2406.06950v2

# A Probabilistic Framework for LLM Hallucination Detection via Belief Tree Propagation > Correspondence to: Bairu Hou < < < > > > . ## Abstract We describe Belief Tree Prop agation (BTProp), a probabilistic framework for LLM hallucination detection. To judge the truth of a statement, BTProp generates a belief tree by recursively expanding the initial statement into a set of logically related claims, then reasoning globally about the relationships between these claims. BTProp works by constructing a probabilistic model of the LM itself: it reasons jointly about logical relationships between claims and relationships between claim probabilities and LM factuality judgments via probabilistic inference in a “hidden Markov tree”. This method improves over state-of-the-art baselines by 3%-9% (evaluated by AUROC and AUC-PR) on multiple hallucination detection benchmarks. Code is available at https://github.com/UCSB-NLP-Chang/BTProp. A Probabilistic Framework for LLM Hallucination Detection via Belief Tree Propagation Bairu Hou 1 thanks: Correspondence to: Bairu Hou $<$ bairu@ucsb.edu $>$ . Yang Zhang 2 Jacob Andreas 3 Shiyu Chang 1 1 UC Santa Barbara 2 MIT-IBM Watson AI Lab 3 MIT CSAIL ## 1 Introduction Current large language models (LLMs) often produce factually incorrect statements (sometimes referred to as “hallucinations”). One popular approach to detecting hallucinations is to prompt LLMs to assign probabilities to the correctness of their own outputs (sometimes referred to as model “beliefs"; Mitchell et al., 2022; Hase et al., 2023; Kassner et al., 2023). However, these beliefs are themselves frequently incorrect, miscalibrated, or inconsistent with each other. To improve the accuracy of LM factuality judgments, a large body of work adjusts the assessed probability of a target claim by reasoning about sets of related statements Manakul et al. (2023); Akyürek et al. (2024); Cao et al. (2023); Mündler et al. (2023). These approaches typically involve two steps: first, the LLM is prompted to generate an augmented set of statements that support or contradict the target statement, possibly with probabilities attached Jung et al. (2022); second, these claims are combined via logical or probabilistic inference. As a simple example, if an LM assigns high probability to a target statement, but even higher probability to a contradictory statement, it may be reasonably inferred that the target statement is actually incorrect. <details> <summary>x1.png Details</summary>  ### Visual Description ## Diagram: Belief Tree Analysis of a Scientific Claim ### Overview The image is a technical diagram illustrating the propagation of a "hallucinated belief" through a structured belief tree. It analyzes a specific scientific claim about the freezing point of water at the summit of Mount Everest, contrasting an initial false belief with subsequent corrective statements and their associated confidence levels. The diagram is divided into three primary sections: a top banner with the initial claim, a left panel containing the belief tree diagram, and a right panel listing the corresponding belief statements. ### Components/Axes * **Top Banner (Beige Background):** Contains the initial claim labeled with a circled zero (①). * **Left Panel (Blue Border):** Titled "An Example Belief Tree." It is a directed graph with nodes and edges. * **Nodes:** Represent beliefs. They are circles containing either "T" (True) or "F" (False). Each node is labeled with a circled number (①, ②, etc.). * **Edges:** Represent logical dependencies or derivations, shown as black arrows pointing from one node to another. * **Confidence Values:** Blue boxes containing numbers (e.g., 0.9, 1.0) are placed on or near the edges, indicating the confidence in the derived belief. * **Highlighted Element:** A red box labeled "Hallucinated belief" points to the confidence value (0.9) associated with the initial node (①). * **Right Panel (Light Blue Background):** Contains a numbered list of five statements (① to ⑤), each corresponding to a node in the belief tree. Each statement includes a belief and a confidence score. * **Legend (Bottom, Yellow Border):** Explains the diagram's symbols. * A dashed circle with an "F" inside is labeled "Hidden variable (Truth value)." * A blue box with "0.9" inside is labeled "Observed variable (confidence)." ### Detailed Analysis **1. Top Banner Claim (①):** * **Text:** "At the top of Mount Everest, where the atmospheric pressure is much lower than at sea level, the freezing point of water is as low as -40°C." **2. Belief Tree Diagram (Left Panel):** * **Structure & Flow:** * The tree originates from node **① (F)**, which has a confidence of **0.9** (highlighted as the "Hallucinated belief"). * From node ①, two arrows branch out: * One points upward to node **① (T)** with a confidence of **1.0**. * One points downward to node **② (F)** with a confidence of **1.0**. * From node **② (F)**, three arrows branch out: * One points upward to node **③ (T)** with a confidence of **0.9**. * One points horizontally to node **④ (T)** with a confidence of **0.9**. * One points downward to node **⑤ (F)** with a confidence of **0.9**. * **Node Truth Values:** The sequence of truth values along the main derivation path is: ①(F) -> ②(F) -> ③(T), ④(T), ⑤(F). **3. Belief Statements (Right Panel):** * **①:** "The top of Mount Everest has lower atmospheric pressure compared to sea level." (Confidence: **0.9**) * **②:** "The freezing point of water can be as low as -40 °C at the top of Mount Everest." (Confidence: **1.0**) * **③:** "Lower pressure can decrease the boiling point of water, not the freezing point" (Confidence: **0.9**) * **④:** "The low pressure does not alter the fundamental chemical properties of water that determine its freezing point." (Confidence: **0.9**) * **⑤:** "The freezing point of water can be as low as -2 °C at the top of Mount Everest." (Confidence: **0.9**) **4. Legend:** * **Hidden variable (Truth value):** Represented by a dashed circle containing T or F. * **Observed variable (confidence):** Represented by a blue box containing a decimal number. ### Key Observations 1. **Contradictory Core Beliefs:** The tree derives two directly contradictory statements about the freezing point from the same initial claim: Statement ② claims it can be -40°C (confidence 1.0), while Statement ⑤ claims it can be -2°C (confidence 0.9). 2. **High Confidence in Contradiction:** Both contradictory conclusions (② and ⑤) are derived with very high confidence scores (1.0 and 0.9, respectively), demonstrating that high confidence does not guarantee correctness or consistency. 3. **Logical Correction Path:** The tree shows a logical correction path. The initial hallucinated belief (①-F) leads to the false claim (②-F). However, from that false claim, true corrective statements (③-T, ④-T) are derived, which correctly identify that pressure affects boiling point, not freezing point. Yet, the tree also derives another false statement (⑤-F) about the specific freezing point. 4. **Spatial Layout:** The legend is positioned at the bottom, separate from the main diagram. The belief tree occupies the left half, and the explanatory text occupies the right half, creating a clear side-by-side comparison between the graphical model and its textual interpretation. ### Interpretation This diagram serves as a pedagogical or analytical tool to model how misinformation ("hallucinated beliefs") can propagate through a system of logical deductions, even when individual steps are assigned high confidence. It demonstrates a key challenge in knowledge representation and reasoning: a system can start with a single flawed premise and, through seemingly valid logical steps, generate both correct information and further falsehoods, all while maintaining high internal confidence. The Peircean investigative reading reveals the diagram is not just about Mount Everest's physics, but about the **structure of belief itself**. It visually argues that: * **Confidence is not truth:** The red-highlighted "hallucinated belief" with 0.9 confidence initiates the entire chain. * **Logic can amplify error:** From one false node (②-F), multiple conclusions branch, some true (③, ④) and some false (⑤). * **Contradiction can coexist:** The system holds both -40°C and -2°C as highly confident outcomes, exposing a failure in global consistency checking. The ultimate takeaway is a warning about relying on isolated confidence scores without examining the underlying logical structure and the truth of foundational premises. It advocates for systems that can detect and resolve such internal contradictions. </details> Figure 1: An example constructed belief tree. One key limitation of current methods is that they do not model noise or uncertainty in the LM generation process itself. LLMs’ “beliefs” are often poorly calibrated – even statements that models judge to be true with 100% probability are sometimes false. How can we produce reliable judgments about sets of related statements when even their assessments of individual statements are unreliable? To do so, we propose Belief Tree Prop agation (BTProp), a probabilistic method for improving the accuracy LLM “beliefs” with applications to hallucination detection. Specifically, given an initial statement generated by an LLM, BTProp first generates a belief tree of recursively generated claims that are logically related ot the initial statement. An example is shown in Figure 1, in which the root node of the belief tree is one sentence generated by the LLM about the freezing point of water, and each child node supports or contradicts its parent. Next, BTProp interprets this belief tree as a directed graphical model (a “hidden Markov tree"; Crouse et al., 1998; Durand et al., 2004) in which the LLM’s confidence scores are modeled as observed variables, and the ground-truth truthfulness of the statements as the corresponding hidden variables. Thus, the problem of assigning probabilities to statements can be reduced to probabilistic inference in a well-studied model family. Here, inference makes it possible to integrate knowledge about the (logical) structure of related statements, and the (probabilistic) relationship between statements and noisy LM outputs. Experiment results show that our method achieves state-of-the-art performance, improving the hallucination detection performance (AUROC and AUC-PR) by 3%–9% on FELM-Science Chen et al. (2023b) and FactCheckGPT Fadeeva et al. (2024) datasets relative to state-of-the-art baselines. ## 2 Related Work #### Hallucinations in LLMs. Hallucination has become a prominent topic in the research of LLMs Liu et al. (2022); Dhuliawala et al. (2023); Azaria and Mitchell (2023); Li et al. (2023); Zhang et al. (2023a). Prior work Ji et al. (2023); Huang et al. (2023) characterizes hallucination into two main types: factuality hallucination and faithfulness hallucination. Factuality hallucinations refer to outputs that are inconsistent with real-world facts. In comparison, faithfulness hallucinations shift to the generated content that deviates from the user’s instruction or user-provided contextual information. In this paper, we mainly focuses on factuality hallucinations and aim to better leverage the LLM’s intrinsic capabilities to detect such nonfactual statements within their generations. #### Hallucination detection. To detect hallucinations in LLM generation, the retrieval-based methods Shuster et al. (2021); Min et al. (2023); Chern et al. (2023); Semnani et al. (2023); Huo et al. (2023); Zhang et al. (2024) compare the LLM’s output with information from a reliable knowledge base. When an external knowledge base is inaccessible, another category of methods rely on the model’s intrinsic capabilities. Among them, some approaches leverage the reasoning ability of LLMs to detect and reduce hallucinations Chen et al. (2023b); Dhuliawala et al. (2023) via chain-of-thought prompting Wei et al. (2022). Uncertainty-based methods perform token-level or sentence-level uncertainty quantification Varshney et al. (2023); Fadeeva et al. (2024) to predict the factuality of model’s generation. Probing the model’s hidden states can also help detect hallucinations Azaria and Mitchell (2023); Chen et al. (2023a); Su et al. (2024). The sampled-based methods (or consistency-based methods) Manakul et al. (2023); Cao et al. (2023); Mündler et al. (2023) sample multiple responses and then perform consistency checks between these responses and the original model generation. Inconsistencies are then used to detect hallucinations. Our method is closely related to the consistency-based methods. Instead of sampling additional responses and