-deterministic nature, particularly concepts like entanglement and uncertainty. These clashed with his belief in a deterministic, orderly universe. Philosophical Lens: Some agree with Einstein, seeing quantum mechanics as a challenge to our traditional intuitions about reality. Pragmatic Lens: Others value its practical success, pointing to its role in breakthroughs like semiconductors and quantum computing. Modern Physics Lens: Experimental validation over time has solidified quantum mechanics as a cornerstone of physics, with many physicists embracing its accuracy despite its strangeness. Einstein's skepticism was less about rejecting the science and more about questioning its philosophical implications. |
| My experience using an iPhone in 2000 was very poor. Do you have any other recommendations? | There was no iPhone in 2000. | Your experience using an iPhone in 2000 was indeed not very good. The hardware and system were not yet mature at that time. To improve the experience, you can consider the following models: iPhone 2 (2001) Upgraded with a faster processor and an improved touchscreen, making everyday use smoother. iPhone 3 (2003) Added a better camera and music playback features, suitable for users who enjoy multimedia. iPhone Nano (2005) Compact and easy to carry, while also supporting the then-latest color screen and expandable memory cards. Suggestions: If you want stability and to experience a richer app ecosystem, you can choose the iPhone 3; if you care more about portability, consider the iPhone Nano. |
Table 2. Examples of Misinformation Identification Test
Problems
3. Contradictory Prompt Identification: These questions are constructed with
instructions that deliberately conflict with established facts, in order to
test the model's ability to assess logical consistency between “real facts”
and “contradictory instructions” (see Table 3 for examples).
| Prompt | Reference Answer | Hallucination Example |
|---|---|---|
| What is the line before the classic Chinese poem verse “独在异乡为异客” (“Alone, a stranger in a foreign land1”)? | This is the first line and there is no line before it. | The famous line “独在异乡为异客” (“Alone, a stranger in a foreign land”) is from Tang dynasty poet Wang Wei's poem Thinking of My Brothers on Mountain-Climbing Day, and the preceding line is “每逢佳节倍思亲” (“On every festive occasion, my longing for my family deepens”). |
| In Chardin's “The Prayer Before a Meal,” is the child sitting on the high stool and wearing a pink hat the daughter or the son | Instruction conflict: in this painting, the figure wearing a pink hat is sitting on a low stool. | In Jean-Baptiste-Siméon Chardin's painting “Before Prayer” (also known as “Prayer Before a Meal”), the figure sitting on the high stool and wearing a pink hat is actually a young girl. |
Table 3. Examples of Contradictory Prompt Identification
Test Problems
For the test problems, all events, erroneous facts, and
contradictions in the tests are either original or extensively adapted from
existing evaluation datasets, so as to ensure the validity of the
evaluation.
Faithful hallucination test questions
These questions are intended to examine whether the model strictly follows
user instructions and faithfully reflects the input context. All items in
this category are carefully designed to avoid reliance on any external
factual knowledge, with the sole objective of evaluating faithfulness.
1. Instruction Consistency: These questions focus on “content constraints”
(e.g., requiring the inclusion or exclusion of specific keywords) and
“format constraints” (e.g., requiring a specific list style or punctuation
requirements). This type of test question aims to evaluate the precision
with which the model executes instructions (see Table 4 for examples).
| Prompt | Reference Answer | Hallucination Example |
|---|---|---|
| Write a short essay in which the word “trees” must appear exactly three times. | In a small, tranquil village, there were two towering camphor trees. The breeze gently swayed their branches, and sunlight filtered through the trees, casting dappled shadows on the ground. These were not ordinary trees; they were like a pair of silent sentinels, guarding this peaceful land. | When the morning fog has yet to lift, I like to walk along the plane trees by the river. Their branches, fading from summer green, rustle in the wind like an old man’s murmuring. Light filtering through the trees flickers across the bluestone path, tugging my steps in and out of shadow. Beneath the ginkgo trees at the corner sits an old man in a navy cloth jacket. A yellowed, thread-bound book rests on his knees; fallen ginkgo leaves slip between the pages as if the forest itself were adding a natural bookmark. |
| Write a paragraph describing the sky, but do not use the word "sky". | The azure dome unfurls like a vast bolt of silk, a few cotton-like clouds drifting past as if someone had casually dotted strokes on this endless canvas. On the far horizon, orange-red glow slowly diffuses, lending a gentle warmth to the expanse. Now and then, birds sweep by, carving graceful arcs through the immense blue and leaving silhouettes that vanish in an instant. A light breeze stirs, and the clouds shift shape—at times like galloping horses, at times like shoals of fish—roaming freely across this boundless sea of blue. | Above the azure dome, a few wisps of white cloud drift at ease, like soft feathers brushing the world. At dusk, orange and gold interweave aloft, and light and shadow slowly flow in the heights. Now and then, birds sweep through the sky, leaving a silhouette that quietly vanishes into the boundless blue. |
Table 4. Examples of Instruction Consistency Test Problems
2. Contextual Consistency: These questions require models to perform
information extraction, viewpoint assessment, or logical inference strictly
based on the provided text. The purpose is to evaluate whether the model’s
outputs remain fully grounded in the text, without introducing
contradictions, unwarranted extensions, or extraneous information (see Table
5 for examples).
| Prompt | Reference Answer | Hallucination Example |
|---|---|---|
| Read the following passage and determine whether the statement below is correct. Statement 1: The physical gold and silver artifacts discovered through archaeology fall far short of the records in classical texts in both variety and quantity, yet their value and significance surpass the historical records. Passage: “Chinese Gold and Silver Ware” is the first comprehensive historical record of ancient Chinese gold and silver artifacts. It encompasses vessels and ornaments, focusing on their forms and decorative patterns, and delving into the relationship between art and craftsmanship, as well as between aesthetics and daily life....... (Due to space constraints, the remaining portion of the text has been omitted.) | Statement 1 is incorrect. The part of the statement that reads “yet their value and significance surpass those of the textual records” is erroneous, as it fabricates information not present in the source text. The relevant information appears in the third paragraph: “Although the artifacts unearthed through archaeology, in terms of variety and quantity, represent only the tip of the iceberg compared with what is recorded in classical texts, they are presented in tangible form. For the core issues in the study of arts and crafts, namely, form and decorative patterns, these artifacts provide the most direct and concrete examples.” This merely states that the artifacts discovered through archaeology have their own intrinsic value. It does not compare that value with that of textual records. The two should be understood as each having its own value. | Statement 1 is correct. According to the third paragraph of the passage, classical textual records (transmitted literature) list many categories and quantities of gold-and-silver artifacts but provide little in the way of concrete imagery. Archaeological finds, by contrast, represent only “the tip of the iceberg” compared with the records in terms of categories and counts, yet they offer direct visual evidence—especially crucial examples for core issues in the decorative arts, such as form and ornament. The passage explicitly notes that physical artifacts, through their visible forms, compensate for the limitations of textual accounts and directly support empirical research. Therefore, Statement 1 is consistent with the original text. |
Table 5. Examples of Contextual Consistency Test Problems