2402.03744

# INSIDE: LLMs’ Internal States Retain the Power of Hallucination Detection **Authors**: - Zhihang Fu, Jieping Ye (Alibaba Cloud  Zhejiang University) > Corresponding Author Abstract Knowledge hallucination have raised widespread concerns for the security and reliability of deployed LLMs. Previous efforts in detecting hallucinations have been employed at logit-level uncertainty estimation or language-level self-consistency evaluation, where the semantic information is inevitably lost during the token-decoding procedure. Thus, we propose to explore the dense semantic information retained within LLMs’ IN ternal S tates for halluc I nation DE tection (INSIDE). In particular, a simple yet effective EigenScore metric is proposed to better evaluate responses’ self-consistency, which exploits the eigenvalues of responses’ covariance matrix to measure the semantic consistency/diversity in the dense embedding space. Furthermore, from the perspective of self-consistent hallucination detection, a test time feature clipping approach is explored to truncate extreme activations in the internal states, which reduces overconfident generations and potentially benefits the detection of overconfident hallucinations. Extensive experiments and ablation studies are performed on several popular LLMs and question-answering (QA) benchmarks, showing the effectiveness of our proposal. Code is available at https://github.com/alibaba/eigenscore 1 Introduction Large Language Models (LLMs) have recently achieved a milestone breakthrough and demonstrated impressive abilities in various applications (Ouyang et al., 2022; OpenAI, 2023). However, it has been widely observed that even the state-of-the-art LLMs often make factually incorrect or nonsense generations (Cohen et al., 2023; Ren et al., 2022; Kuhn et al., 2022), which is also known as knowledge hallucination (Ji et al., 2023). The potentially unreliable generations make it risky to deploy LLMs in practical scenarios. Therefore, hallucination detection, that is, accurately detecting and rejecting responses when hallucinations occur in LLMs, has attracted more and more attention from the academic community (Azaria & Mitchell, 2023; Ren et al., 2022; Kuhn et al., 2022). The token-level uncertainty estimation (e.g., predictive confidence or entropy) has shown its efficacy in hallucination detection on conventional NLP tasks (Malinin & Gales, 2020; Huang et al., 2023). However, how to derive the sentence-level uncertainty from the token-level remains a challenge, especially for modern auto-regressive LLMs whose response contents are generally diverse and sophisticated (Malinin & Gales, 2020; Kuhn et al., 2022; Duan et al., 2023). Thus, to avoid complicated token-to-sentence uncertainty derivation, researchers propose to evaluate the sentence uncertainty by the output languages directly (Kadavath et al., 2022; Yin et al., 2023; Zhou et al., 2023). Among the recent advancements, prompting LLMs to generate multiple responses to the same question and evaluating the self-consistency of those responses has been proven effective in hallucination detection (Wang et al., 2022; Shi et al., 2022). However, such a post-hoc semantic measurement on decoded language sentences is inferior to precisely modeling the logical consistency/divergence Manakul et al. (2023); Zhang et al. (2023). Hence, instead of logit-level or language-level uncertainty estimation, this paper proposes to leverage the internal states of LLMs to conduct hallucination detection. The motivation is intuitive: LLMs preserve the highly-concentrated semantic information of the entire sentence within their internal states (Azaria & Mitchell, 2023), allowing for the direct detection of hallucinated responses in the sentence embedding space. In particular, with the generalized framework of IN ternal S tates for halluc I nation DE tection (INSIDE), this paper performs hallucination detection from two perspectives. First, skipping secondary semantic extraction via extra models, we directly measure the self-consistency/divergence of the output sentences using internal states of LLMs. In order to explore semantic consistency in the embedding space, Section 3.1 introduces an EigenScore metric regarding the eigenvalues of sentence embeddings’ covariance matrix. Second, to handle the self-consistent (overconfident) hallucinations, we propose to rectify abnormal activations of the internal states. Specifically, Section 3.2 develops a feature clipping approach to truncate extreme features, which tends to prevent overconfident generations during the auto-regressive procedure. In Section 4, the effectiveness of our method is validated through extensive experiments on several well-established QA benchmarks. The main contributions of our work are as follows: - We propose a generalized INSIDE framework that leverages the internal states of LLMs to perform hallucination detection. - We develop an EigenScore metric to measure the semantic consistency in the embedding space, and demonstrate that the proposed EigenScore represents the differential entropy in the sentence embedding space. - A test time feature clipping approach is introduced to truncate extreme activations in the feature space, which implicitly reduces overconfident generations and helps identify the overconfident hallucinations. - We achieve state-of-the-art hallucination detection performance on several QA benchmarks, and conduct extensive ablation studies to verify the efficacy of our method. 2 Background on Hallucination Detection In this work, we mainly focus on the knowledge hallucination detection of natural language generation based on LLMs, especially for Q&A task (Reddy et al., 2019; Kwiatkowski et al., 2019). Given an input context $\bm{x}$ , a typical LLM (Zhang et al., 2022; Touvron et al., 2023a) parameterized with $\bm{\theta}$ is able to generate output sequences in autoregressive manner $y_{t}=f(\bm{x},y_{1},y_{2},·s,y_{t-1}|\bm{\theta})$ , where $\bm{y}=[y_{1},y_{2},·s,y_{T}]$ denotes the output sequence and $y_{t}$ denotes the t- $th$ output token. We denote $p(y_{t}|y_{<t},\bm{x})$ the Maximum Softmax Probability (MSP) of $t$ -th token. For a traditional classification model, the MSP measures the confidence level of the classification result and has been widely used as an uncertainty measure of predictions (Hendrycks & Gimpel, 2016). Therefore, for sequence generation task, a straightforward sequence uncertainty can be defined as the joint probability of different tokens, which is known as Perplexity (Ren et al., 2022), $$ P(\bm{y}|\bm{x},\bm{\theta})=-\frac{1}{T}\log\prod_{t}p(y_{t}|y_{<t},\bm{x})=-% \frac{1}{T}\sum_{t}\log p(y_{t}|y_{<t},\bm{x}) \tag{1} $$ As shorter sequences generally have lower perplexity, the length of the output sequence $T$ is utilized to normalize the joint probability. Since different tokens contribute differently to the semantics of the sentence (Raj et al., 2023; Duan et al., 2023), the perplexity defined by averaging token-level uncertainty cannot effectively capture the uncertainty of the entire sequence. It has been demonstrated that utilizing multiple generations for one input is beneficial to estimate the sequence-level uncertainty (Malinin & Gales, 2020; Kuhn et al., 2022; Manakul et al., 2023). We denote $\mathcal{Y}=[\bm{y}^{1},\bm{y}^{2},·s,\bm{y}^{K}]$ as $K$ generated responses for input context $\bm{x}$ . For a given LLM, multiple responses could be easily obtained by the top-p/top-k sampling strategy during inference time (Touvron et al., 2023a; Kadavath et al., 2022). In Malinin & Gales (2020), the Length Normalized Entropy is proposed to measure the sequence-level uncertainty by making use of multiple generations, which is defined as $$ H(\mathcal{Y}|\bm{x},\bm{\theta})=-\mathbb{E}_{\bm{y}\in\mathcal{Y}}\frac{1}{T% _{\bm{y}}}\sum_{t}\log p(y_{t}|y_{<t},\bm{x}) \tag{2} $$ When a model is uncertain about its response, it generates hallucination context, resulting in an answer distribution with a high entropy (Kadavath et al., 2022). It has been shown that the length-normalized entropy performs better than the non-normalized one (Lin et al., 2023). In addition to the predictive uncertainty or entropy, the semantic consistency (Lin et al., 2023; Raj et al., 2023) among multiple responses has also been widely explored to measure the hallucination degree of LLMs, which hypothesis that the LLMs are expected to generate similar outputs if they know the input context and they are sure about the answers (Wang et al., 2022; Manakul et al., 2023). An intuitive semantic consistency metric is Lexical Similarity (Lin et al., 2022; 2023), which explores the average similarity across multiple answers as consistency measure $$ S(\mathcal{Y}|\bm{x},\bm{\theta})=\frac{1}{C}\sum_{i=1}^{K}\sum_{j=i+1}^{K}sim% (\bm{y}^{i},\bm{y}^{j}) \tag{3} $$ where $C=K·(K-1)/2$ and $sim(·,·)$ is the similarity defined by Rouge-L Lin (2004). 