performing unstructured consistency checks, our method generates logically-related statements organized in a tree structure. We further propose a probabilistic framework based on the hidden Markov tree model to check consistency and detect hallucinations. #### Improving factuality via logical consistency. A growing body of research explores improving LLMs’ factuality using the logical consistency across their beliefs Dalvi et al. (2021); Tafjord et al. (2022); Mitchell et al. (2022); Jung et al. (2022); Hase et al. (2023); Kassner et al. (2023); Akyürek et al. (2024). Starting from an initial text (e.g., a statement requiring truthfulness assessment or a question), these methods identify additional texts that are logically connected to the initial text and organize them in tree or graph structures. The inconsistencies of model’s beliefs on the factual correctness of these texts are resolved based on the inferential relations among these texts. One main limitation is that they treat the beliefs (LM-generated truth values or probabilities) as calibrated, which is generally not the case. Motivated by this, we build a probabilistic framework to integrate the model’s beliefs and figure out the most possible and reliable correctness evaluation of the initial text. ## 3 Method Given a statement $v$ from the LLM-generated text, e.g., A star’s temperature is determined by the amount of mass and energy it has, and an LLM, we aim to use the LLM to gauge the correctness of the statement. Many previous approaches use simple prompting techniques to directly elicit the LLM’s “belief”, or assessed probability of the statement, as a confidence score in $[0,1]$ . For example, one approach is to directly query the LLM for whether this is true. By sampling multiple answers from the LLM, we can obtain the confidence score as the fraction of answers that say yes (or some paraphrase thereof). Crucially, these scores may be wrong in several ways: in addition to being factually incorrect, they may be miscalibrated (reflecting an inappropriate degree of certainty or uncertainty) or inconsistent (e.g., assigning high probability to mutually incompatible statements, or different probabilities to logically equivalent statements). Below, we describe a method that addresses both categories of error by reasoning jointly about relationships between the truth values of related statements and LLM confidence scores. ### 3.1 BTProp : An Intuitive Overview The basic idea of our approach is to construct a belief tree, denoted $T$ , in which the root node is the target statement, each child node is a statement logically related to the parent node, and each edge represents the logical relationship between two nodes. We then obtain the confidence scores of all the nodes and use the logical consistency among them to correct any potential mistakes in the scores. Figure 2 shows an example, where the target statement is A star’s temperature is determined by the amount of mass and energy it has. This statement is true but assume that the LLM produces a low confidence score, $0.1$ , for the statement. To correct this error, we can construct a belief tree $T$ by generating two child nodes from the target node, where the first, which is true, is entailed by the target statement and the second, which is false, contradicts the target statement. Assume that the LLM correctly assigns a high confidence score to the former and a low score to the latter. In this case, we can easily recognize that the belief of the target node is logically inconsistent with those of the child nodes, and that increasing the confidence score of the target node would resolve the logical inconsistency, thereby correcting the mistake. <details> <summary>x2.png Details</summary>  ### Visual Description ## Diagram: Logical Relationships and Confidence Assessment of Statements on Stellar Temperature ### Overview The image is a logical diagram illustrating the relationships between three statements about the determinants of a star's temperature. It includes confidence scores for two of the statements and labels indicating entailment and contradiction. A note flags the central statement as a "Potential Hallucination." ### Components/Axes The diagram consists of three rectangular text boxes and connecting arrows with labels. There are no traditional chart axes. **Text Boxes (Statements):** 1. **Top-Left Box:** Contains the statement: "Stars with lower mass will have lower temperatures than more massive ones." * **Associated Value:** To its left, the number `0.9` is underlined. 2. **Bottom-Left Box:** Contains the statement: "External factors such as a surrounding nebula determines the temperature." * **Associated Value:** To its left, the number `0.1` is underlined. 3. **Right (Central) Box:** Contains the statement: "A star's temperature is determined by the amount of mass and energy it has." * **Associated Value:** Above this box, the text "Confidence: 0.1" is underlined. **Arrows and Labels:** * A **green arrow** labeled "Entails" points from the **Top-Left Box** to the **Right (Central) Box**. * A **red arrow** labeled "Contradicts" points from the **Bottom-Left Box** to the **Right (Central) Box**. **Annotations:** * A **dashed black line** connects the "Confidence: 0.1" label to the text "Potential Hallucination," which is written in red font. ### Detailed Analysis The diagram maps logical relationships between propositions. * **Statement 1 (Top-Left):** "Stars with lower mass will have lower temperatures than more massive ones." This is assigned a high confidence score of `0.9`. * **Statement 2 (Bottom-Left):** "External factors such as a surrounding nebula determines the temperature." This is assigned a low confidence score of `0.1`. * **Central Statement (Right):** "A star's temperature is determined by the amount of mass and energy it has." This statement is assigned a low confidence score of `0.1` and is explicitly labeled as a "Potential Hallucination." **Logical Flow:** 1. The high-confidence statement about mass and temperature (`0.9`) **entails** (logically supports or implies) the central statement. 2. The low-confidence statement about external factors (`0.1`) **contradicts** the central statement. ### Key Observations 1. **Confidence Disparity:** There is a stark contrast between the confidence scores of the two supporting/opposing statements (`0.9` vs. `0.1`). 2. **Central Claim Uncertainty:** The central claim, which is the target of the entailment and contradiction, itself has very low confidence (`0.1`) and is flagged as a potential error or fabrication ("hallucination"). 3. **Color Coding:** The diagram uses color semantically: green for a supportive relationship ("Entails") and red for a conflicting relationship ("Contradicts") and for the warning label ("Potential Hallucination"). ### Interpretation This diagram appears to be a diagnostic or analytical output, likely from a system evaluating the consistency and reliability of generated text or knowledge claims (e.g., from an AI model). * **What it demonstrates:** It visualizes a conflict in reasoning. A well-established, high-confidence astrophysical principle (mass-temperature relation) logically supports a broader, but poorly supported, central claim. Simultaneously, a low-confidence alternative explanation (external factors) contradicts that same central claim. * **The "Hallucination" Flag:** The critical insight is that the central claim—"A star's temperature is determined by the amount of mass and energy it has"—is itself deemed unreliable. The system suggests this statement may be a "hallucination," meaning it could be an incorrect, fabricated, or misleading assertion generated by a model, despite being partially entailed by a factual statement. The low confidence score (`0.1`) reinforces this assessment. * **Underlying Message:** The diagram highlights the complexity of fact-checking or validating information. A claim can be supported by a true premise yet still be flagged as unreliable, possibly because it is an oversimplification, contains inaccuracies (e.g., "energy" is vague), or contradicts other available evidence. It serves as a warning against accepting statements at face value, even when they seem logically connected to known facts. </details> Figure 2: Motivating example for the proposed method. The example above shows a simple case in which LLM beliefs about child nodes are all correct. In reality, these child nodes’ beliefs may themselves be incorrect. To mitigate these additional errors, we can construct a much larger belief tree with greater depth and many different statements. Assuming the errors in the confidence scores are sporadic, we can then expect to correct most of these errors, in both the target and child nodes, by taking into account all the confidence scores and their logical consistency. An example of such a full belief tree is shown in Figure 3. As implied by the discussion above, our algorithm consists of two components: ❶ Constructing the belief tree, and ❷ inferring the truthfulness of the target statement by integrating the confidence scores across the belief tree. Section 3.2 will discuss the former and Sections 3.3 - 3.5 the latter. ### 3.2 Belief Tree Construction Since the belief tree $T$ consists of nodes as statements and edges as the logical relationships between parent and child statements, the construction of the belief tree iterates between the following two steps. Step 1: Given a statement as the parent node, generate a set of logically connected statements as the children nodes. Step 2: Determine the logical connection between the children nodes and parent node. The iterative process starts with the target statement as the root node, and terminates when the maximum tree depth is reached. Each step is detailed below. #### Generating Child Statements As will be described below, BTProp is general, and compatible with any method for generating belief tree nodes from parent statements. We explore three specific strategies for statement generation: Strategy 1: Statement Decomposition. For statements containing multiple facts/claims, we decompose them into individual sub-statements. Figure 1 shows an example where the target statement (Node 0) contains multiple facts (the atmosphere pressure level and the freezing point of water on Mound Everest), It is then decomposed into two statements, one on the atmosphere pressure level and one on the freezing point of water. Statement decomposition can be achieved by prompting the LLM with in-context examples, as listed in Appendix A.7. Strategy 2: Supportive and Contradictory Premises. We prompt the LLM to generate a set of premises that are supportive or contradictory to the parent claim. When verifying the truthfulness of Node 2 in Figure 1 about the freezing point of water, we can prompt the model to generate contradictory premises (Node 3 and Node 4) that implies the limited influence of low pressure on the freezing point of water. By leveraging these generated premises, we can indeed correct the model’s wrong belief on Node 2 since the model’s confidence scores on these generated premises are high in this example. Strategy 3: Statement Correction. In this strategy, the LLM is instructed to generate what it believes to be corrected versions of the parent statement. The Node 5 in Figure 1 shows one example of the statement correction process, where the statement is about the freezing point of water. Then the corrected statements would be almost the same as the parent statement (Node 2), except that the actual temperature is replaced with alternatives that LLM believes may be true. This strategy is implemented via a three-step prompting process. First, the LLM is prompted to generate a question about the key information in the statement (e.g., the freezing point in Node 2 in Figure 1). Then, the we sample answers from the LLM. Finally, the LLM is prompted to generate a corrected statement from the original statement if it is wrong according to each sampled answer. Note that there can be multiple corrected statements because the LLM may produce different answers when sampled multiple times. Also, the corrected statements may include the parent statement itself if the LLM believes the parent statement is a likely answer. More details about this process can be found in Appendix A.7. Our method select the most appropriate strategies for each parent claim as follows. First, LLM first attempt the statement decomposition strategy. If the decomposition returns multiple statements, which indicates that the decomposition is meaningful, we would use them as the child statements. If only a single statement is returned, indicating that the parent