3 Method <details> <summary>x1.png Details</summary>  ### Visual Description ## Flowchart: Question-Answering System with Language Model ### Overview The image depicts a technical workflow for a question-answering system using a language model (LLM). It illustrates how input tokens are processed through a decoder architecture, generates candidate answer embeddings, and uses an eigenvector-based scoring mechanism to select the final output. The system includes confidence-based response validation. ### Components/Axes 1. **Input Section** (Leftmost block): - Text: "On what date in 1969 did Neil Armstrong first set foot on the Moon?" - Color coding: - Yellow: Token Embedding - Red: Current Token Embedding - Pink: Output Logit 2. **LLM Processing Block** (Central): - Components: - FC Layer (Feature Clip) - Decoder - Input Tokens: Represented as vertical bars with color gradients - Output: Three answer embeddings (purple, orange, yellow) labeled "Embedding of answer 1", "Embedding of answer 2", ..., "Embedding of answer K" 3. **Eigenvector Processing** (Middle-right): - Input: Matrix of answer embeddings - Output: Eigenvector with directional arrows (red for positive, blue for negative) 4. **Output Section** (Rightmost): - Decision diamond: "High EigenScore?" - Two possible outputs: - "The answer is 20th July." - "Sorry we don't support answer for this question." ### Detailed Analysis - **Token Processing**: Input tokens are color-coded (yellow for embeddings, red for current token) and fed into the LLM's decoder architecture. - **Answer Generation**: The decoder produces K candidate answer embeddings, visualized as horizontal bars with gradient colors. - **Eigenvector Analysis**: The eigenvector component processes embeddings through a matrix operation, with directional arrows indicating vector relationships. - **Confidence Threshold**: A binary decision node evaluates the EigenScore to determine output validity. ### Key Observations 1. The system uses a confidence threshold (EigenScore) to filter unsupported answers. 2. Color coding distinguishes different processing stages: - Yellow/Red: Input token representations - Purple/Orange/Yellow: Answer embeddings - Pink: Output logits 3. The eigenvector component suggests a mathematical approach to answer selection. 4. The flowchart implies a probabilistic or vector-based similarity matching mechanism. ### Interpretation This system demonstrates a hybrid approach combining: 1. **Neural Language Modeling**: For answer generation through token embeddings and decoder architecture 2. **Linear Algebra**: Using eigenvectors to analyze answer embeddings 3. **Confidence Scoring**: Implementing a threshold mechanism for response validation The eigenvector-based scoring suggests the system measures answer relevance through vector space relationships, potentially identifying the most semantically similar answer to the question. The confidence threshold indicates an awareness of answer reliability, preventing responses to unsupported queries. The color-coded visualization aids in understanding the multi-stage processing pipeline from raw input to final output. </details> Figure 1: Illustration of our proposed hallucination detection pipeline. During inference time, for a given question, the extreme features in the penultimate layer are truncated and the EigenScore is computed based on the sentence embeddings across multiple responses. In this section, we introduce the details of our proposed INSIDE framework for hallucination detection. The whole pipeline is illustrated as Fig. 1. In section 3.1, we demonstrate a simple but effective EigenScore metric by exploring sentence-level semantics in the internal states of LLMs. In section 3.2, a test-time feature clipping approach is introduced to effectively alleviate the issue of overconfident generation, thereby aiding in the identification of self-consistent hallucinations 3.1 Hallucination Detection by EigenScore The existing uncertainty or consistency based hallucination detection metrics are exploited in the logit or language space, which neglect the dense semantic information that is retained within the internal states of LLMs. To better exploit the dense semantic information, we propose to measure the semantic divergence in the sentence embedding space. For the $t$ -th output token $y_{t}$ , we denote the hidden embedding in the $l$ -th layer as $\bm{h}^{l}_{t}∈\mathbb{R}^{d}$ , where $d$ is the dimension of the hidden embedding ( $d=4096$ for LLaMA-7B and $d=5120$ for LLaMA-13B). According to Ren et al. (2022); Azaria & Mitchell (2023), the sentence embedding can be obtained by averaging the token embedding $\bm{z}=\frac{1}{T}\sum_{t=1}^{T}\bm{h}_{t}$ , or taking the last token embedding as sentence embedding $\bm{z}=\bm{h}_{T}$ . In our main experiments, we use the embedding of the last token in the middle layer as the sentence embedding, as it effectively captures the sentence semantic (Azaria & Mitchell, 2023). The comparison results of using different sentence embeddings are demonstrated in the ablation studies 4.3. For $K$ generated sequences, the covariance matrix of $K$ sentence embeddings can be computed as $$ \bm{\Sigma}=\mathbf{Z}^{\top}\cdot\mathbf{J}_{d}\cdot\mathbf{Z} \tag{4} $$ where $\bm{\Sigma}∈\mathbb{R}^{K× K}$ represents the covariance matrix that captures the relationship between different sentences in the embedding space, $\mathbf{Z}=[\bm{z}_{1},\bm{z}_{2},·s,\bm{z}_{K}]∈\mathbb{R}^{d× K}$ represents the embedding matrix of different sentences, $\mathbf{J}_{d}=\bm{I}_{d}-\frac{1}{d}\mathbf{1}_{d}\mathbf{1}_{d}^{→p}$ is the centering matrix and $\mathbf{1}_{d}∈\mathbb{R}^{d}$ is the all-one column vector. Then, the proposed EigenScore can be defined as the logarithm determinant (LogDet) of the covariance matrix, $$ E(\mathcal{Y}|\bm{x},\bm{\theta})=\frac{1}{K}\log\text{det}(\bm{\Sigma}+\alpha% \cdot\mathbf{I}_{K}) \tag{5} $$ Here, $\text{det}(\mathbf{X})$ represents the determinant of matrix $\mathbf{X}$ , and a small regularization term $\alpha·\mathbf{I}_{K}$ is added to the covariance matrix to explicitly make it full rank. Since the matrix determinant can be obtained by solving the eigenvalues, the EigenScore can be computed as $$ E(\mathcal{Y}|\bm{x},\bm{\theta})=\frac{1}{K}\log(\prod_{i}\lambda_{i})=\frac{% 1}{K}\sum_{i}^{K}\log(\lambda_{i}) \tag{6} $$ where $\lambda=\{\lambda_{1},\lambda_{2},·s,\lambda_{K}\}$ denotes the eigenvalues of the regularized covariance matrix $\bm{\Sigma}+\alpha·\mathbf{I}$ , which can be solved by Singular Value Decomposition (SVD). Eq. 6 shows that the hallucination degree of LLM’s generation can be measured by the average logarithm of the eigenvalues. The conclusion is intuitive, as the eigenvalues of covariance matrix capture the divergence and correlation relationship between embeddings of different sentences. When the LLM is confident to the answers and $K$ generations have similar semantic, the sentence embeddings will be highly correlated and most eigenvalues will be close to 0. On the contrary, when the LLM is indecisive and hallucinating contents, the model will generate multiple sentences with diverse semantics leading to more significant eigenvalues. The following remark is also provided to explain why the proposed EigenScore is a good measure of knowledge hallucination. Remark 1. LogDet of covariance matrix represents the differential entropy in the sentence embedding space. Differential Entropy is the natural extension of discrete Shannon Entropy $H_{e}(X)=-\sum_{X}-p(x)\log p(x)$ . The differential entropy $H_{de}(X)$ in continuous space can be defined by replacing the probability function with its density function $f(x)$ and integrating over $x$ , i.e., $H_{de}(X)=-∈t_{x}f(x)\log f(x)dx$ . In principle (Zhouyin & Liu, 2021), for a multivariate Gaussian distribution $X\sim N(\bm{\mu},\mathbf{\Sigma})$ , the differential entropy can be represented as $$ H_{de}(X)=\frac{1}{2}\log\text{det}(\mathbf{\Sigma})+\frac{d}{2}(\log 2\pi+1)=% \frac{1}{2}\sum_{i=1}^{d}\log\lambda_{i}+C \tag{7} $$ where $d$ is the dimension of variables and $C$ is a constant. Therefore, the differential entropy is determined by the eigenvalues (LogDet) of the covariance matrix. According to Remark 1, the proposed EigenScore defined by Eq. 6 represents the differential entropy in the sentence embedding space, which offers valuable insight into using EigenScore as a semantic divergence measure. Compared to existing uncertainty or consistency metrics that obtained in logit or language space (Malinin & Gales, 2020; Huang et al., 2023; Lin et al., 2022), the advantages of EigenScore are: (1) It captures the semantic divergence (entropy) in the dense embedding space, which is expected to retain highly-concentrated semantic information compared to logits or languages (Reimers & Gurevych, 2019). (2) Representing semantic divergence in embedding space can effectively solve the semantic equivalence (linguistic invariances) problem (Kuhn et al., 2022) in natural language space. (3) Fine-grained semantic relationship among different responses can be exploited by using eigenvalues of covariance matrix. Therefore, through the exploration of dense semantic information in the internal states, the EigenScore is expected to outperform existing uncertainty and consistency metrics, resulting in improved hallucination detection performance. <details> <summary>x2.png Details</summary>  ### Visual Description ## Line Chart: Neuron Activation Distribution ### Overview The image depicts a line chart titled "Neuron Activation Distribution," visualizing neuron activation values across a range of neuron indexes. The chart features a single cyan-colored line representing activation data, with fluctuations in activation levels across the neuron indexes. ### Components/Axes - **Title**: "Neuron Activation Distribution" (centered at the top). - **X-Axis**: Labeled "Neuron Indexes," scaled from 0 to 4000 in increments of 1000. - **Y-Axis**: Labeled "Neuron Activations," scaled from -30 to 30 in increments of 10. - **Legend**: No explicit legend is visible in the image. The cyan line is the sole data series. - **Line**: A single cyan line traversing the chart, with no markers or annotations. ### Detailed Analysis - **Neuron Indexes (X-Axis)**: The horizontal axis spans 0 to 4000, with evenly spaced tick marks at 0, 1000, 2000, 3000, and 4000. - **Neuron Activations (Y-Axis)**: The vertical axis ranges from -30 to 30, with values distributed symmetrically around zero. - **Line Behavior**: The cyan line exhibits significant variability, with peaks reaching approximately +25 and troughs dipping to around -25. The line oscillates irregularly, showing no consistent upward or downward trend. Notable spikes occur near neuron indexes 500, 1500, 2500, and 3500, while deeper troughs are observed near 1000 and 3000. ### Key Observations 1. **High Variability**: Activation values fluctuate widely, with no clear pattern or correlation between adjacent neuron indexes. 