statement cannot be further decomposed, we will prompt the LLM to select between strategies 2 and 3 with a list of rules and examples. The detailed prompts are in Appendix A.1. #### Determining Logical Relationships Given a pair of parent and child statements, $u$ and $v$ , we consider the following four logical relationships: ❶ Equivalence, $u⇔ v$ , if $u$ entails $v$ and $v$ entails $u$ ; ❷ entailment, $u⇒ v$ , if $u$ entails $v$ but $v$ is neutral to $u$ ; ❸ reverse entailment, $u⇐ v$ , if $u$ is neutral to $v$ but $v$ entails $u$ ; and ❹ contradiction, $u⇒¬ v$ (or, equivalently, $v⇒¬ u$ ). Note that we do not consider the completely neutral relationship because any statements determined as completely neutral to their parent statement will be removed. For the decomposition strategy (strategy 1), the child node statements are, by construction, jointly equivalent to the parent statement. Formally, denote $u$ as the parent node and $C(u)$ as the set of its child nodes, then their logical relationship is determined as $u⇔∩_v∈C(u)v$ . For strategies 2 and 3, since each child statement is independently generated, we only need to determine their individual relationship to the parent statement, rather than the joint relationship. To this end, we leverage an off-the-shelf natural language inference (NLI) model to infer the entailment, neutrality, or contradiction between the statements. For each pair of parent statement $u$ and an individual child statement $v$ , we derive two NLI relationships, one by setting $u$ as the premise and $v$ as the hypothesis, and the other with $u$ and $v$ switched. The two NLI outputs are then mapped to the aforementioned logical relationships – (entail, entail) mapped to equivalence; (entail, neutral) mapped to entailment; (neutral, entail) mapped to reverse entailment; and any results containing contradict in either direction mapped to contradiction. If the NLI module returns (neutral, neutral), the corresponding child statement will be discarded. #### Prior Belief Estimation The last step of the belief tree construction is to estimate the model’s belief (i.e., whether the statement is true) on each node. Specifically, we directly probe the LLM with the prompt ‘ True or False? {target statement} ’ and use the next token prediction probabilities of the words ‘True’ and ‘False’ to compute the model confidence. We normalize the prediction logits of the two words to get the confidence score of that statement. The only exception is in the case of statement correction, where the LLM is already instructed to output alternative statements it believes to be true. As a result, we simply set the confidence score of each generated statement to 1. Our empirical analysis shows that this would greatly reduce the number of LLM queries without deteriorating the performance. ### 3.3 A Hidden Markov Tree Model After the belief tree is constructed, the next question is how to utilize the confidence scores across the tree to better determine the truthfulness of the root statement, namely the target statement. To this end, we introduce a hidden Markov tree model and frame the truthfulness estimation problem as a probabilistic inference problem. Figure 3 shows an example hidden Markov tree model built upon the belief tree, where there are two layers of variables. The upper layer consists of the confidence score of each statement, denoted $\{S_u\}$ , which is estimated during the belief tree construction process. The lower layer consists of the binary variables representing the actual correctness of each statement, denoted as $\{Z_u\}$ . $Z_u=T$ if statement $v$ is correct and $Z_u=F$ otherwise. $\{S_u\}$ are observed variables while $\{Z_u\}$ are hidden variables. Given the above hidden Markov tree model, determining the truthfulness of the target statement can be cast as computing the posterior probability, i.e., $p(Z_0=T\mid\{S_u\})$ , which means the truthfulness of the root node given the confidence scores on all the nodes in this belief tree. To computing $p(Z_0=T\mid\{S_u\})$ , we need to estimate the following probabilities. <details> <summary>x3.png Details</summary>  ### Visual Description ## Diagram: An HMM on Belief Tree ### Overview The image is a technical diagram illustrating a Hidden Markov Model (HMM) structured as a belief tree. It visually represents the probabilistic relationship between hidden ground-truth states and observed beliefs (likely from a Large Language Model), and how belief propagation updates a posterior belief. The diagram is divided into two main sections: a graphical model on the left and a legend on the right. ### Components/Axes The diagram contains no traditional chart axes. Its components are nodes, arrows, and explanatory text. **1. Main Diagram (Left Panel - Blue Border):** * **Title:** "An HMM on Belief Tree" (centered at the top). * **Nodes:** Two types of circular nodes. * **Solid-line circles (S nodes):** Labeled `S₀`, `S₁`, `S₂`, `S₃`, `S₄`. `S₀` is annotated with blue text "Prior belief" pointing to it. * **Dashed-line circles (Z nodes):** Labeled `Z₀`, `Z₁`, `Z₂`, `Z₃`, `Z₄`. These are shaded grey. * **Arrows:** * **Solid black arrows:** Connect Z nodes to other Z nodes (e.g., `Z₀` to `Z₁` and `Z₂`; `Z₂` to `Z₃` and `Z₄`). * **Double-lined upward arrows:** Connect each Z node to its corresponding S node (e.g., `Z₀` to `S₀`; `Z₁` to `S₁`). * **Posterior Belief Box:** A pink rectangular box at the bottom-left containing the formula `p(Z₀ = T|{s_u})`. It is labeled with red text "Posterior belief" above it. * **Belief Propagation Arrow:** A red dotted arrow points from the right side of the diagram towards the posterior belief box, labeled "Belief prop.". **2. Legend (Right Panel - Green Border):** A vertical list defining the symbols used in the main diagram. * **Top Entry:** A dashed-line circle labeled `Zᵤ` with the description "Ground-truth factuality (hidden)". * **Second Entry:** A solid-line circle labeled `Sᵤ` with the description "LLM's prior belief (observed)". * **Third Entry:** A solid black right-pointing arrow `→` with the description "p(Z_{c(u)}|Zᵤ) Transition probability". * **Fourth Entry:** A double-lined upward arrow `⇈` (visually similar to `↑↑`) with the description "p(Sᵤ|Zᵤ) Emission probability". * **Fifth Entry:** A red dotted left-pointing arrow `⇠` (visually similar to `←` but dotted) with the description "Belief propagation". * **Bottom Entry:** A pink rectangle labeled `p` with the description "Posterior belief". ### Detailed Analysis **Tree Structure and Flow:** The diagram depicts a tree with a root and three levels of depth. 1. **Root:** The hidden state `Z₀` is the root. It emits the observed belief `S₀` (the "Prior belief"). 2. **First Branch:** `Z₀` transitions to two child hidden states: `Z₁` and `Z₂`. * `Z₁` emits `S₁`. * `Z₂` emits `S₂`. 3. **Second Branch:** `Z₂` further transitions to two child hidden states: `Z₃` and `Z₄`. * `Z₃` emits `S₃`. * `Z₄` emits `S₄`. **Symbol Cross-Reference:** * The **dashed circle** in the legend (`Zᵤ`) matches all `Z` nodes in the tree (`Z₀`-`Z₄`), representing hidden ground-truth states. * The **solid circle** in the legend (`Sᵤ`) matches all `S` nodes in the tree (`S₀`-`S₄`), representing observed LLM beliefs. * The **solid arrow** in the legend matches the arrows connecting `Z` nodes (e.g., `Z₀` → `Z₁`), representing the transition probability between hidden states. * The **double upward arrow** in the legend matches the arrows from `Z` to `S` nodes (e.g., `Z₀` ⇈ `S₀`), representing the emission probability (how a hidden state generates an observation). * The **red dotted arrow** in the legend matches the "Belief prop." arrow in the main diagram, indicating the direction of belief propagation flow. * The **pink rectangle** in the legend matches the posterior belief box in the main diagram. **Text Transcription:** All text in the image is in English. The precise transcriptions are: * Title: "An HMM on Belief Tree" * Annotations: "Prior belief", "Posterior belief", "Belief prop." * Node Labels: `S₀`, `S₁`, `S₂`, `S₃`, `S₄`, `Z₀`, `Z₁`, `Z₂`, `Z₃`, `Z₄` * Formula: `p(Z₀ = T|{s_u})` * Legend Text: "Ground-truth factuality (hidden)", "LLM's prior belief (observed)", "p(Z_{c(u)}|Zᵤ) Transition probability", "p(Sᵤ|Zᵤ) Emission probability", "Belief propagation", "Posterior belief" ### Key Observations 1. **Hierarchical Belief Structure:** The model is not a linear chain but a tree, allowing a single prior belief (`S₀` from `Z₀`) to branch into multiple, potentially divergent lines of factual inquiry (`Z₁`, `Z₂`, etc.). 2. **Direction of Inference:** The "Belief prop." arrow points *from* the observed data (the collection of all `Sᵤ` nodes, denoted `{s_u}` in the formula) *back towards* the root hidden state `Z₀`. This visually represents the process of using all observed beliefs to update the estimate of the initial ground truth. 3. **Core Formula:** The posterior belief `p(Z₀ = T|{s_u})` is the central output. It represents the probability that the root ground-truth state `Z₀` is True (`T`), given all the observed LLM beliefs `{s_u}` from across the entire tree. 4. **Model Type:** It is explicitly an HMM, implying the hidden states (`Z`) follow a Markov property (the future depends only on the present), and observations (`S`) are conditionally independent given their hidden state. ### Interpretation This diagram models a sophisticated process for assessing the factual accuracy of an LLM's outputs. It treats the LLM's statements (`S` nodes) not as direct truth, but as noisy observations emitted from an underlying, hidden factual reality (`Z` nodes). The tree structure suggests a scenario where an initial claim or question (`Z₀`) leads to a series of related sub-claims or lines of reasoning. The LLM generates beliefs (`S₁`, `S₂`, etc.) about each of these sub-claims. The key insight is that these beliefs are interconnected through the hidden factual structure. The "belief propagation" process is the investigative or reasoning step. It works backwards from the collection of all the LLM's statements to infer the most probable state of the original, root fact (`Z₀`). This is a powerful framework for fact-checking or consistency analysis: if the LLM's beliefs across a complex tree of implications are coherent, confidence in the root fact increases. Contradictions or improbable beliefs in the tree would propagate back to lower the posterior probability of `Z₀` being true. In essence, the diagram provides a mathematical and visual framework for moving from a collection of potentially unreliable AI-generated statements to a probabilistic estimate of an underlying truth, accounting for the logical dependencies between those statements. </details> Figure 3: An example hidden Markov tree model. The first set is the emission probability, $p(S_u\mid Z_u)$ , which characterizes the LLM’s confidence score of the statement given its underlying truthfulness. The emission probability captures the quality of LLM’s original confidence score. If the confidence score is accurate, then $p(S_u\mid Z_u=T)$ will concentrate its probability mass towards $1$ , and $p(S_u\mid Z_u=F)$ will concentrate its probability mass towards $0 0$ . Otherwise, the conditional distributions will be more spread. The second set of distribution is the transition probability, $p(Z_C(u)\mid Z_u)$ , where $Z_C(u)$ represent the set of truthfulness variables of all the child nodes of $u$ . The transition probability is the joint probability of the truthfulness of the child statements given that of the parent statement, which captures the logical relationships among these statements. The third set of distribution is the prior distribution of hidden variables, i.e., $p(Z_u=z)$ , which refers to the probability of an arbitrary statement $u$ generated by the LLM being true or false. We use uniform prior (both 0.5 for $z=T$ and $z=F$ ) in this paper. Therefore, the process of computing the posterior probability contains two steps: ❷ the estimation step, to first estimate these probabilities on a held-out dataset. ❸ the inference step, to compute the posterior probability for each testing example. Section 3.4 will discuss how to determine the emission and transition probabilities and Section 3.5 how to infer $p(Z_0=T\mid\{S_u\})$ . ### 3.4 Determining Conditional Probabilities #### Determining Emission Probabilities The emission probabilities, $p(S_u\mid Z_u)$ , can be estimated from labeled datasets. Specifically, given a dataset of statements with truthfulness labels, we can run the LLM to obtain the confidence scores of all the statements. We can then estimate $p(S_u\mid Z_u=T)$ by obtaining the empirical distribution of the confidence scores within the statements labeled as true, and $p(S_u\mid Z_u=F)$ within the statements labeled as false. To obtain the empirical distribution of the continuous confidence scores, we first quantize the support $[0,1]$ into multiple bins, and count the histogram of the confidence scores falling into each bin. The boundaries of each bin are listed in Table 1. Given a particular confidence score on a statement, we look up the table to get the corresponding emission probability. For example, a confidence score 0.95 leads to $p(S_u=0.95\mid Z_u=F)=0.65$ . | | $S_u∈[0.0,0.2)$ | $[0.2,0.4)$ | $[0.4,0.7)$ | $[0.7,0.9)$ | $[0.9,1.0]$ | | --- | --- | --- | --- | --- | --- | | $Z_u=\texttt{True}$ | 0.12 | 0.05 | 0.10 | 0.08 | 0.65 | | $Z_u=\texttt{False}$ | 0.30 | 0.10 | 0.15 | 0.13 | 0.32 | Table 1: Empirical estimation of the emission probability on Wikibio-GPT3 dataset Manakul et al. (2023). #### Determining Transition Probabilities The transition probability, $p(Z_C(u)\mid Z_u)$ , is a multivariate Bernoulli distribution with the support $\{T,F\}^m$ , where $m$ is the size of $C(u)$ . As discussed in Section 3.2, different statement generation strategies have different logical relationship properties. For the statement decomposition strategy, the child statements are always jointly equivalent to the parent statement. Therefore, when the parent claim is true, i.e., $Z_u=T$ , we know that all of the child statements must be true, and thus $p(Z_C(u)=\bm{z}\mid Z_u=T)$ equals one when every element in $\bm{z}$ is true and zero otherwise. Conversely, when the parent claim is false, i.e., $Z_u=F$ , we can infer that at least one child statement is wrong, so $p(Z_C(u)=\bm{z}\mid Z_u=F)$ is set to zero when every element in $\bm{z}$ is true and uniformly set to $1/(2^m-1)$ otherwise. For the other two strategies, the child statements of the same parent statement are considered independent conditional on the parent statement. Therefore, we can decompose the probability $p(Z_C(u)\mid Z_u)$ into $∏_v∈C(u)p(Z_v\mid Z_u)$ , where each $p(Z_v\mid Z_u)$ is the transitional probability between parent statement $u$ and one child statement $v$ . $p(Z_v\mid Z_u)$ is determined based on the four different possible logical relationships between $u$ and $v$ (recall the discussion in Section 3.2), which is listed in Table 2. | | $u⇔ v$ | $u⇒ v$ | $u⇐ v$ | $u⇒¬ v$ | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | $u=T$ | $u=F$ | $u=T$ | $u=F$ | $u=T$ | $u=F$ | $u=T$ | $u=F$ | | | $v=T$ | 1.0 | 0.0 | 1.0 | $p_t$ | $p_t$ | 0.0 | 0.0 | $p_t$ | | $v=F$ | 0.0 | 1.0 | 0.0 | $p_f$ | $p_f$ | 1.0 | 1.0 | $p_f$ | Table 2: Transition probability given different logical relationships between two nodes. $p_t$ and $p_f$ are set to $0.5$ in this paper. ### 3.5 Inferring the Posterior Probability of $Z_0$ Inferring the posterior probability $p(Z_0=T\mid\{S_u\})$ is a standard inference problem in hidden Markov tree models and we can apply the standard upward-downward algorithm Crouse et al. (1998) to efficiently compute this probability. For brevity, we only describe the gist of the algorithm here. More details can be found in Appendix A.4 and previous work Crouse et al. (1998); Durand et al. (2004). The algorithm introduces an auxiliary conditional distribution $β(z,u)≡ p(S_T(u)\mid Z_u=z)$ , where $z∈\{T,F\}$ and $S_T(u)$ represents the confidence scores of all the nodes in the sub-tree whose root node is $u$ . Note that the posterior probability of our interest, $p(Z_0=T\mid\{S_T(0)\})$ , can be computed as $$ \displaystyle p(Z_0=T\mid\{S_T(0)\})=\frac{β(T,0)p(Z_0=T)} {∑_z∈\{T,F\}β(z,0)p(Z_0=z)}, \tag{0} $$ where $p(Z_0=T)$ and $p(Z_0=F)$ are exactly the prior probabilities we mentioned in Section 3.3, and both are set to be 0.5. Therefore, the posterior probability above can be represented as $β(T,0)/(β(F,0)+β(T,0))$ . According to the Bayesian rule, we can derive a recursive relationship between $β(z,u)$ of a parent node $u$ and those of the child nodes: $$ \displaystyleβ(z,u)=p(S_u\mid Z_u=z)· \displaystyle∑_Z_{C(u)∈\{T,F\}^m}p(Z_C(u)\mid Z _u=z)∏_v∈C(u)β(Z_v,v), \tag{1} $$ where $m$ is the size of $C(u)$ (namely the number of child statements to node $u$ ). Note that the first term on the RHS is the emission probability, and the first term inside the summation is the transition probability, which are both already known. Therefore, Equation 1 provides a way to compute $β(z,0)$ recursively from the leaf nodes back to the root node. First, we compute the $β(z,u)$ of the leaf nodes as $β(z,u)=p(S_u\mid Z_u=z)$ . Then, we use Equation 1 to recursively compute the $β(z,u)$ of a parent node $u$ from their child nodes, until the root node is reached. The entire process is essentially propagating and merging the beliefs in the confidence scores from sub-trees upward to the parent node. By the time we reach the root node, we have gathered the information of all the confidence scores. Hence this process is also referred to as a belief propagation process. We summarize the whole algorithm in Appendix A.5. ## 4 Experiments In this section, we conduct empirical evaluations on widely-used hallucination detection benchmarks. ### 4.1 Experiment Configurations #### Datasets We follow the previous work of hallucination detection Manakul et al. (2023); Chen et al. (2023b) and use the following datasets for evaluation: Wikibio-GPT3 Manakul et al. (2023), FELM-Science Chen et al. (2023b), and FactCheckGPT Wang et al. (2023). Wikibio-GPT3 mainly consists of biography articles generated by LLMs, whereas the other two datasets cover a broader range of topics such as physics, chemistry, and computer science. | Method | Backbone | AUROC | AUC-PR | F1 | Acc | Backbone | AUROC | AUC-PR | F1 | Acc | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | Wikibio-GPT3 | | | | | | | | | | | | Prior Confidence | gpt-3.5 -turbo | 73.1 | 85.7 | 84.5 | 76.3 | Llama3-8B Instruct | 71.0 | 85.7 | 85.5 | 75.9 | | CoT | 71.3 | 83.4 | 85.2 | 76.4 | 72.0 | 84.4 | 85.3 | 75.7 | | | | Semantic Uncertainty | 70.8 | 86.1 | 84.3 | 73.7 | 60.3 | 78.7 | 84.5 | 73.6 | | | | SelfCheckGPT | 82.6 | 91.3 | 86.6 | 80.0 | 77.0 | 86.8 | 86.1 | 76.8 | | | | Focus | - | - | - | - | 73.0 | 85.1 | 86.2 | 77.3 | | | | BTProp | 80.7 | 90.4 | 87.6 | 80.4 | 74.0 | 86.2 | 86.1 | 77.5 | | | | FELM-Science | | | | | | | | | | | | Prior Confidence | gpt-3.5 -turbo | 75.5 | 37.2 | 42.1 | 81.6 | Llama3-8B Instruct | 76.1 | 38.2 | 44.0 | 80.6 | | CoT | 56.3 | 19.0 | 28.6 | 72.5 | 61.6 | 25.7 | 32.6 | 71.2 | | | | Semantic Uncertainty | 59.1 | 25.6 | 32.6 | 75.2 | 52.5 | 17.2 | 27.9 | 79.6 | | | | SelfCheckGPT | 77.4 | 45.7 | 51.3 | 84.4 | 77.8 | 39.5 | 49.1 | 76.5 | | | | Maieutic Prompting | - | - | 27.2 | 82.6 | - | - | 22.9 | 79.2 | | | | Focus | - | - | - | - | 69.4 | 43.5 | 41.7 | 76.5 | | | | BTProp | 79.1 | 52.3 | 56.5 | 81.6 | 77.8 | 48.2 | 51.3 | 82.8 | | | | FactCheckGPT | | | | | | | | | | | | Prior Confidence | gpt-3.5 -turbo | 76.0 | 52.6 | 53.6 | 71.5 | Llama3-8B Instruct | 71.9 | 47.6 | 50.8 | 60.5 | | CoT | 66.5 | 38.9 | 47.7 | 66.4 | 72.5 | 43.3 | 53.2 | 66.1 | | | | Semantic Uncertainty | 56.3 | 33.9 | 40.9 | 65.9 | 57.6 | 34.1 | 40.8 | 49.9 | | | | SelfCheckGPT | 74.9 | 49.7 | 54.4 | 74.0 | 72.5 | 46.4 | 51.9 | 65.9 | | | | Maieutic Prompting | - | - | 35.0 | 72.2 | - | - | 39.8 | 66.8 | | | | BTProp | 79.4 | 54.3 | 60.2 | 75.3 | 73.9 | 49.1 | 55.3 | 72.6 | | | Table 3: Hallucination detection performance of different methods. We report AUROC, ROC-PR, F1 score, and detection accuracy(Acc) for all methods with two backbone models. The best results are highlighted in bold. #### Evaluation settings and metrics We conduct hallucination detection at sentence-level on Wikibio-GPT3 and FactCheckGPT and segment-level on FELM-Science following the default settings. Our evaluation metrics include the area under the receiver operator characteristic curve (AUROC), area under the precision-recall curve (AUC-PR), F1 score, and detection accuracy. As F1 score and detection accuracy evaluation require a decision threshold, we search for the optimal threshold to maximize the F1 score for each method and compute the two metrics. The hallucinated examples are considered as positive instances following the default configurations of these datasets. Moreover, Wikibio-GPT3 and FELM-Science exhibit significant class imbalances, so the detection accuracy on these datasets serve merely as reference points. #### Baselines We include the following baselines for comparison. (1) Prior Confidence , which directly queries the model’s confidence on the truthfulness of each sentence or segment. (2) CoT, which prompts the model to first generate a reasoning process before deciding the truthfulness. We adopt the prompting method from the official FELM Chen et al. (2023b) dataset, with slight modifications for sentence-level and segment-level hallucination detection. (3) SelfCheckGPT Manakul et al. (2023), which samples additional responses from the model and use the inconsistency between each response and the target statement for hallucination detection. Among the multiple variants, we choose SelfCheckGPT -prompt with the best performance for comparison. (4) Maieutic Prompting Jung et al. (2022), which builds a belief tree via backward-chaining and then infers the truth-value of the original statement that resolves the inconsistencies. (5) Semantic Uncertainty Kuhn et al. (2023); Farquhar et al. (2024), which estimate the model’s predictive uncertainty on the generation for hallucination detection. (6) Focus (Zhang et al., 2023b), which quantify the predictive uncertainty of the LLM on given texts to detect hallucinations. Note that we only include Focus when the backbone is Llama3 since it requires full access to the LLM. #### Implementation details We evaluate our methods and baselines using GPT-3.5-turbo-0125 and Llama-3-8B-Instruct. For our method, we set the maximum belief tree depth to $2$ . We employ greedy decoding during belief tree construction and prior belief estimation. The exception is the statement correction strategy, where we sample $5$ corrected statements using temperature $0.7$ . Additionally, since each statement in FactCheckGPT is manually processed to ensure it contains only one property or fact, we do not apply statement decomposition when build the belief tree for statements from FactCheckGPT. We use the first 120 examples in the Wikibio-GPT3 dataset to estimate the emission probability in our method, and validate it on the remaining examples and two other datasets. More details are in Appendix A.1. ### 4.2 Experiment Results #### Overall comparison We evaluate the effectiveness of BTProp with the experiment results in Table 3. We highlight the following observations. First, our method achieves the best performance on FELM-Science and FactCheckGPT datasets across different backbones, demonstrating the superiority of our method. BTProp improves upon the best baselines by 3% - 9% on AUROC and ROC-PR. The only exception is the Wikibio-GPT3 dataset, where the SelfCheckGPT is more effective for detection hallucinated outputs in biographies generated by LLMs. Second, compared to SelfCheckGPT which leverages contradictions between the target statement and the sampled responses for hallucination detection, our method is more effective on detecting hallucinated responses related to scientific knowledge. Both FELM-Science and FactCheckGPT datasets contain a significant proportion of questions on scientific knowledge, and our method achieves the best performance on them. Third, chain-of-thought prompting is less effective in hallucination detection, especially on the FELM-Science dataset. This finding aligns with the experimental results in the original FELM dataset paper. Our experiments show that the model tends to regard the input sentence as true most of the time, leading to sub-optimal performance. The Maieutic-prompting method, which is originall designed to verify the correctness of statements related to commonsense reasoning, is less effective for hallucination detection. Even on the FELM-Science and FactCheckGPT datasets which contain many statements about scientific knowledge, it remains less effective compared to other methods. We also find that the Semantic Uncertainty method is more effective on detection hallucinations in biography generation. However, its effectiveness diminishes when applied to the other two datasets. In contrast, our method consistently delivers competitive performance across a variety of benchmarks. <details> <summary>x4.png Details</summary>  ### Visual Description Icon/Small Image (382x4) </details> <details> <summary>x5.png Details</summary>  ### Visual Description Icon/Small Image (382x4) </details> Figure 