2. **Outliers**: Peaks and troughs exceed the typical range of ±10, suggesting sporadic high or low activation events. 3. **Symmetry**: The distribution of activations is roughly symmetric around zero, indicating balanced positive and negative activation events. ### Interpretation The chart suggests a heterogeneous distribution of neuron activations, with individual neurons exhibiting diverse activation levels. The lack of a systematic trend implies that neuron activity is either uncorrelated or influenced by external factors not represented in the data. The outliers (extreme peaks/troughs) could represent neurons with heightened or suppressed activity, potentially critical for specific neural processes. The symmetry around zero may indicate a baseline activation state, with deviations reflecting dynamic responses. This distribution could be relevant to studies of neural plasticity, signal processing, or pathological conditions involving abnormal neural firing patterns. </details> (a) Neuron Activation <details> <summary>x3.png Details</summary>  ### Visual Description ## Histogram: Neuron Activation Distribution ### Overview The image displays a histogram representing the distribution of neuron activation values. The chart uses a teal-colored bar plot to visualize the density of normalized feature values across a range from -0.75 to 1.00. The distribution peaks near the center (0.0) and tapers off symmetrically toward both ends. ### Components/Axes - **Title**: "Neuron Activation Distribution" (top-center, black text). - **X-axis**: Labeled "Normalized Features," with tick marks at intervals of 0.25 (-0.75, -0.50, -0.25, 0.00, 0.25, 0.50, 0.75, 1.00). - **Y-axis**: Labeled "Density," with tick marks at intervals of 0.5 (0.0, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0). - **Legend**: Positioned at the top-right, indicating the teal color corresponds to "Neuron Activation Distribution." ### Detailed Analysis - **Data Distribution**: The histogram shows a bell-shaped curve, with the highest density (~3.0) centered at 0.0. Density decreases linearly toward both extremes, reaching near-zero values at -0.75 and 1.00. - **Key Data Points**: - At -0.75: Density ≈ 0.0. - At -0.50: Density ≈ 0.1. - At -0.25: Density ≈ 0.5. - At 0.00: Density ≈ 3.0 (peak). - At 0.25: Density ≈ 2.5. - At 0.50: Density ≈ 1.5. - At 0.75: Density ≈ 0.3. - At 1.00: Density ≈ 0.0. - **Bar Structure**: Bars are evenly spaced along the x-axis, with heights proportional to density values. No gaps or missing data points are observed. ### Key Observations 1. **Symmetry**: The distribution is nearly symmetric around 0.0, suggesting balanced activation patterns. 2. **Peak Dominance**: The majority of neuron activations cluster tightly around the mean (0.0), indicating low variability. 3. **Tapered Extremes**: Very few neurons exhibit extreme activation values (beyond ±0.5). ### Interpretation The data suggests that neuron activations in this dataset are predominantly centered around a normalized value of 0.0, with a Gaussian-like distribution. This symmetry implies that neural responses are balanced, with minimal skew toward hyper- or hypo-activation states. The sharp decline in density at the extremes (-0.75 and 1.00) indicates that extreme activation events are rare, which could reflect stable neural processing or effective regulatory mechanisms. The absence of outliers reinforces the reliability of the dataset for modeling purposes. The teal color coding aligns with the legend, confirming consistency in data representation. </details> (b) Feature Distribution Figure 2: Illustration of activation distributions in the penultimate layer of LLaMA-7B. (a) Activation distribution in the penultimate layer for a randomly sampled token. (b) Activation distribution for a randomly sampled neuron activation of numerous tokens. 3.2 Test Time Feature Clipping Recent works have shown that the LLMs are subject to the risks of self-consistent (overconfident) hallucinations (Ren et al., 2022; Ji et al., 2023), which has not been considered by existing consistency based methods. Therefore, to address those failure cases caused by overconfident generation, a test time feature clipping approach is introduced during the computation of EigenScore. As shown in Figure. 2, we illustrate the activation distribution in the penultimate layer of LLaMA-7B. An intuitive observation is that the penultimate layer of LLMs tends to exhibit numerous extreme features, consequently increasing the likelihood of generating overconfident and self-consistent generations. Inspired by prior works that rectify internal activations to reduce overconfident prediction for Out-of-Distribution (OOD) detection (Sun et al., 2021; Djurisic et al., 2022; Chen et al., 2024), we introduce a test time feature clipping (FC) method to prevent LLMs generate overconfident hallucinations. To rectify those extreme features, the FC operation is defined as the following piecewise function $$ FC(h)=\begin{cases}h_{min},&h<h_{min}\\ h,&h_{min}\leq h\leq h_{max}\\ h_{max}&h>h_{max}\end{cases} \tag{8} $$ where $h$ represents the feature of the hidden embeddings in the penultimate layer of the LLMs, $h_{min}$ and $h_{max}$ are two thresholds for determining the minimum and maximum truncation activations. When $h_{min}=-∞$ and $h_{max}=+∞$ , the output feature embedding is equivalent to the original output. For the determination of the optimal truncation thresholds, a memory bank which dynamically pushes and pops element in it, is utilized to conserve $N$ token embeddings during test time. Then, for each hidden neuron, the thresholds $h_{min}$ and $h_{max}$ are set to the top and bottom $p$ -th percentiles of the features in the memory bank. Refer to the three-sigma-rule Pukelsheim (1994), we set $p=0.2$ in all cases. This implies that the activations falling within the largest and smallest top 0.2% in the memory bank are identified as abnormal features and subsequently truncated for reducing overconfident generation. 4 Experiments 4.1 Experimental Setup Datasets. We utilize four widely used question answering (QA) datasets for evaluation, including two open-book conversational QA datasets CoQA (Reddy et al., 2019) and SQuAD (Rajpurkar et al., 2016), as well as two closed-book QA datasets TriviaQA (Joshi et al., 2017) and Natural Questions (NQ) (Kwiatkowski et al., 2019). We follow Lin et al. (2023) to utilize the development split of CoQA with 7983 QA pairs, the validation split of NQ with 3610 QA pairs and the validation split of the TriviaQA (rc.nocontext subset) with 9,960 deduplicated QA pairs. For the SQuAD dataset, we filter out the QA pairs with their flag is_impossible = True, and utilize the subset of the development-v2.0 split with 5928 QA pairs. The lengths of the sequences vary in the four datasets. Specifically, the ground truth answers in CoQA and SQuAD are relatively longer, while and TriviaQA typically consists of answers that are only with one or two words. Models. We use two representative open source LLMs, including LLaMA (Touvron et al., 2023a) and OPT (Zhang et al., 2022) in our experiments. Specifically, we consider off-the-shelf LLaMA-7B https://huggingface.co/decapoda-research/llama-7b-hf, LLaMA-13B https://huggingface.co/decapoda-research/llama-13b-hf, OPT-6.7B https://huggingface.co/facebook/opt-6.7b and their corresponding tokenizer provided by Hugging Face. We use the pre-trained wights and do not finetune these models in all cases. Evaluation Metrics. Following prior work Kuhn et al. (2022); Ren et al. (2022), we evaluate the hallucination detection ability of different methods by employing them to determine whether the generation is correct or not. Therefore, the area under the receiver operator characteristic curve (AUROC) and Pearson Correlation Coefficient (PCC) are utilized as the performance measure. AUROC is a popular metric to evaluate the quality of a binary classifier and uncertainty measure (Ren et al., 2022; Lin et al., 2023). Higher AUROC scores are better. PCC is utilized to measure the correlation between the hallucination detection metric and the correctness measure, which is usually defined as the ROUGE score (Lin, 2004) or semantic similarity (Reimers & Gurevych, 2019) between the generated answers and ground truth answers. A higher PCC score is better. Baselines. We compare our proposal with the most popular uncertainty-based methods Perplexity Ren et al. (2022) and Length-normalized Entropy (LN-Entropy) Malinin & Gales (2020), and the consistency-based metric Lexical Similarity (Lin et al., 2022). Besides, in order to investigate whether traditional OOD detection methods can be used for hallucination detection, we also introduce a popular OOD detection method Energy score (Liu et al., 2020) as a comparison method. Correctness Measure. We follow Kuhn et al. (2022); Lin et al. (2023) to utilize both the ROUGE-L (Lin, 2004) and the semantic similarity (Reimers & Gurevych, 2019) as the correctness measure. ROUGE-L https://github.com/google-research/google-research/tree/master/rouge is an n-gram based metric that computes the longest common subsequence between two pieces of text. The generation is regarded as correct when the ROUGE-L (f-measure) is large than a given threshold, which we set to 0.5 in our main experiments. Besides, we also use the embedding similarity as the correctness measure. The sentence embeddings of model generation and the ground truth answer are extracted by the nli-roberta-large model https://huggingface.co/sentence-transformers/nli-roberta-large, and the generation is regarded as true when the cosine similarity between two embeddings is larger than 0.9. Implementation Details. Implementation of this work is based on pytorch and transformers libraries. For the hyperparameters that are used for sampling strategies of LLMs’ decoder, we set temperature to 0.5, top-p to 0.99 and top-k to 5 through the experiments. The number of generations is set to $K=10$ . For the sentence embedding used in our proposal, we use the last token embedding of the sentence in the middle layer, i.e., the layer index is set to int(L/2). For the regularization term of the covariance matrix, we set $\alpha=0.001$ . For the memory bank used to conserve token embeddings, we set $N=3000$ . When implement the Energy Score, we average the token-level energy score as the sentence-level energy score. 