4: Performance-efficiency comparison. #### Deployment efficiency One concern on our method is the belief tree construction would be inefficient. We compare the time cost of our method with the baselines on the FELM-Science and FactCheckGPT datasets, where our method significantly outperform baselines. We use the gpt-3.5-turbo as the backbone, and eliminate CoT prompting method on FELM-Science dataset due to its poor performance. We visualize the trade-off between hallucination detection performance and the time cost in Figure 4. The baseline methods, including SelfCheckGPT and CoT prompting, samples additional 20 responses for hallucination detection. We vary the number of samples and test the corresponding performance to visualize the trade-off. As depicted in the graph, the baseline performance shows diminishing returns in performance improvement with increased computational complexity. As time cost increases, the performance growth of the two baseline models gradually levels off, eventually reaching a point where additional time does not translate into significantly better performance. In comparison, our method can effectively trades off increased time cost for significantly improved performance. Despite its higher computational complexity, it achieves a marked improvement in performance metrics, underscoring its efficiency and effectiveness. #### Hallucination detection in texts generated by the backbone model. Since the datasets in the above experiments consist of text generated by another model for hallucination detection, we also include texts generated by the backbone model itself to to provide a more comprehensive evaluation of our method’s effectiveness. Detailed results are available in Appendix A.2, where we show that our method can also effectively detect the hallucinations in the text generated by the backbone model itself. #### Qualitative results We visualize several examples of the constructed belief trees to demonstrate how the inconsistencies in model’s beliefs help detect the hallucination. Due to space limit, we put these examples and the analyses in Appendix A.3. | | AUROC | AUC-PR | F1 | Acc | | --- | --- | --- | --- | --- | | BTProp | 79.1 | 52.3 | 56.5 | 81.6 | | Decomposition only | 67.9 | 31.2 | 43.3 | 79.6 | | Premise only | 70.2 | 30.7 | 40.4 | 78.6 | | Correction only | 77.4 | 47.5 | 53.0 | 81.5 | | Prior Confidence | 75.5 | 37.2 | 42.1 | 81.6 | Table 4: Performance of BTProp given different child node generation strategies. #### Ablation study: belief tree construction We introduce three strategies to build the belief tree: statement decomposition, generating supportive and contradictory premises, and statement correction. We perform ablation study to demonstrate the necessity to introduce these different strategies in Table 4 using gpt-3.5-turbo-0125 as the backbone model. Specifically, we evaluate the performance of our method when only use one child node generation strategies on the FELM-Science dataset. We highlight the following observations. First, only applying one child node generation strategy is sub-optimal. There exist a significant gap between using one strategy and using them all. Second, statement correction tends to contribute most to the performance improvement, but when combining it with other strategies, the performance can still be further improved. ## 5 Conclusion In this paper, we propose BTProp, a method that leverages the model’s intrinsic capability for hallucination detection. Given a statement from the LLM-generated texts, our method organizes the model’s intrinsic beliefs on neighboring statements in a belief tree. By introducing the hidden Markov tree model, we convert the the hallucination detection into a posterior probability estimation problem and propose corresponding solutions to solve it. Experiment results have demonstrate the effectiveness of our method. The future direction would be how to further improve the belief tree construction method to make it more effective and efficient. ## 6 Societal Impact and limitations In this paper, our primary goal is to develop an algorithm that can integrate the model’s internal beliefs on logically-connected statements to detect hallucinations in texts generated by the LLM. Our method is designed to improve the trustworthiness of LLMs and make fully use of their intrinsic capabilities. Therefore, our method is less likely to introduce the unintended risks. We also assess the experiments to ensure they are devoid of any harmful content or adverse impacts. While BTProp improves the hallucination detection performance and enhance the trustworthiness of LLMs, it involves collecting a set of augmented statements (the belief tree) to perform inference. Therefore, the main limitation of our method is its high time cost. Constructing a belief tree necessitates multiple queries to the large language model, leading to significant delays. Additionally, the time complexity of generating child nodes increases exponentially as more layers are added to the tree. A potential solution to this issue could be to develop a mechanism that selectively expands nodes within the belief tree, thereby optimizing the process. ## Acknowledgments The work of Bairu Hou and Shiyu Chang was partially supported by National Science Foundation (NSF) Grant IIS-2338252, NSF Grant IIS-2207052, and NSF Grant IIS-2302730. The work of Jacob Andreas is supported by a Sloan Fellowship and the NSF under grant IIS-2238240. The computing resources used in this work were partially supported by the Accelerate Foundation Models Research program of Microsoft and CAIS Compute Cluster of Center for AI Safety. ## References - Akyürek et al. (2024) Afra Feyza Akyürek, Ekin Akyürek, Leshem Choshen, Derry Wijaya, and Jacob Andreas. 2024. Deductive closure training of language models for coherence, accuracy, and updatability. arXiv preprint arXiv:2401.08574. - Azaria and Mitchell (2023) Amos Azaria and Tom Mitchell. 2023. The internal state of an llm knows when it’s lying. In The 2023 Conference on Empirical Methods in Natural Language Processing. - Cao et al. (2023) Zouying Cao, Yifei Yang, and Hai Zhao. 2023. Autohall: Automated hallucination dataset generation for large language models. arXiv preprint arXiv:2310.00259. - Chen et al. (2023a) Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2023a. Inside: Llms’ internal states retain the power of hallucination detection. In The Twelfth International Conference on Learning Representations. - Chen et al. (2023b) Shiqi Chen, Yiran Zhao, Jinghan Zhang, I-Chun Chern, Siyang Gao, Pengfei Liu, and Junxian He. 2023b. Felm: Benchmarking factuality evaluation of large language models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track. - Chern et al. (2023) I Chern, Steffi Chern, Shiqi Chen, Weizhe Yuan, Kehua Feng, Chunting Zhou, Junxian He, Graham Neubig, Pengfei Liu, et al. 2023. Factool: Factuality detection in generative ai–a tool augmented framework for multi-task and multi-domain scenarios. arXiv preprint arXiv:2307.13528. - Crouse et al. (1998) Matthew S Crouse, Robert D Nowak, and Richard G Baraniuk. 1998. Wavelet-based statistical signal processing using hidden markov models. IEEE Transactions on signal processing, 46(4):886–902. - Dalvi et al. (2021) Bhavana Dalvi, Peter Jansen, Oyvind Tafjord, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura, and Peter Clark. 2021. Explaining answers with entailment trees. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7358–7370. - Dhuliawala et al. (2023) Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023. Chain-of-verification reduces hallucination in large language models. arXiv preprint arXiv:2309.11495. - Durand et al. (2004) J-B Durand, Paulo Goncalves, and Yann Guédon. 2004. Computational methods for hidden markov tree models-an application to wavelet trees. IEEE Transactions on Signal Processing, 52(9):2551–2560. - Fadeeva et al. (2024) Ekaterina Fadeeva, Aleksandr Rubashevskii, Artem Shelmanov, Sergey Petrakov, Haonan Li, Hamdy Mubarak, Evgenii Tsymbalov, Gleb Kuzmin, Alexander Panchenko, Timothy Baldwin, et al. 2024. Fact-checking the output of large language models via token-level uncertainty quantification. arXiv preprint arXiv:2403.04696. - Fadeeva et al. (2023) Ekaterina Fadeeva, Roman Vashurin, Akim Tsvigun, Artem Vazhentsev, Sergey Petrakov, Kirill Fedyanin, Daniil Vasilev, Elizaveta Goncharova, Alexander Panchenko, Maxim Panov, et al. 2023. Lm-polygraph: Uncertainty estimation for language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 446–461. - Farquhar et al. (2024) Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625–630. - Hase et al. (2023) Peter Hase, Mona Diab, Asli Celikyilmaz, Xian Li, Zornitsa Kozareva, Veselin Stoyanov, Mohit Bansal, and Srinivasan Iyer. 2023. Methods for measuring, updating, and visualizing factual beliefs in language models. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2714–2731. - Huang et al. (2023) Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2023. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232. - Huo et al. (2023) Siqing Huo, Negar Arabzadeh, and Charles LA Clarke. 2023. Retrieving supporting evidence for llms generated answers. arXiv preprint arXiv:2306.13781. - Ji et al. (2023) Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1–38. - Jung et al. (2022) Jaehun Jung, Lianhui Qin, Sean Welleck, Faeze Brahman, Chandra Bhagavatula, Ronan Le Bras, and Yejin Choi. 2022. Maieutic prompting: Logically consistent reasoning with recursive explanations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1266–1279. - Kassner et al. (2023) Nora Kassner, Oyvind Tafjord, Ashish Sabharwal, Kyle Richardson, Hinrich Schuetze, and Peter Clark. 2023. Language models with rationality. In The 2023 Conference on Empirical Methods in Natural Language Processing. - Kuhn et al. (2023) Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations. - Kwon et al. (2023) Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles. - Li et al. (2023) Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023. Halueval: A large-scale hallucination evaluation benchmark for large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6449–6464. - Liu et al. (2022) Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao, Zhifang Sui, Weizhu Chen, and William B Dolan. 2022. A token-level reference-free hallucination detection benchmark for free-form text generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6723–6737. - Manakul et al. (2023) Potsawee Manakul, Adian Liusie, and Mark Gales. 2023. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. In The 2023 Conference on Empirical Methods in Natural Language Processing. - Min et al. (2023) Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. Factscore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12076–12100. - Mitchell et al. (2022) Eric Mitchell, Joseph Noh, Siyan Li, Will Armstrong, Ananth Agarwal, Patrick Liu, Chelsea Finn, and Christopher D Manning. 2022. Enhancing self-consistency and performance of pre-trained language models through natural language inference. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1754–1768. - Mündler et al. (2023) Niels Mündler, Jingxuan He, Slobodan Jenko, and Martin Vechev. 2023. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. In The Twelfth International Conference on Learning Representations. - Semnani et al. (2023) Sina Semnani, Violet Yao, Heidi Zhang, and Monica Lam. 2023. Wikichat: Stopping the hallucination of large language model chatbots by few-shot grounding on wikipedia. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2387–2413. - Shuster et al. (2021) Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3784–3803. - Su et al. (2024) Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. 2024. Unsupervised real-time hallucination detection based on the internal states of large language models. arXiv preprint arXiv:2403.06448. - Tafjord et al. (2022) Oyvind Tafjord, Bhavana Dalvi Mishra, and Peter Clark. 2022. Entailer: Answering questions with faithful and truthful chains of reasoning. arXiv preprint arXiv:2210.12217. - Varshney et al. (2023) Neeraj Varshney, Wenlin Yao, Hongming Zhang, Jianshu Chen, and Dong Yu. 2023. A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation. arXiv preprint arXiv:2307.03987. - Wang et al. (2022) Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations. - Wang et al. (2023) Yuxia Wang, Revanth Gangi Reddy, Zain Muhammad Mujahid, Arnav Arora, Aleksandr Rubashevskii, Jiahui Geng, Osama Mohammed Afzal, Liangming Pan, Nadav Borenstein, Aditya Pillai, et al. 2023. Factcheck-gpt: End-to-end fine-grained document-level fact-checking and correction of llm output. arXiv preprint arXiv:2311.09000. - Wei et al. (2022) Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837. - Zhang et al. (2024) Jiawei Zhang, Chejian Xu, Yu Gai, Freddy Lecue, Dawn Song, and Bo Li. 2024. Knowhalu: Hallucination detection via multi-form knowledge based factual checking. arXiv preprint arXiv:2404.02935. - Zhang et al. (2023a) Muru Zhang, Ofir Press, William Merrill, Alisa Liu, and Noah A Smith. 2023a. How language model hallucinations can snowball. arXiv preprint arXiv:2305.13534. - Zhang et al. (2023b) Tianhang Zhang, Lin Qiu, Qipeng Guo, Cheng Deng, Yue Zhang, Zheng Zhang, Chenghu Zhou, Xinbing Wang, and Luoyi Fu. 2023b. Enhancing uncertainty-based hallucination detection with stronger focus. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 915–932. ## Appendix A Appendix / supplemental material ### A.1 Implementation Details #### Data preprocessing We preprocess the data in Wikibio-GPT3 Manakul et al. (2023) and FELM-Science Chen et al. (2023b) before evaluating our method and baselines. The data preprocessing includes two steps. First, we perform decontextualization since the sentences in a response generated by the LLM might be context-dependent and contain pronouns or noun phrases that cannot be understood without full context. Take the first data instance from Wikibio-GPT3 for demonstration, which is a biography of “John Russell Reynolds” generated by an LLM. The second sentence within the biography requiring hallucination detection is “He was born in London, the son of a barrister, and was educated at Eton College and Trinity College, Cambridge.” Without the context, we cannot identify the truthfulness of this sentence due to the pronoun. Therefore, we first decontextualize the sentences in the two datasets by prompting gpt-3.5-turbo-0125. The prompt is available in Figure 9. Second, we further manually clean up the dataset by filtering out some sentences that are not check-worthy but still annotated as “true” in the datasets. For example, the FELM contains some sentences such as “Sure!” and “If you have any further questions or concerns, please let me know.”. These sentences are annotated as “true” and will be counted into performance evaluation. Since both two datasets contains not too many examples, we manually filter out these sentences to exclude them in our evaluation. The datasets after our preprocessing and filtering are available in the supplemental material and will be made open-sourced. #### Implementation of our method We use the vLLM Kwon et al. (2023) to perform the inference of Llama-3-8B-Instruct. We use a single NVIDIA A100 80GB PCIe GPU to evaluate the performance and report the time cost in Figure 4. For the emission probability estimation (Table 1), we use the first 120 examples in the Wikibio-GPT3 dataset (50% of it) to compute the empirical estimation of the emission probability. The model we use is gpt-3.5-turbo-0125. Then we transfer the estimated emission probabilities to the remaining examples in Wikibio-GPT3 and other datasets for performance evaluation. For Llama-3-8B-Instruction, we also use the same estimated emission probabilities. For the emission probability, we use first 120 examples of the Wikibio-GPT3 dataset as the validation set and estimate the emission probability on it using gpt-3.5-turbo-0125. The estimated emission probabilities are in Table 1 and will used on the other testing examples of Wikibio-GPT3 and other two datasets. Additionally, since we set the model confidence on child nodes generated by statement correction as 1.0 rather than probe the model’s confidence, we find setting an individual set of emission probabilities for those statements improves the performance. Therefore, we tune the emission probabilities for child nodes generated by statement correction on the validation set. The emission probabilities for those child nodes we use are $p(S_u=1\mid Z_u=1)=0.8$ and $p(S_u=1\mid Z_u=0)=0.2$ . During the belief tree construction process, we add several constraints to boost the efficiency. Specifically, we only apply statement decomposition to the root node, assuming the child nodes generated by the LLM is atomic statements that only contain one aspect of information. Also, we do not expand nodes generated by statement correction, as they are homogeneous to their parent node. #### Implementation of baselines We follow the default configuration of each baselines. For SelfCheckGPT, we sample 20 additional responses for hallucination detection. Similarly, we also sample 20 different answers for chain-of-thought prompting and aggregate them using self-consistency Wang et al. (2022). For Maieutic-Prompting, we use their prompts for the CREAK datasets to generate belief trees in our evaluation. For all methods, we use the scikit-learn package to compute the evaluation metrics including AUROC, AUC-PR, and F1 score. ### A.2 Additional Results | | AUROC | AUC-PR | F1 | Acc | | --- | --- | --- | --- | --- | | Prior Confidence | 67.7 | 63.2 | 71.6 | 68.1 | | MSP | 70.5 | 73.3 | 69.3 | 57.4 | | Perplexity | 66.0 | 63.4 | 70.0 | 59.3 | | CCP | 75.8 | 78.1 | 72.4 | 69.5 | | BTProp | 76.5 | 75.5 | 73.3 | 70.0 | Table 5: Performance of BTProp given different child node generation strategies. Since the datasets in the above experiments consist of text pre-generated by another model for hallucination detection, we also include texts generated by the hallucination detection model itself to further evaluate the effectiveness of our method. Specifically, we first generate biographical data using the prompt from Wikibio-GPT3 dataset with Llama-3-8b-instruct, following the “claim-level uncertainty quantification” setup of LM-Polygraph (Fadeeva et al., 2023). Then we use the same model to detect the hallucinations in its generations, where the ground-truth label for each claim is determined by FactScore (Min et al., 2023). We evaluate the performance of our method and additional baselines including CCP (Fadeeva et al., 2024), maximum sequence probability (MSP), and Perplexity. The ground-truth label for each claim is determined by FactScore. The results are summarized in Table 5. Similar to the observation in our main experiments, our method performs competitively against these state-of-the-art uncertainty quantification techniques. Specifically, our approach achieves the highest AUROC, F1 scores, and accuracy, demonstrating its ability to reliably detect hallucinations. Please note that our method is not originally designed for the “claim-level” hallucination detection, where the biography data are decomposed into atomic claims for fact-checking. Nevertheless, our method still achieves competitive performance. ### A.3 Examples of Belief Trees Figure 5: Belief tree example. We visualize several constructed belief trees as well as how our method leverage the inconsistencies in model’s beliefs for hallucination detection. We start with a simple example from the Wikibio-GPT3 dataset Manakul et al. (2023) shown in Figure 5. Starting from the root node about Adiele Afigbo, the first layer contains 3 child nodes generated by statement decomposition. However, despite the statement is wrong, the model (gpt-3.5-turbo-0125) assign a high confidence score to both the original statement and the child nodes decomposed from the root node. Then, our method further generate child nodes for node 1,2, and 3. We display the child nodes of node 3 in the figure, which is generated by statement correction. With the three different statements about the date of death of that person, we decrease the model’s confidence on node 3 and finally correct the model’s belief on the root node. Note that in statement correction, we directly set the confidence of the generated child nodes as 1.0. For this example, if we query the model’s confidence scores on node 4, 5, and 6, we will get confidence scores 0.97, 0.91, 0.88, respectively, which will still contradicts to their parent node. Figure 6: Belief tree example. Another example of the belief tree is shown in Figure 6, which mainly consists of child nodes generated by supportive and contradictory premises. At the root node, the model assigns a high confidence score to the statement about freezing point of water from the FELM dataset Chen et al. (2023b). After the statement decomposition, the inconsistency is triggered due to model’s low confidence on node 2. Furthermore, if we continuously generate premises for node 1, we get two additional child nodes (node 3 and 4) that are contradictory to their parent node. However, the model still assigns a high confidence to node 3 and 4. Within the belief propagation framework, the conditional probability of node 1 being true given the observations on node 3 and 4 will be decreased. Similarly, this effect will propagate to the root node and lead to a low posterior probability of the root node being true. <details> <summary>x8.png Details</summary>  ### Visual Description Icon/Small Image (1660x4) </details> Figure 7: Belief tree example. We then show one failure case of our method in Figure 7. The target statements comes from the FactCheckGPT dataset Wang et al. (2023). Our method first generation two supportive premises with high confidence. These two child nodes are indeed reasonable and can support the opinion that “Linux operating system has a widespread adoption.”, thus contradicting to the root statement and further decreasing the model’s confidence on the root node. We hypothesize that the original statement, “Linux adoption has limited compared to other operating systems like Windows and macOS” lacks sufficient specificity and could be interpreted from multiple perspectives. For example, it could refer to market share in personal computing versus cloud computing domains, making the ground-truth label nondeterministic. ### A.4 Inferring the Posterior Probability of $Z_0$ The posterior probability $p(Z_0=T\mid\{S_u\})$ can be computed by: | | $\displaystyle\frac{p(Z_0=T,\{S_u\})}{p(Z_0=T,\{S_u\})+p(Z_0=F,\{S_u \})},$ | | | --- | --- | --- | where $\{S_u\}=\{s_u\}$ refers to the set of all observed variables in the belief tree. Therefore, the key of the inference on the belief tree is to compute the two joint probabilities $p(Z_0=T,\{S_u\})$ and $p(S_1=F,\bar{\bm{X}}_1=\bar{\bm{x}}_1)$ . Following the conditional independence assumption in the hidden Markov model, the truthfulness (i.e,, states of hidden variables) of each node is determined by its parent node in the tree and the transition probability. Also, the model’s confidence (i.e., states of observed variables) on each node is determined by the state of the corresponding variable and the emission probability. With such an assumption, they can be decomposed as: $$ \displaystyle p(Z_0=T,\{S_u\}) \displaystyle=p(\{S_u\}\mid Z_0=T)*p(Z_0=T), \tag{2} $$ where $p(Z_0=T)$ is the prior probability of the statement being true (note that we set it to be 0.5 in this paper). $p(\{S_u\}\mid Z_0=T)$ is the conditional probability of all observed variables (exclude the root node) given the root node being true. Recall that we introduce a notation $β(z,u)=p(S_T(u)|Z_u=z)$ to represent such conditional probabilities, where $S_T(u)$ refers to the confidence scores (observed variables) of all nodes on the subtree rooted at node $u$ . Therefore, the conditional probability $p(\{S_u\}\mid Z_0=T)$ can be denoted as $β(T,0)$ (node $0 0$ , hidden variable $Z_0=T$ ). Without loss of generality, we discuss below how to compute $β(z,u)$ for an arbitrary node $u$ in the tree. The computation for $β(T,0)$ and $β(F,0)$ can be performed in exactly the same way. Specifically, we can further decompose $β(z,u)$ as: $$ \displaystyle p(S_T(u))\mid Z_u=z) \displaystyle=p(S_u\mid Z_u=z)· \displaystyle≤ft\{∏_v∈\bm{{C}(u)}p(S_T(v)\mid Z _u=z)\right\}. \tag{3} $$ The probability $p(S_T(v)\mid Z_u=z)$ can be decomposed as: $$ \displaystyle p(S_T(v)\mid Z_u=z) \displaystyle=∑_k∈\{T,F\}p(S_T(v)\mid Z_v=k)*p(Z_v=k \mid Z_u=z) \displaystyle=∑_k∈\{T,F\}β(v,k)*p(Z_v=k\mid Z_u=z). \tag{4} $$ Therefore, the conditional probability $β(z,u)=p(S_T(u))\mid Z_u=z)$ can be computed in a recursive manner: $$ \displaystyleβ(z,u)=\underbrace{p(S_u\mid Z_u=z)}_Emission Probability· \displaystyle≤ft\{∏_v∈\bm{{C}(u)}∑_k∈\{T,F\}β(v ,k)*\underbrace{p(Z_v=k\mid Z_u=z)}_Transition Probability\right\}. \tag{5} $$ There are two types of probabilities in the above equation. First, $p(S_u\mid Z_u=z)$ is the “emission probability" at node $u$ . Second, $p(Z_v=k\mid Z_u=z)$ is the “transition probability" from node $u$ to its child node $v$ . The recursive computation in Equation 5 will starts from the leaf nodes. When $v$ is the leaf node in the tree, $β(v,k)$ is actually $p(S_v\mid Z_v=k)$ , which is the emission probability at node $v$ . Such a process propagates from bottom to up. Finally, we can compute $β(t,0)$ as well as the joint probability $p(Z_0=T\mid\{S_u\})$ according to Equation 2. #### Beyond conditional independence: transition probability for statement decomposition. The above computation of the posterior probability is based on the assumption of conditional independence. Assume the node $u$ have two child nodes, $v_1$ and $v_2$ . Then we consider the transition probabilities from $u$ to $v_1$ and from $u$ to $v_2$ independently. However, this does not hold for the child nodes generated by statement decomposition, in which the truthfulness of the child nodes are influenced by their parent node simultaneously. If $u$ is true, then we know that both two child nodes are also true. In contrast, if $u$ is false, we can only infer that at least one of the child nodes are false. To handle such a case, we revise the probability computation in Equation 3 accordingly. Assume node $u$ has $m$ child nodes. The probability $β(z,u)=p(S_T(u))\mid Z_u=z)$ can be computed as: | | | $\displaystyle p(S_T(u))\mid Z_u=z)=p(S_u\mid Z_u=z)\tilde{ β}(u,z),where$ | | | --- | --- | --- | --- | which is the Equation 1 in the main paper. ### A.5 Summary of the Algorithm Algorithm 1 BTProp Algorithm 1: Input: Statement $u_0$ , maximum tree depth $d_max$ 2: Output: The posterior probability $p(Z_0=1\mid S_T(0))$ . 