4.2 Main Results Table 1: Hallucination detection performance evaluation of different methods on four QA tasks. AUROC (AUC) and Pearson Correlation Coefficient (PCC) are utilized to measure the performance. $\text{AUC}_{s}$ represents AUROC score with sentence similarity as correctness measure, and $\text{AUC}_{r}$ represents AUROC score with ROUGE-L score as correctness measure. All numbers are percentages. | LLaMA-7B Energy LN-Entropy | Perplexity 51.7 68.7 | 64.1 54.7 73.6 | 68.3 1.0 30.6 | 20.4 45.1 70.1 | 57.5 47.6 70.9 | 60.0 -10.7 30.0 | 10.2 64.3 72.8 | 74.0 64.8 73.7 | 74.7 18.2 29.8 | 30.1 66.8 83.4 | 83.6 67.1 83.2 | 83.6 29.1 54.0 | 54.4 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | Lexical Similarity | 74.8 | 77.8 | 43.5 | 74.9 | 76.4 | 44.0 | 73.8 | 75.9 | 30.6 | 82.6 | 84.0 | 55.6 | | | EigenScore | 80.4 | 80.8 | 50.8 | 81.5 | 81.2 | 53.5 | 76.5 | 77.1 | 38.3 | 82.7 | 82.9 | 57.4 | | | LLaMA-13B | Perplexity | 63.2 | 66.2 | 20.1 | 59.1 | 61.7 | 14.2 | 73.5 | 73.4 | 36.3 | 84.7 | 84.5 | 56.5 | | Energy | 47.5 | 49.2 | -5.9 | 36.0 | 39.2 | -20.2 | 59.1 | 59.8 | 14.7 | 71.3 | 71.5 | 36.7 | | | LN-Entropy | 68.8 | 72.9 | 31.2 | 72.4 | 74.0 | 36.6 | 74.9 | 75.2 | 39.4 | 83.4 | 83.1 | 54.2 | | | Lexical Similarity | 74.8 | 77.6 | 44.1 | 77.4 | 79.1 | 48.6 | 74.9 | 76.8 | 40.3 | 82.9 | 84.3 | 57.5 | | | EigenScore | 79.5 | 80.4 | 50.2 | 83.8 | 83.9 | 57.7 | 78.2 | 78.1 | 49.0 | 83.0 | 83.0 | 58.4 | | | OPT-6.7B | Perplexity | 60.9 | 63.5 | 11.5 | 58.4 | 69.3 | 8.6 | 76.4 | 77.0 | 32.9 | 82.6 | 82.0 | 50.0 | | Energy | 45.6 | 45.9 | -14.5 | 41.6 | 43.3 | -16.4 | 60.3 | 58.6 | 25.6 | 70.6 | 68.8 | 37.3 | | | LN-Entropy | 61.4 | 65.4 | 18.0 | 65.5 | 66.3 | 22.0 | 74.0 | 76.1 | 28.4 | 79.8 | 80.0 | 43.0 | | | Lexical Similarity | 71.2 | 74.0 | 38.4 | 72.8 | 74.0 | 39.3 | 71.5 | 74.3 | 23.1 | 78.2 | 79.7 | 42.5 | | | EigenScore | 76.5 | 77.5 | 45.6 | 81.7 | 80.8 | 49.9 | 77.9 | 77.2 | 33.5 | 80.3 | 80.4 | 0.485 | | Effectiveness of EigenScore. In Table. 1, we compare our proposed EigenScore with several representative reliability evaluation methods on three LLMs and four QA datasets. The results show that: (1) In both LLaMA and OPT models, our proposed EigenScore consistently outperforms other comparison methods by a large margin in CoQA, SQuAD and NQ datasets under different evaluation metrics. In particular, the EigenScore outperforms Lexical Similarity by 5.6% in CoQA and 8.9% in SQuAD with AUROC metric at most. (2) It’s interesting to see that the Perplexity performs best in TriviaQA dataset but performs poorly on other datasets, especially for CoQA and SQuAD. This is because the generations and ground truth answers on TrivaiQA dataset is very simple, with only one or two words in the most cases. Therefore, the performance of different methods in TriviaQA is close and by simply averaging the token-level confidence as uncertainty measure performs well. (3) On average, the performance in LLaMA-13B is better than that in LLaMA-7B and OPT-6.7B, while the performances in LLaMA-7B is slightly better than that in OPT-6.7B. It demonstrates that better hallucination detection performance can be achieved with a more powerful pre-trained LLM. Effectiveness of Feature Clipping. To demonstrate the effectiveness of the introduced test-time feature clipping, we compare the hallucination detection performance of different methods with and without applying the feature clipping technique. The results are shown in Table 2. As can be seen, the introduced feature clipping consistently improves the performance of different methods, with the largest improvement being 1.8% in AUROC. Table 2: Hallucination detection performance evaluation of different methods with and without (w/o) applying feature clipping (FC). ”+FC” denotes applying feature clipping and EigenScore (w/o) denotes EigenScore without applying feature clipping. All numbers are percentages. | Model | LLaMA-7B | OPT-6.7B | | | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | Datasets | CoQA | NQ | CoQA | NQ | | | | | | Methods | AUC s | PCC | AUC s | PCC | AUC s | PCC | AUC s | PCC | | LN-Entropy | 68.7 | 30.6 | 72.8 | 29.8 | 61.4 | 18.0 | 74.0 | 28.4 | | LN-Entropy + FC | 70.0 | 33.4 | 73.4 | 31.1 | 62.6 | 21.4 | 74.8 | 30.3 | | Lexical Similarity | 74.8 | 43.5 | 73.8 | 30.6 | 71.2 | 38.4 | 71.5 | 23.1 | | Lexical Similarity + FC | 76.6 | 46.3 | 74.8 | 32.1 | 72.6 | 40.2 | 72.4 | 24.2 | | EigenScore (w/o) | 79.3 | 48.9 | 75.9 | 38.3 | 75.3 | 43.1 | 77.1 | 32.2 | | EigenScore | 80.4 | 50.8 | 76.5 | 38.3 | 76.5 | 45.6 | 77.9 | 33.5 | 4.3 Ablation Studies <details> <summary>x4.png Details</summary>  ### Visual Description ## Line Chart: AUROC Performance Across Number of Generations ### Overview The chart illustrates the performance of three evaluation metrics (LN-Entropy, Lexical Similarity, EigenScore) measured by Area Under the Receiver Operating Characteristic curve (AUROC) across varying numbers of generations (5–40). All metrics show distinct trends, with EigenScore achieving the highest AUROC values throughout. ### Components/Axes - **X-axis**: "Number of Generations" (discrete markers at 5, 10, 15, 20, 30, 40). - **Y-axis**: "AUROC" (continuous scale from 72 to 80). - **Legend**: Located in the top-left corner, mapping: - Gray diamond: LN-Entropy - Teal circle: Lexical Similarity - Orange star: EigenScore ### Detailed Analysis 1. **LN-Entropy (Gray Diamonds)**: - Starts at ~72.3 at 5 generations. - Peaks at ~73.1 at 15 generations. - Declines slightly to ~72.9 at 40 generations. - Trend: Minor fluctuations with no significant growth. 2. **Lexical Similarity (Teal Circles)**: - Begins at ~72.8 at 5 generations. - Rises steadily to ~75.2 at 40 generations. - Trend: Consistent upward trajectory with minimal plateauing. 3. **EigenScore (Orange Stars)**: - Starts at ~74.5 at 5 generations. - Sharp increase to ~77.5 by 15 generations. - Plateaus between ~77.6–77.8 from 20–40 generations. - Trend: Rapid initial improvement followed by stabilization. ### Key Observations - **EigenScore** consistently outperforms other metrics, achieving the highest AUROC values (74.5–77.8). - **Lexical Similarity** shows the second-best performance, with a steady increase (72.8–75.2). - **LN-Entropy** remains the lowest-performing metric, with values clustered between 72.3–73.1. - EigenScore’s sharp rise (5–15 generations) suggests rapid early gains, while Lexical Similarity’s linear growth indicates sustained improvement. ### Interpretation The data suggests that **EigenScore** is the most effective metric for evaluating performance in this context, particularly in early generations. Its plateau at ~77.8 implies diminishing returns after 15 generations. **Lexical Similarity** demonstrates reliable, incremental gains, making it a viable alternative for long-term evaluation. **LN-Entropy**’s stagnation highlights potential limitations in capturing performance improvements over time. The divergence between EigenScore and Lexical Similarity at later generations (e.g., 30–40) may indicate differing sensitivities to model complexity or data distribution shifts. </details> <details> <summary>x5.png Details</summary>  ### Visual Description ## Bar Chart: AUROC Performance Across Layer Indexes ### Overview The chart displays the Area Under the Receiver Operating Characteristic curve (AUROC) performance metric across five distinct layer indexes (5, 10, 20, 30, 33). AUROC values range from 75 to 82 on the y-axis, with two horizontal reference lines at 79 (gray dashed) and 80.5 (yellow dashed). The bars represent performance at each layer index, with teal-colored bars centered under their respective x-axis labels. ### Components/Axes - **X-axis (Layer Indexes)**: Discrete categories labeled 5, 10, 20, 30, 33. Bars are evenly spaced and centered under their labels. - **Y-axis (AUROC)**: Continuous scale from 75 to 82, with increments of 1. Two horizontal reference lines: - Gray dashed line at 79.0 - Yellow dashed line at 80.5 - **Bars**: Teal-colored vertical bars representing AUROC values for each layer index. ### Detailed Analysis - **Layer 5**: AUROC ≈ 78.8 (below the 79.0 reference line). - **Layer 10**: AUROC ≈ 80.2 (between 79.0 and 80.5 lines). - **Layer 20**: AUROC ≈ 80.7 (highest value, just above the 80.5 reference line). - **Layer 30**: AUROC ≈ 80.4 (slightly below Layer 20, near the 80.5 line). - **Layer 33**: AUROC ≈ 79.2 (between 79.0 and 80.5 lines). ### Key Observations 1. **Peak Performance**: Layer 20 achieves the highest AUROC (80.7), surpassing the 80.5 reference line. 2. **Decline After Layer 20**: Performance drops at Layer 30 (80.4) and further at Layer 33 (79.2), suggesting diminishing returns or potential overfitting in deeper layers. 