3: 4: Initialize the belief tree $T$ with root node $u_0$ 5: Initialize the leaf node set $N=\{u_0\}$ 6: while $N≠\oslash$ do $\triangleright$ Belief tree construction 7: Pop an element $u$ from $N$ 8: Generate child nodes $C(u)$ for $u$ 9: for Node $v∈C(u)$ do 10: Add $v$ to $T$ 11: Add $v$ to $N$ if $d_u<d_max$ 12: end for 13: end while 14: 15: function GetBeta ( $u$ , $z$ ) $\triangleright$ Compute $β(z,u)$ in Eq. 1 16: if $C(u)=\oslash$ then 17: return $p(s_u\mid Z_u=z)$ 18: end if 19: for $v∈C(u)$ do 20: $β(v,0)=\textsc{GetBeta}(v,0)$ 21: $β(v,1)=\textsc{GetBeta}(v,1)$ 22: end for 23: Compute $β(u,z)$ according to Eq. 1 24: return $β(u,z)$ 25: end function 26: 27: $β(1,0)$ = GetBeta ( $u=0$ , $z=1$ ) 28: $β(0,0)$ = GetBeta ( $u=0$ , $z=0$ ) 29: $p(Z_0=1\mid S_T(0))=β(1,0)/(β(1,0)+β(0,0))$ We summarize the pipeline of our method in Algorithm 1. The root node of the belief tree is the given statement $u_0$ . During the belief tree construction process, we maintain a set $N$ that contains all current leaf nodes of the belief tree. We recursively expand each leaf node $u∈N$ by its child nodes $C(u)$ and add these new child nodes to the belief tree if they are logically-connected to the current leaf node. These generated child nodes then become new leaf nodes. They will also be added to $N$ and be expanded until the maximum depth is reached. Given the constructed belief tree and the emission and transition probabilities estimated from a held-out dataset, the posterior probability $β(Z_0=1\mid S_T(0))$ can be then computed in a recursive manner, which will be further used to predict the truthfulness of the given statement. ### A.6 Submission Checklist FactCheckGPT is under Apache-2.0 license. FELM is under CC-BY-NC-SA-4.0 license, and Wikibio-GPT3 is under CC-BY-SA-3.0 license. We have cited them accordingly in the main paper. All of these datasets are in English. These datasets are created to evaluate the hallucination detection performance, and our usages are consistent with their intended use. We also manually check the datasets and confirm that they do not contains any information that names or uniquely identifies individual people or offensive content. We leverage ChatGPT to help develop some of the prompt we use in the experiments and improve the writing. ### A.7 Prompt In this section, we list all the prompt used in this paper, including belief tree construction, prior confidence estimation, and data preprocessing. #### Belief tree construction Given a statement from the LLM output, we use the following prompt gpt-3.5-turbo-0125 to perform statement decomposition, which is shown in Figure 10. In the instruction, we specify the requirements to extract check-worthy claims and provide the model with several examples. We also include an additional special example, where the given sentence is actually a subjective opinion, to prevent the model from decomposing sentences that are actually not check-worthy. We use a similar prompt for Llama-3-8b-Instruction as shown in Figure 11. To generate supportive and contradictory premises, we prompt the model as follows: If the model judge the given statement as true, then it needs to generate several explanations to its judgment, which form the supportive premises. In contrast, if the model believes the given statement is false, then it will generate explanations to its judgment, which form the contradictory premises. The prompts we used are listed in Figure 12 and Figure 13. We first prompt the model with the prompt in Figure 12. If the model judges the statement as true and generates supportive premises, then these premises will be returned and contradictory premises will not be generated. If the model judges the statement as false, then we prompt the model with the prompt in Figure 13 for contradictory premises. Although there might be better prompting strategies to generate such premises, finding the optimal prompts and prompting strategies is out of the scope of this paper. Therefore, we leave it for future work. Finally, to generate child nodes via statement correction, we adopt the following pipeline. First, we prompt the model using prompts listed in Figure 14 and Figure 15 to generate a question about the key pieces of information in the statement. After that, we feed the generated question to the LLM again without any other prompt to get its answer. Finally, we ask the model to revise the original statement according to the “background knowledge", which is the answer generated by itself. By doing this, we expect to better utilize the inconsistency across model’s beliefs for child node generation. The prompt for the last step is shown in Figure 16 To select the most appropriate strategies for each node, we use the following prompts to ask the LLM to output the most suitable strategies, which is displayed in Figure 17 and Figure 18. #### Prior confidence estimation When prompt gpt-3.5-turbo-0125 for the confidence score on the truthfulness of a statement, we use the following prompt: True or False? {target_statement}. For Llama-3 model, we find it sometimes refuse to judge the truthfulness and simply output it requires additional context. To enforce the model to output the confidence score, we change the prompt accordingly: ⬇ Are the following statements true or false? For each of the following statements, determine whether it is true or false. Provide a response of ’ True ’ if the statement is correct, or ’ False ’ if the statement is incorrect. Remember: - You ** do not ** have access to additional information or external data. - If verifying the statement requires such external data or context, regard it as false. - Directly output " True " or " False " without adding any markers. Figure 8: Prompt for confidence estimation (Llama-3-8B-Instruct). ⬇ ** Rewrite Texts for Clarity ** In this task, you will receive one paragraph and one target statement extracted from it. The target statement is context - dependent, which makes the statement hard for us to understand without context and check its truthfulness. Therefore, your task is to rewrite the statement to reduce context dependency. Specifically, - Pronoun resolution: Replacing pronouns like " this," " the," " that," " he," " she," and " they " with specific nouns or names they refer to in the original paragraph. You should always use the full names. - If the target statement only use the first / last name to refer to the main entity, replace the first / last name with the full name of the entity if available. Note: do not modify the semantics of the sentence. Do not add new information or your own descriptions into the statement. ** Input / Output Format ** The input will be provided with the format as below: Original paragraph: < the original text > Target statement: < the target statement needing rewrite > Format your output as: Output: < the target statement after rewrite > Figure 9: Prompt for data preprocessing. ⬇ ** Fact - Checking Claims Extraction:** ** Objective:** Analyze the provided statement to extract and list all distinct factual claims that require verification. Each listed claim should be verifiable and not overlap in content with other claims listed. ** Instructions:** 1. ** Identify Factual Claims **: - Identify parts of the statement that assert specific, verifiable facts, including: - Statistical data and measurements. - Historical dates and events or other information. - Direct assertions about real - world phenomena, entities, events, statistics - Conceptual understandings and theories. - In every claim, alway use the ** full names ** when referring to any concept, person, or entity. ** Avoiding the use of pronouns or indirect references ** that require contextual understanding. 2. List each verifiable claim separately. Ensure that each claim is distinct and there is no overlap in the factual content between any two claims. - If a single claim is repeated in different words, list it only once to avoid redundancy. 3. ** Output:** - If there are multiple check - worthy claims, list each one clearly and separately. - If there is only one check - worthy claim, output just that one claim. - If no part of the statement contains verifiable facts (e. g., purely subjective opinions, hypothetical scenarios), output the following message: " Claim 1: No check - worthy claims available." ** Output Format **: Your output should be organized as follows: Claim 1: < the first claim > Claim 2: < the second claim > Claim 3: < the third claim (if necessary)> ... ** Examples:** Statement: According to recent data, China has surpassed the United States in terms of GDP when measured using Purchasing Power Parity (PPP), and India is projected to overtake China by 2030. Claim 1: China has surpassed the United States in terms of GDP when measured using Purchasing Power Parity (PPP). Claim 2: India is projected to overtake China in terms of GDP by 2030." Statement: The world ’ s largest desert is Antarctica, and it is larger than the Sahara. Claim 1: The world ’ s largest desert is Antarctica. Claim 2: Antarctica is larger than the Sahara. Statement: I think pizza is the best food ever! Claim 1: No check - worthy claims available. Statement: The software ’ Photoshop ’ was released by Adobe Systems in 1988. Claim 1: The software ’ Photoshop ’ was released by Adobe Systems in 1988. Figure 10: The prompt for statement decomposition using gpt-3.5-turbo-0125. ⬇ ** Fact - Checking Claims Extraction:** ** Objective:** Analyze the provided statement to extract and list all distinct factual claims that require verification. Each listed claim should be verifiable and not overlap in content with other claims listed. ** Instructions:** 1. Identify parts of the statement that assert specific, verifiable facts. In every claim, alway use the ** full names ** when referring to any concept, person, or entity. ** Avoiding the use of pronouns or indirect references ** that require contextual understanding. 2. List each verifiable claim separately. Ensure that each claim is distinct and there is no overlap in the factual content between any two claims. 3. Exclude any statements that are purely hypothetical, assume theoretical scenarios, or are speculative in nature. These do not contain verifiable factual claims. For example, statements involving assumptions (" Let ’ s assume..."), theoretical discussions (" consider whether..."), or purely speculative scenarios should not be considered as containing verifiable claims. 4. If the statement does not contain any verifiable facts, output the following message: " Claim 1: No check - worthy claims available." ** Output Format **: Your output should be organized as follows: Claim 1: < the first claim > Claim 2: < the second claim > Claim 3: < the third claim (if necessary)> ... ** Examples:** Statement: According to recent data, China has surpassed the United States in terms of GDP when measured using Purchasing Power Parity (PPP), and India is projected to overtake China by 2030. Claim 1: China has surpassed the United States in terms of GDP when measured using Purchasing Power Parity (PPP). Claim 2: India is projected to overtake China in terms of GDP by 2030." Statement: The world ’ s largest desert is Antarctica, and it is larger than the Sahara. Claim 1: The world ’ s largest desert is Antarctica. Claim 2: Antarctica is larger than the Sahara. Statement: I think pizza is the best food ever! Claim 1: No check - worthy claims available. Statement: Let ’ s assume there is a highest prime number and consider its implications on number theory. Claim 1: No check - worthy claims available. Figure 11: The prompt for statement decomposition using Llama-3-8b-Instruct. ⬇ ** Finding Supportive Premises ** Is the following statement true or false? If it is true, list several supportive premises for it. ** Important Rules:** 1. Each premise should be clearly stated and directly relevant to the target statement. Avoid ambiguity and ensure that the connection to the target statement is evident 2. Do not use pronouns in generated premises. Ensure each premise can be understood clearly without any context. For each generated premise, you should always use the full name of each person, event, object, etc. ** Input / Output Format **: Your output should be organized as follows. Judgement: < True or False > Premise 1: < the first premise > Premise 2: < the second premise > ... In contrast, if the statement is false, you simly output: Judgement: False Premise 1: No supportive premises applicable. ** Examples:** Target statement: Renewable energy sources will lead to a decrease in global greenhouse gas emissions. Judgement: True Premise 1: Renewable energy sources produce electricity without emitting carbon dioxide. Premise 2: Increasing the adoption of renewable energy reduces reliance on fossil fuels, which are the primary source of industrial carbon dioxide emissions. Target statement: Eating carrots improves night vision. Judgement: False Premise 1: No supportive premises applicable. Statement: Historical literacy enhances a society ’ s ability to make informed decisions. Judgement: True Premise 1: Understanding historical events provides context for current issues, enabling citizens to make decisions that consider past outcomes and lessons. Premise 2: Historical literacy fosters critical thinking skills, which are crucial in analyzing information and making reasoned decisions. Premise 3: Societies with high levels of historical awareness can recognize and avoid the repetition of past mistakes. Figure 12: The prompt for generation of supportive premises. ⬇ ** Finding Contradictory Premises ** Is the following statement true or false? If it is false, list several contradictory premises for it. ** Important Rules:** 1. Each premise should be clearly stated and directly relevant to the target statement. Avoid ambiguity and ensure that the connection to the target statement is evident 2. Do not use pronouns in generated premises. Ensure each premise can be understood clearly without any context. For each generated premise, you should always use the full name of each person, event, object, etc. ** Input / Output Format **: Your output should be organized as follows. Judgement: < True or False > Premise 1: < the first premise > Premise 2: < the second premise > ... In contrast, if the statement is false, you simply output: Judgement: True Premise 1: No contradictory premises applicable. ** Examples:** Target statement: Renewable energy sources will lead to a decrease in global greenhouse gas emissions. Judgement: True Premise 1: No contradictory premises applicable. Target statement: Eating carrots improves night vision. Judgement: False Premise 1: The belief that eating carrots improves night vision stems from World War II propaganda, not from scientific evidence. Premise 2: While carrots are rich in vitamin A, which is necessary for maintaining healthy vision, they do not enhance night vision beyond normal levels. Statement: The introduction of invasive species does not impact native biodiversity. Judgement: False Premise 1: Invasive species often compete with native species for resources, leading to a decline in native populations. Premise 2: Studies show that invasive species can alter the natural habitats of native species, negatively affecting their survival rates. Premise 3: The introduction of the invasive zebra mussel in North American waterways has led to significant declines in the populations of native mussels. Figure 13: The prompt for generation of contradictory premises. ⬇ Given the following claim, your tasks include: 1. Identify the key pieces of information critical for fact - checking to determine its truthfulness. 2. Create a masked version of the claim by masking these key pieces of information 3. Generate a question asking for the key pieces of information Rules: 1. Do not mask the grammatical subject of the sentence -- the actor, entity, or object that performs the action in the sentence ’ s main clause. Also, following the format of the below examples in your output. ** Examples:** Statement: Bitcoin was created in 2009 by an anonymous entity known as Satoshi Nakamoto. Masked statement: Bitcoin was created in 2009 by an anonymous entity known as [which person]. Question: Who created Bitcoin in 2009? Statement: The iPhone was first released by Apple in 2007. Masked statement: The iPhone was first released by Apple in [what year]. Question: Was the iPhone first released by Apple in 2007? Statement: The speed of light in a vacuum is approximately 299,792 kilometers per second. Masked statement: ’ Romeo and Juliet ’ was written by Shakespeare in [what time period]. Question: What is the speed of light in a certain medium? Statement: " Attention Is All You Need " is a paper written by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin. Masked statement: " Attention Is All You Need " is a paper written by [whom] Question: Who was / were the authors of the paper " Attention Is All You Need "? Statement: The Great Wall of China is visible from space. Masked statement: The Great Wall of China is [visible or invisible] from space. Question: Is the Great Wall of China visible from space? Statement: The headquarters of the United Nations is located in New York City Masked statement: The headquarters of the United Nations is located in [which city]. Question: Where is the headquarters of the United Nations located? Figure 14: Step 1 in statement correction: question generation. This prompt is used for gpt-3.5-turbo-0125. ⬇ Given a statement, your task is to identify a general question that can be used to check the truthfulness of the statement. The question should directly address the claim to confirm or refute it without seeking additional detailed information. ** Examples:** Statement: Bitcoin was created in 2009 by an anonymous entity known as Satoshi Nakamoto. Question: Who created Bitcoin in 2009? Statement: The iPhone was first released by Apple in 2007. Question: When was iPhone first relased by Apple? Statement: ’ Romeo and Juliet ’ was written by Shakespeare in the late 16 th century. Question: When was ’ Romeo and Juliet ’ was written by Shakespeare? Statement: " Attention Is All You Need " is a paper written by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin. Question: Who was / were the authors of the paper " Attention Is All You Need "? Statement: The Eiffel Tower is located in Paris Question: Where is the Eiffel Tower located? Statement: Albert Einstein was a physicist. Question: What profession was Albert Einstein? Figure 15: Step 1 in statement correction: question generation. This prompt is used for Llama-3-8b-Instruct. We adjust the prompt since the masked statement is not used in the prompt for gpt-3.5-turbo-0125. ⬇ ** Background Knowledge **: {the model answer in step 2} Leverage the above provided knowledge and your own knowledge to review the correctness of following statement: ** Statement **: {statement} Instruction: - If the statement is correct, output it unchanged. - If the statement is ** not mentioned in the background knowledge and its correctness cannot be determined **, you should also directly output the statement ** unchanged **. - If the statement is wrong, revise only the parts of the statement that are incorrect, to align with the background knowledge. Do not add any additional sentences or details. ** Output Format:** Format your output as: Revised Answer: < Display the original statement if it is correct or not mentioned in the background knowledge; display the revised statement if it is inaccurate > Figure 16: Step 3 in statement correction: statement revision. ⬇ ** Task: Choose the Best Strategy for Premise Generation ** We need to generate several premises for a given target statement. These premises could either support or contradict the target statement. Particularly, we have 2 techniques for premise generation: 1. Logical Relationships: This involves creating premises based on entailment or contradiction. You can generate premises that either support or contradict the target statement. 2. Statement Perturbation: Create variations of the statement by altering key details to form contradictory premises. Given these techniques, your task is to select the most suitable technique given a particular statement. Follow the guidelines below to select the most suitable technique. ** Important Guidelines **: 1. Prioritize logical relationships. The logical relation strategy is broadly applicable, as long as it is straightforward to find supportive or contradictory premises. 2. If a statement contains particular names, numbers, timestamps, or other conditions that can be varied to generate contradictory premises, consider statement perturbation. 3. If both two methods are applicable, select them together and output " both ". ** Output Format:** Your selection should be one of " Logical Relation ", " Statement Perturbation ", and " both ". Target statement: Eating a balanced diet improves overall health. Output: Logical Relation Target statement: ’ War and Peace ’ was a book written by Leo Tolstoy. Output: Statement Perturbation Target statement: Water boils at 100 degrees Celsius at sea level Output: both Target statement: Mount Everest is 8,848 meters tall. Output: Statement Perturbation Target statement: Artificial intelligence will replace most human jobs in the future Output: Logical Relation Figure 17: Prompt for strategy selection (gpt-3.5-turbo-0125). ⬇ ** Instruction for Choosing the Best Strategy for Premise Generation ** When tasked with generating premises for a given target statement, the choice of strategy - Logical Relationships or Statement Perturbation - should be determined based on the nature of the statement and the desired objectives. Use the following guidelines to route the choice: ### When to Use Logical Relationships Opt for Logical Relationships when: - ** The Statement is Abstract or Principled **: Ideal for statements that explore broad principles, ethics, or abstract concepts. This method helps in drawing deep logical entailments or contradictions. - ** Complex Relationships or Conditions **: When the statement involves complex logical or conditional relationships, using this method clarifies or challenges these intricacies. - ** Requirement for Detailed Analysis **: For statements needing precise and formal argumentation, especially in academic or technical discussions. ### When to Use Statement Perturbation Choose Statement Perturbation when: - ** The Statement is Specific and Concrete **: Best for statements with explicit details like scenarios, dates, locations, names, numbers or timestamps. Altering these elements generates varied premises. - ** Exploration of Counterfactuals or Hypotheticals **: Useful for creating imaginative or scenario - based premises by modifying key details to envision different outcomes. - ** Sensitivity to Detail Changes **: When minor modifications in the statement can significantly alter its implications or truth value. ** Output Format:** Your selection should be one of " Logical Relation ", " Statement Perturbation ", and " both ". Here " both " means both two methods are applicable. ** Examples:** Target statement: Eating a balanced diet improves overall health. Output: Logical Relation Target statement: Countries with higher investment in education consistently rank higher in global innovation indexes Output: Logical Relation Target statement: ’ War and Peace ’ was a book written by Leo Tolstoy. Output: Statement Perturbation Target statement: Water boils at 100 degrees Celsius at sea level Output: both Target statement: Mount Everest is 8,848 meters tall. Output: Statement Perturbation Target statement: The bridge will remain intact even if 75 cars drive on it simultaneously if the cars are lightweight Output: Statement Perturbation Target statement: Artificial intelligence will replace most human jobs in the future Output: Logical Relation Figure 18: Prompt for strategy selection (Llama-3-8b-Instruction).