3. **Baseline Comparison**: All layers except Layer 5 exceed the 79.0 reference line, indicating generally strong performance. ### Interpretation The data suggests that increasing layer complexity initially improves model performance (up to Layer 20), but further layers degrade AUROC. This could indicate an optimal layer count at 20, beyond which the model may overfit or lose generalization ability. The 80.5 reference line (yellow) may represent a target threshold, with Layer 20 being the only index to exceed it. The decline at Layer 33 aligns with common patterns in deep learning architectures, where excessive depth can harm performance without proper regularization. </details> Figure 3: (a) Performance in LLaMA-7B and NQ dataset with different number of generations. (b) Performance in LLaMA-7B and CoQA dataset with sentence embedding in different layers. Orange line indicates using the last token’s embedding in the middle layer (layer 17) as sentence embedding. Gray line indicates using the averaged token embedding in the last layer as sentence embedding. The performance is measured by $\text{AUROC}_{s}$ . Number of Generations. For the methods that explore semantic consistency for hallucination detection, the number of generations $K$ is a key factor to the performance. Therefore, to evaluate the impact of the number of generations, we select $K$ from $\{5,10,15,20,30,40\}$ and perform experiments with LLaMA-7B and the NQ dataset. The performance in Figure 3 shows that: (1) Our proposed EigenScore consistently outperforms LN-Entropy and Lexical Similarity by a large margin for different $K$ . (2) When $K<15$ , the performance of different methods increases as $K$ increases and when $K>15$ , the performance tends to remain stable. The results suggeste that setting K to 20 provides the optimal trade-off between performance and inference cost. (3) Compared to EigenScore and Lexical Similarity, LN-Entropy is less sensitive to the number of generations, which demonstrates that Lexical Similarity and our EigenScore are more effective at utilizing the information in different generations. How EigenScore Performs with Different Sentence Embeddings. In the main experiments, we employ the embedding of the last token in the middle layer as sentence embedding. Here, we also investigate how the model performs with different sentence embeddings. In Figure. 3, we show the hallucination detection performance by using sentence embedding from different layers. The results show that using the sentence embedding in the shallow and final layers yields significantly inferior performance compared to using sentence embedding in the layers close to the middle. Besides, another interesting observation is that utilizing the embedding of the last token as the sentence embedding achieves superior performance compared to simply averaging the token embeddings, which suggests that the last token of the middle layers retain more information about the truthfulness. Sensitivity to Correctness Measures. It’s difficult to develop automatic metrics for QA task that correlate well with human evaluations. Therefore, the choice of correctness measures is a crucial component of hallucination detection evaluation. In this section, we evaluate the performance with different correctness measure thresholds in LLaMA-7B and CoQA dataset. The experimental results are presented in Table. 3. It shows that the threshold has a great influence on the final hallucination detection performance. Significantly, our proposed EigenScore consistently outperforms comparison methods in different thresholds. Besides, the results also indicate that the hallucination detection performance of different methods will be better under a rigorous correctness measure. Table 3: Performance evaluation with different correctness measure thresholds in LLaMA-7B and CoQA dataset. The ROUGE-L (f-measure) score and Sentence Similarity with different thresholds are employed to measure the correctness of the generated answers. | Perplexity | 65.2 | 68.3 | 68.1 | 63.7 | 63.5 | 64.1 | | --- | --- | --- | --- | --- | --- | --- | | LN-Entropy | 67.4 | 73.6 | 74.1 | 65.2 | 65.6 | 68.7 | | Lexical Similarity | 75.8 | 77.8 | 79.3 | 72.8 | 73.9 | 74.8 | | EigenScore | 76.4 | 80.8 | 83.5 | 75.9 | 77.2 | 80.4 | Sensitivity to Hyperparameters. The hyperparameters, including temperature, top-k and top-p, of the LLMs’ decoder determine the diversity of the generations. To evaluate the impact of those hyperparameters. We provide a sensitivity analysis in Figure 4. As observed, the performance is greatly influenced by temperature but shows little sensitivity to top-k. The performance of the consistency based methods (EigenScore and Lexical Similarity) drops significantly when the temperature is greater than 1. The optimal temperature can be selected from $[0.1,1.0]$ . <details> <summary>x6.png Details</summary>  ### Visual Description ## Line Chart: Sensitivity to Temperature ### Overview The chart illustrates the sensitivity of four evaluation metrics (Perplexity, LN-Entropy, Lexical Similarity, EigenScore) to temperature variations, measured via Area Under the Receiver Operating Characteristic curve (AUROC). Temperature ranges from 0.1 to 5.0, with AUROC scores plotted on a 40–100 scale. All metrics show distinct trends, with some peaking at intermediate temperatures before declining. ### Components/Axes - **X-axis (Temperature)**: Logarithmic scale from 0.1 to 5.0, with markers at 0.1, 0.3, 0.5, 1, 3, and 5. - **Y-axis (AUROC)**: Linear scale from 40 to 100, with increments of 10. - **Legend**: Located in the top-right corner, associating: - **Blue (dashed line with stars)**: Perplexity - **Gray (solid line with diamonds)**: LN-Entropy - **Teal (dashed line with circles)**: Lexical Similarity - **Orange (solid line with stars)**: EigenScore ### Detailed Analysis 1. **Perplexity (Blue)**: - Starts at ~65 AUROC at 0.1 temperature. - Remains flat until 0.5, then drops sharply to ~60 by 5.0. - Minimal variability across temperatures. 2. **LN-Entropy (Gray)**: - Begins at ~67 AUROC at 0.1. - Peaks at ~69 around 0.5, then declines to ~66 by 5.0. - Slight upward trend until 0.5, then gradual decline. 3. **Lexical Similarity (Teal)**: - Starts at ~70 AUROC at 0.1. - Rises to ~75 at 0.5, then plummets to ~58 by 5.0. - Sharp decline after 0.5 temperature. 4. **EigenScore (Orange)**: - Begins at ~72 AUROC at 0.1. - Peaks at ~80 around 0.5, then falls to ~58 by 5.0. - Steepest decline among all metrics after 0.5. ### Key Observations - **EigenScore** achieves the highest AUROC (~80) at 0.5 temperature but drops sharply to ~58 at 5.0. - **Lexical Similarity** shows the most dramatic decline (~75 → ~58) after 0.5 temperature. - **LN-Entropy** and **Perplexity** exhibit relative stability, with LN-Entropy maintaining higher scores than Perplexity across all temperatures. - All metrics decline significantly at higher temperatures (3–5), suggesting reduced performance under extreme conditions. ### Interpretation The data suggests that **EigenScore** and **Lexical Similarity** are highly sensitive to temperature changes, performing optimally at intermediate temperatures (0.5) but degrading sharply at higher values. This could indicate overfitting or instability in these metrics under extreme conditions. In contrast, **LN-Entropy** and **Perplexity** demonstrate robustness, maintaining consistent performance across temperatures. The sharp declines in EigenScore and Lexical Similarity at high temperatures may highlight their reliance on specific linguistic patterns that become less reliable as temperature increases. These findings could inform metric selection in temperature-sensitive applications, favoring stability over peak performance. </details> <details> <summary>x7.png Details</summary>  ### Visual Description ## Line Chart: Sensitivity to Top-K ### Overview The chart illustrates the sensitivity of four evaluation metrics (Perplexity, LN-Entropy, Lexical Similarity, EigenScore) to varying Top-K values (3–50) using Area Under the Receiver Operating Characteristic curve (AUROC) as the performance metric. All metrics exhibit relatively stable performance across Top-K ranges, with EigenScore consistently outperforming others. ### Components/Axes - **X-axis (Top-K)**: Discrete values at 3, 5, 10, 20, 30, 50. - **Y-axis (AUROC)**: Scale from 40 to 90, with increments of 10. - **Legend**: Located in the bottom-right corner, mapping: - Blue crosses (×): Perplexity - Gray diamonds (◆): LN-Entropy - Teal circles (●): Lexical Similarity - Orange stars (★): EigenScore ### Detailed Analysis 1. **Perplexity (Blue ×)**: - Flat line at ~65 AUROC across all Top-K values. - No significant variation observed. 2. **LN-Entropy (Gray ◆)**: - Slight upward trend from ~67 (Top-K=3) to ~69 (Top-K=50). - Minimal fluctuation between intermediate Top-K values. 3. **Lexical Similarity (Teal ●)**: - Stable at ~75 AUROC for Top-K=3–20. - Minor increase to ~76 at Top-K=50. 4. **EigenScore (Orange ★)**: - Consistently highest performance (~80 AUROC) across all Top-K. - Slight dip to ~79 at Top-K=10, then recovery to ~80. ### Key Observations - **EigenScore** maintains the highest AUROC (79–80) regardless of Top-K, indicating robustness. - **Perplexity** is the least sensitive metric, showing no change across Top-K. - **LN-Entropy** exhibits the weakest sensitivity, with a marginal 2-point increase. - **Lexical Similarity** remains stable until Top-K=50, where it marginally improves. ### Interpretation The data suggests that **EigenScore** is the most reliable metric for evaluating model performance across varying Top-K configurations, as it consistently achieves the highest AUROC. **Perplexity** and **Lexical Similarity** demonstrate stability but lower performance, while **LN-Entropy** shows minimal sensitivity. The flat trends imply that Top-K adjustments have limited impact on these metrics, though EigenScore’s slight dip at Top-K=10 warrants further investigation into potential anomalies. This analysis is critical for optimizing model evaluation strategies in natural language processing tasks. </details> Figure 4: (a) Performance sensitivity to temperature. (b) Performance sensitivity to top-k. The performance is measured by $\text{AUROC}_{s}$ . 5 Related Work Reliability Evaluation of LLMs During real-world deployments, the reliability of LLMs poses a substantial challenge, as LLMs reveal their propensity to exhibit unreliable generations (Ji et al., 2023; Zhang et al., 2023). Therefore, considerable efforts has been made to address the security and reliability evaluation of LLMs (Huang et al., 2023; Malinin & Gales, 2020; Kuhn et al., 2022; Kadavath et al., 2022; Cohen et al., 2023; Azaria & Mitchell, 2023). Among those methods, uncertainty based metric has been widely explored, which typically involves predictive confidence or entropy of the output token (Malinin & Gales, 2020; Kuhn et al., 2022; Duan et al., 2023). Besides, consistency based methods also play an important role in reliability evaluation, which hypothesizes that LLMs tend to generate logically inconsistent responses to the same question when they are indecisive and hallucinating contents (Kuhn et al., 2022; Raj et al., 2023; Manakul et al., 2023). Based on the consistency hypothesis, researchers also found it is feasible to prompt the LLMs to evaluate their responses themselves (Kadavath et al., 2022; Cohen et al., 2023; Manakul et al., 2023). Eigenvalue as Divergence Measure The eigenvalue or determinant of covariance matrix captures the variability of the data and has been widely explored as divergence measure in a wide range of machine learning tasks (Wold et al., 1987; Kulesza & Taskar, 2011; Xu et al., 2021; Zhouyin & Liu, 2021; Cai et al., 2015). For instance, in Wold et al. (1987), the authors proposed the well-known Principal Components Analysis (PCA) and demonstrates that the most largest eigenvalues of sample covariance matrix corresponds to the principle semantic of sample set. Besides, the determinant of covariance matrix, determined by the eigenvalues, has been utilized to sample a diversity subset in determinantal point processes (DDP) (Kulesza & Taskar, 2011) and activation learning (Xu et al., 2021) tasks, which demonstrates the determinant of covariance matrix is a good diversity measure. Besides, several studies also proposed to approximate the differential entropy with the logarithm determinant of covariance matrix (Zhouyin & Liu, 2021; Klir & Wierman, 1999). 6 Conclusion Measuring the hallucination degree of LLM’s generation is of critical importance in enhancing the security and reliability of LLM-based AI systems. This work presents an INSIDE framework to exploit the semantic information that are retained within the internal states of LLMs for hallucination detection. Specifically, a simple yet effective EigenScore is proposed to measure the semantic consistency across different generations in the embedding space. Besides, to identify those self-consistent (overconfident) hallucinations which have been overlooked by previous methods, a feature clipping technique is introduced to reduce overconfident generations by truncating extreme features. Significant performance improvement has been achieved in several popular LLMs and QA benchmarks. Although our experiments focus on QA task, our method does not make any assumptions about the task modality, and we believe our method is widely applicable to other tasks, such as summarization and translation. We hope that our insights inspire future research to further explore the internal semantics of LLMs for hallucination detection. References - Almazrouei et al. (2023) Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Maitha Alhammadi, Mazzotta Daniele, Daniel Heslow, Julien Launay, Quentin Malartic, et al. The falcon series of language models: Towards open frontier models. Hugging Face repository, 2023. - Azaria & Mitchell (2023) Amos Azaria and Tom Mitchell. The internal state of an llm knows when its lying. arXiv preprint arXiv:2304.13734, 2023. - Bai et al. (2022) Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. - Cai et al. (2015) T Tony Cai, Tengyuan Liang, and Harrison H Zhou. Law of log determinant of sample covariance matrix and optimal estimation of differential entropy for high-dimensional gaussian distributions. Journal of Multivariate Analysis, 137:161–172, 2015. - Chang et al. (2023) Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Kaijie Zhu, Hao Chen, Linyi Yang, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. arXiv preprint arXiv:2307.03109, 2023. - Chen et al. (2024) Chao Chen, Zhihang Fu, Kai Liu, Ze Chen, Mingyuan Tao, and Jieping Ye. Optimal parameter and neuron pruning for out-of-distribution detection. Advances in Neural Information Processing Systems, 36, 2024. - Cohen et al. (2023) Roi Cohen, May Hamri, Mor Geva, and Amir Globerson. Lm vs lm: Detecting factual errors via cross examination. arXiv e-prints, pp. arXiv–2305, 2023. - Djurisic et al. (2022) Andrija Djurisic, Nebojsa Bozanic, Arjun Ashok, and Rosanne Liu. Extremely simple activation shaping for out-of-distribution detection. In The Eleventh International Conference on Learning Representations, 2022. - Duan et al. (2023) Jinhao Duan, Hao Cheng, Shiqi Wang, Chenan Wang, Alex Zavalny, Renjing Xu, Bhavya Kailkhura, and Kaidi Xu. Shifting attention to relevance: Towards the uncertainty estimation of large language models. arXiv preprint arXiv:2307.01379, 2023. - Hendrycks & Gimpel (2016) Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. - Huang et al. (2023) Yuheng Huang, Jiayang Song, Zhijie Wang, Huaming Chen, and Lei Ma. Look before you leap: An exploratory study of uncertainty measurement for large language models. arXiv e-prints, pp. arXiv–2307, 2023. - 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. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1–38, 2023. - Joshi et al. (2017) Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1601–1611, 2017. - Kadavath et al. (2022) Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield Dodds, Nova DasSarma, Eli Tran-Johnson, et al. Language models (mostly) know what they know. arXiv e-prints, pp. arXiv–2207, 2022. - Klir & Wierman (1999) George Klir and Mark Wierman. Uncertainty-based information: elements of generalized information theory, volume 15. Springer Science & Business Media, 1999. - Kuhn et al. (2022) Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations, 2022. - Kulesza & Taskar (2011) Alex Kulesza and Ben Taskar. k-dpps: Fixed-size determinantal point processes. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 1193–1200, 2011. - Kwiatkowski et al. (2019) Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453–466, 2019. - Li et al. (2023) Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. arXiv preprint arXiv:2306.03341, 2023. - Liang et al. (2022) Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, et al. Holistic evaluation of language models. arXiv preprint arXiv:2211.09110, 2022. - Lin (2004) Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pp. 74–81, 2004. - Lin et al. (2023) Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. Generating with confidence: Uncertainty quantification for black-box large language models. arXiv e-prints, pp. arXiv–2305, 2023. - Lin et al. (2022) Zi Lin, Jeremiah Zhe Liu, and Jingbo Shang. Towards collaborative neural-symbolic graph semantic parsing via uncertainty. Findings of the Association for Computational Linguistics: ACL 2022, 2022. - Liu et al. (2020) Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in neural information processing systems, 33:21464–21475, 2020. - Malinin & Gales (2020) Andrey Malinin and Mark Gales. Uncertainty estimation in autoregressive structured prediction. In International Conference on Learning Representations, 2020. - Manakul et al. (2023) Potsawee Manakul, Adian Liusie, and Mark JF Gales. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896, 2023. - OpenAI (2023) OpenAI. Gpt-4 technical report, 2023. - Ouyang et al. (2022) Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022. - Pukelsheim (1994) Friedrich Pukelsheim. The three sigma rule. The American Statistician, 48(2):88–91, 1994. - Raj et al. (2023) Harsh Raj, Vipul Gupta, Domenic Rosati, and Subhabrata Majumdar. Semantic consistency for assuring reliability of large language models. arXiv preprint arXiv:2308.09138, 2023. - Rajpurkar et al. (2016) Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 2383–2392, 2016. - Reddy et al. (2019) Siva Reddy, Danqi Chen, and Christopher D Manning. Coqa: A conversational question answering challenge. Transactions of the Association for Computational Linguistics, 7:249–266, 2019. - Reimers & Gurevych (2019) Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Association for Computational Linguistics, 2019. - Ren et al. (2022) Jie Ren, Jiaming Luo, Yao Zhao, Kundan Krishna, Mohammad Saleh, Balaji Lakshminarayanan, and Peter J Liu. Out-of-distribution detection and selective generation for conditional language models. In The Eleventh International Conference on Learning Representations, 2022. - Shi et al. (2022) Freda Shi, Daniel Fried, Marjan Ghazvininejad, Luke Zettlemoyer, and Sida I Wang. Natural language to code translation with execution. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp. 3533–3546, 2022. - Sun et al. (2021) Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems, 34:144–157, 2021. - Touvron et al. (2023a) Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023a. - Touvron et al. (2023b) Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023b. - Wang et al. (2022) Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. - Wold et al. (1987) Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37–52, 1987. - Xu et al. (2021) Xinyi Xu, Zhaoxuan Wu, Chuan Sheng Foo, and Bryan Kian Hsiang Low. Validation free and replication robust volume-based data valuation. Advances in Neural Information Processing Systems, 34:10837–10848, 2021. - Yin et al. (2023) Zhangyue Yin, Qiushi Sun, Qipeng Guo, Jiawen Wu, Xipeng Qiu, and Xuanjing Huang. Do large language models know what they don’t know? arXiv preprint arXiv:2305.18153, 2023. - Zhang et al. (2022) Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. - Zhang et al. (2023) Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. Siren’s song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. - Zhou et al. (2023) Kaitlyn Zhou, Dan Jurafsky, and Tatsunori Hashimoto. Navigating the grey area: Expressions of overconfidence and uncertainty in language models. arXiv preprint arXiv:2302.13439, 2023. - Zhouyin & Liu (2021) Zhanghao Zhouyin and Ding Liu. Understanding neural networks with logarithm determinant entropy estimator. arXiv preprint arXiv:2105.03705, 2021. Appendix A Performance Evaluation on TruthfulQA TruthfulQA is an important benchmark to evaluate the truthfulness of LLMs (Joshi et al., 2017). Therefore, we also compare our proposal with the baseline methods in the TruthfulQA benchmark. The optimal classification thresholds is determined by maximizing the G-Mean value, which is defined as $\textbf{G-Mean}=\sqrt{TPR*(1-FPR)}$ . The results are presented in Table 4. For the ITI Li et al. (2023), which trains multiple binary classifiers with the internal embeddings for hallucination detection, we report the best performance in their paper. As can be seen, our proposal consistently outperforms the baseline methods and achieves comparable performance as ITI when we utilize 50 in-distribution prompts. It’s worth nothing that the ITI relies on training 1024 binary classifiers in TruthQA datasets, and they report the best performance (83.3) in the validation set. Therefore, their best performance is better than our proposal which has not been trained on TruthfulQA. However, training on the validation set also limits the generalization of their method on other domains (Li et al., 2023). As TruthfulQA is a very challenging dataset for LLMs, zero-shot inference results in poor performance. Therefore, we follow previous work (Bai et al., 2022) to utilize different number of in-distribution prompts during inference time. The results show that the performance could be significantly improved when we increase the number of prompts, which also explains why ITI performs good. Table 4: Performance comparison of different methods on TruthfulQA dataset. LexialSim denotes Lexical Similarity and SelfCKGPT denotes SelfCheckGPT. Hallucination detection accuracy is reported. # Prompt denotes the number of prompt templates. For ITI Li et al. (2023), we report the best number in their paper directly. All numbers are percentages. | 5 20 50 | 70.0 76.4 73.1 | 71.2 77.7 77.9 | 73.6 77.9 73.6 | 74.2 76.8 78.3 | 83.3 83.3 83.3 | 76.7 79.5 81.3 | | --- | --- | --- | --- | --- | --- | --- | Appendix B Comparison with More Competitive Methods To demonstrate the effectiveness of our proposal, we also compare our EigenScore with several competitive methods, including Semantic Entropy (SemanticEnt) (Kuhn et al., 2022), Shifting Attention to Relevance (SentSAR) (Duan et al., 2023) and SelfCheckGPT (SelfCKGPT) (Manakul et al., 2023). We follow the experimental setting in Duan et al. (2023) to set the number of generation to $N=10$ for OPT-6.7B and $N=5$ for LLaMA. For the results of SementicEnt and SentSAR, we report the number in Duan et al. (2023) directly. For the implementation of SelfCheckGPT, we leverage the SelfCheckBERTScore provided in the official code package https://github.com/potsawee/selfcheckgpt. The comparison results in Table 5 demonstrate that our EigenScore significantly outperforms the competitors. Additionally, both SentSAR and SelfCheckGPT exhibit comparable performance, which is much superior to Semantic Entropy. Note that both SentSAR, SelfCheckGPT and our proposal evaluate the quality of LLMs’ generation by exploring the self-consistency across multiple outputs. However, compared to Semantic Entropy (Kuhn et al., 2022) or SelfCheckGPT (Manakul et al., 2023) which relies on another language model for sentence embedding extraction, our approach leverages the internal states of LLMs, which retain highly-concentrated semantic information. Besides, the EigenScore defined by the LogDet of the sentence covariance matrix is able to capture the semantic consistency more effectively compared to the sentence-wise similarity (Manakul et al., 2023). Furthermore, the proposed feature clipping strategy allows our model to identify the overconfident hallucinations, which has not been investigated by previous works (Kuhn et al., 2022; Manakul et al., 2023) Table 5: Performance comparison of EigenScore and and several state-of-the-art methods on CoQA dataset. AUC s represents AUROC with the sentence similarity as correctness measure, and AUC r represents using ROUGE-L as correctness measure. All numbers are percentages. | OPT-6.7B | 63.1 | 71.7 | 69.8 | 72.2 | 70.2 | 74.1 | 71.9 | 77.5 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | LLaMA-7B | 64.9 | 68.2 | 70.4 | 65.8 | 68.7 | 72.9 | 71.2 | 75.7 | | LLaMA-13B | 65.3 | 66.7 | 71.4 | 64.7 | 68.1 | 77.0 | 72.8 | 79.8 | Appendix C Performance Evaluation on More LLMs In the main experiments, we evaluate the performance of different methods in LLaMA-7B, LLaMA-13B and OPT-6.7B. To demonstrate the robustness of our method across different models, we also provide the performance comparison in the recent LLaMA2-7B (Touvron et al., 2023b) and Falcon-7B models (Almazrouei et al., 2023). Table 6 reveals that our proposal consistently exhibits superior performance compared to the other methods across different LLMs. Table 6: Performance evaluation on LLaMA2-7B and Falcon-7B. LexicalSim denotes Lexical Similarity and SelfCKGPT denotes SelfCheckGPT. AUC s and AUC r are utilized as correctness measure. Other experimental settings are consistent with Table 1. | LLaMA2-7b | CoQA | 62.2 | 66.6 | 69.9 | 75.2 | 74.4 | 77.5 | 72.4 | 75.1 | 78.6 | 80.7 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | NQ | 70.8 | 70.2 | 72.1 | 71.2 | 72.1 | 72.9 | 69.1 | 68.1 | 74.4 | 73.7 | | | Falcon-7b | CoQA | 57.0 | 60.6 | 62.6 | 63.2 | 74.8 | 76.4 | 76.7 | 77.9 | 80.8 | 80.6 | | NQ | 74.3 | 74.7 | 74.6 | 74.7 | 73.8 | 75.4 | 74.7 | 74.0 | 76.3 | 75.7 | | Appendix D Computational Efficiency Analysis As our proposal is a sampling based approach, additional inference cost is required to generate multiple outputs for accurate hallucination detection. We compare our proposal with the base LLM and other comparing methods in LLaMA-7B and LLaMA-13B. All experiments are performed on NVIDIA-A100 and we set the number of generations to $N=10$ through the experiments. The average inference time per question is shown in Fig. 5. As observed, our EigenScore is about 10 times more efficient than the methods that rely on another large model to measure the self-consistency (such as SelfCheckGPT (Manakul et al., 2023)), and shares the similar computational overhead with the LN-Entropy and Lexical Similarity. Compared to the computational overhead of generating multiple outputs, the cost of feature clipping and EigenScore computation is negligible (0.06s). It is worth noting that the inference overhead required to generate multiple results is not linearly proportional to the time required to generate a single output, owing to the sampling and decoding strategy of the autoregressive LLM model. <details> <summary>x8.png Details</summary>  ### Visual Description ## Bar Chart: Computational Cost Comparison in LLaMA-7B ### Overview The chart compares the computational cost (inference time per question) of six methods used in the LLaMA-7B framework. The y-axis represents inference cost in seconds per question, while the x-axis lists the methods. The data shows significant variation in computational efficiency across methods. ### Components/Axes - **Title**: "Computational Cost Comparison in LLaMA-7B" - **X-axis (Categories)**: - BaseLLM - Perplexity - LN-Entropy - LexicaSim - SelfCKGPT - EigenScore - **Y-axis (Scale)**: - Label: "Inference Cost (Second/Question)" - Range: 0 to 12 (increments of 2) - **Bars**: - Colored in blue (no legend present) - Heights correspond to labeled values on top of each bar ### Detailed Analysis - **BaseLLM**: 0.24 seconds/question (shortest bar) - **Perplexity**: 0.24 seconds/question (tied with BaseLLM) - **LN-Entropy**: 0.80 seconds/question - **LexicaSim**: 0.81 seconds/question - **SelfCKGPT**: 10.68 seconds/question (tallest bar, 13x higher than LN-Entropy) - **EigenScore**: 0.81 seconds/question (tied with LexicaSim) ### Key Observations 1. **Outlier**: SelfCKGPT exhibits a computational cost **13.35x higher** than the next most expensive method (LN-Entropy). 2. **Efficiency Cluster**: BaseLLM, Perplexity, LN-Entropy, LexicaSim, and EigenScore all operate within a narrow range (0.24–0.81 seconds/question). 3. **Symmetry**: LexicaSim and EigenScore share identical costs (0.81), while BaseLLM and Perplexity are identical (0.24). ### Interpretation The data suggests **SelfCKGPT is an outlier in computational demand**, potentially due to architectural complexity or iterative processing requirements. The clustering of other methods around 0.24–0.81 seconds/question indicates they are similarly optimized for efficiency. This disparity highlights trade-offs between accuracy (if SelfCKGPT offers superior performance) and resource constraints in LLaMA-7B deployments. The absence of a legend implies all methods use the same metric, but the lack of error bars or confidence intervals limits conclusions about statistical significance. </details> (a) LLaMA-7B <details> <summary>x9.png Details</summary>  ### Visual Description ## Bar Chart: Computational Cost Comparison in LLaMA-13B ### Overview The chart compares the computational cost (inference time per question) of six methods used in LLaMA-13B. The y-axis represents inference cost in seconds per question, ranging from 0 to 12. The x-axis lists six methods: BaseLLM, Perplexity, LN-Entropy, LexicaSim, SelfCKGPT, and EigenScore. All bars are blue, with no legend present. ### Components/Axes - **Title**: "Computational Cost Comparison in LLaMA-13B" - **X-axis**: Categories (methods) labeled as: - BaseLLM - Perplexity - LN-Entropy - LexicaSim - SelfCKGPT - EigenScore - **Y-axis**: "Inference Cost (Second/Question)" with increments at 0, 2, 4, 6, 8, 10, 12. - **Bars**: Six vertical bars, each labeled with a numerical value above it. ### Detailed Analysis - **BaseLLM**: 0.31 seconds/question (shortest bar). - **Perplexity**: 0.31 seconds/question (tied with BaseLLM). - **LN-Entropy**: 1.27 seconds/question. - **LexicaSim**: 1.28 seconds/question. - **SelfCKGPT**: 10.26 seconds/question (tallest bar, ~8x higher than LN-Entropy). - **EigenScore**: 1.27 seconds/question (tied with LN-Entropy). ### Key Observations 1. **Outlier**: SelfCKGPT has a computational cost **~8x higher** than the next most expensive method (LN-Entropy/LexicaSim/EigenScore). 2. **Low-Cost Methods**: BaseLLM and Perplexity share the lowest cost (0.31s/question). 3. **Similarity**: LN-Entropy, LexicaSim, and EigenScore cluster tightly (1.27–1.28s/question). 4. **Scale Disparity**: SelfCKGPT’s bar is visually dominant, emphasizing its inefficiency. ### Interpretation The data highlights **SelfCKGPT as a computational outlier**, suggesting it is significantly less efficient than other methods in LLaMA-13B. This could stem from architectural complexity, algorithmic overhead, or task-specific demands. The near-identical costs of LN-Entropy, LexicaSim, and EigenScore imply comparable efficiency, while BaseLLM and Perplexity represent the most resource-light approaches. The absence of a legend simplifies interpretation but limits contextual understanding of the methods’ purposes. The chart underscores trade-offs between computational cost and potential performance gains, critical for optimizing LLaMA-13B deployments in resource-constrained environments. </details> (b) LLaMA-13B Figure 5: Inference cost comparison of different methods in LLaMA-7B and LLaMA-13B. BaseLLM denotes the LLM without using any hallucination detection metrics. LexicalSim denotes Lexical Similarity and SelfCKGPT denotes SelfCkeckGPT. Appendix E Evaluation with Exact Match In the main experiments, we employ the ROUGE and sentence similarity as correctness measure, which are widely used for natural language generation evaluation (Chang et al., 2023; Kuhn et al., 2022; Huang et al., 2023). In order to facilitate the comparison of our work’s performance with other works, we also provide the evaluation results by employing exact match (Liang et al., 2022) as the correctness score, which is much more strict to determine a generation as correct. The results in Table 7 show similar conclusions to those in Table 1, which demonstrates that our proposal significantly outperforms the compared methods in most cases. Table 7: Performance evaluation with Exact Match as correctness measure. LexicalSim denotes the Lexical Similarity. The experimental settings are consistent with Table 1. | LLaMA-7B | CoQA | 63.7 | 70.7 | 76.1 | 83.0 | | --- | --- | --- | --- | --- | --- | | SQuAD | 57.3 | 72.1 | 76.9 | 83.9 | | | NQ | 75.3 | 75.6 | 75.8 | 80.1 | | | TriviaQA | 82.5 | 83.4 | 81.8 | 82.4 | | | OPT-6.7B | CoQA | 59.4 | 61.7 | 71.8 | 79.4 | | SQuAD | 56.7 | 65.2 | 72.7 | 82.9 | | | NQ | 79.8 | 78.1 | 73.2 | 79.8 | | | TriviaQA | 83.8 | 81.3 | 79.3 | 82.7 | | Appendix F More visualization and ablation for Feature Clipping In Fig. 6, we illustrate the distributions of neuron activation from four selected tokens. As can be seen, the distribution changes a lot across samples. Therefore, it is risky to determine the clipping threshold with only the current input sample (EigenScore-C). A feasible solution is to pre-compute the optimal threshold based on a batch of input samples (EigenScore-P). Besides, another solution is to dynamically record the activation values and determine the threshold during the inference process (EigenScore-MB). We have experimented with both solutions and the experimental results are presented in Table 8. The results demonstrate that determining the thresholds with a memory bank works slightly better. We attribute this variability to potential differences in the activation distributions across various datasets. Table 8: Ablation study of determining the clipping threshold with different technique. EigenScore-C indicates determining the threshold with the current input sample. EigenScore-P indicates pre-computing the threshold with a batch of samples. EigenScore-MB denotes using memory bank to determine the optimal threshold. AUC s is reported. | EigenScore-C | 78.1 | 74.8 | | --- | --- | --- | | EigenScore-P | 79.9 | 75.3 | | EigenScore-MB | 80.4 | 76.5 | <details> <summary>x10.png Details</summary>  ### Visual Description ## Line Graph: Neuron Activation Distribution ### Overview The image depicts a line graph titled "Neuron Activation Distribution," visualizing neuron activation values across neuron indexes. The graph shows a teal-colored line fluctuating vertically between approximately -6 and 8, with the x-axis representing neuron indexes (0 to 4000) and the y-axis representing neuron activation values (-10 to 10). The data exhibits irregular oscillations without a clear directional trend. ### Components/Axes - **Title**: "Neuron Activation Distribution" (centered at the top). - **X-axis**: Labeled "Neuron Indexes," scaled from 0 to 4000 in increments of 1000. - **Y-axis**: Labeled "Neuron Activations," scaled from -10 to 10 in increments of 2. - **Line**: A single teal-colored line representing neuron activation values. No legend is present to explain the line's meaning. - **Gridlines**: Implied by the structured spacing of axis ticks, but not explicitly drawn. ### Detailed Analysis - **Neuron Indexes (X-axis)**: - Range: 0 to 4000 (4001 total points, assuming integer indexing). - Key markers: 0, 1000, 2000, 3000, 4000. - **Neuron Activations (Y-axis)**: - Range: -10 to 10 (21 possible discrete values). - Notable extremes: Peaks near 8 (e.g., ~3500) and troughs near -6 (e.g., ~2000). - **Line Behavior**: - Fluctuates irregularly, with no consistent upward or downward trend. - Average activation hovers near 0, but with significant variability (±3–4 units). - Sharp spikes and dips occur at irregular intervals (e.g., ~1500, ~2500, ~3500). ### Key Observations 1. **High Variability**: Activations deviate widely from the mean (0), suggesting diverse or noisy responses. 2. **Extreme Values**: - Maximum activation: ~8 at neuron index ~3500. - Minimum activation: ~-6 at neuron index ~2000. 3. **No Clear Pattern**: The line lacks periodicity or symmetry, indicating randomness or complex underlying dynamics. 4. **Spatial Distribution**: Peaks and troughs are distributed unevenly across the neuron index range. ### Interpretation The graph likely represents the output of a neural network layer or biological neuron population under specific stimuli. The absence of a clear trend implies: - **Noise or Complexity**: Activations may reflect stochastic processes, external perturbations, or high-dimensional interactions. - **Potential Outliers**: Spikes at ~3500 and ~2000 could indicate critical neurons or anomalies in the dataset. - **Biological/Technical Relevance**: In neuroscience, such variability might correlate with adaptive responses; in machine learning, it could signal overfitting or insufficient regularization. No textual content, additional languages, or embedded diagrams are present. The graph focuses solely on quantitative neuron activation data. </details> <details> <summary>x11.png Details</summary>  ### Visual Description ## Line Chart: Neuron Activation Distribution ### Overview The image depicts a line chart titled "Neuron Activation Distribution," visualizing the activation levels of neurons across a range of indexes. The chart shows significant variability in neuron activations, with values fluctuating between approximately -10.0 and +7.5. The data spans 4,000 neuron indexes, with no discernible pattern or trend. ### Components/Axes - **Title**: "Neuron Activation Distribution" (top-center, black text). - **X-Axis**: Labeled "Neuron Indexes," scaled from 0 to 4,000 in increments of 1,000. - **Y-Axis**: Labeled "Neuron Activations," scaled from -10.0 to +7.5 in increments of 2.5. - **Data Series**: A single cyan line representing neuron activation values across indexes. - **Legend**: No explicit legend is present in the image. The cyan line is the only data series. ### Detailed Analysis - **Neuron Indexes (X-Axis)**: The horizontal axis spans 0 to 4,000, with evenly spaced tick marks at 0, 1,000, 2,000, 3,000, and 4,000. - **Neuron Activations (Y-Axis)**: The vertical axis ranges from -10.0 (minimum) to +7.5 (maximum), with intermediate ticks at -7.5, -5.0, -2.5, 0.0, 2.5, 5.0, and 7.5. - **Data Trends**: The cyan line exhibits erratic fluctuations, with peaks reaching up to +7.5 and troughs dipping to -10.0. No consistent upward or downward trend is observed. The line crosses the zero-axis multiple times, indicating alternating positive and negative activation values. - **Notable Features**: - The highest peak occurs near Neuron Index 500 (~+7.5). - The lowest trough occurs near Neuron Index 2,500 (~-10.0). - The majority of activations cluster between -5.0 and +5.0, with extreme values occurring sporadically. ### Key Observations 1. **High Variability**: Neuron activations show no uniformity, with values spanning the full range of the Y-axis. 2. **Extreme Outliers**: Two extreme values (-10.0 and +7.5) stand out as potential outliers. 3. **No Systematic Pattern**: The lack of a clear trend suggests randomness or noise in the data. ### Interpretation The chart likely represents a snapshot of neural activity in a computational or biological model, where activations vary unpredictably across neurons. The absence of a trend implies that the system may lack coherence or that the data reflects transient states (e.g., noise, external stimuli). The extreme outliers could indicate neurons with unusually high or low sensitivity, or errors in data collection. This distribution might be used to analyze neural network stability, identify anomalous neurons, or study dynamic activation patterns in response to specific inputs. </details> <details> <summary>x12.png Details</summary>  ### Visual Description ## Line Chart: Neuron Activation Distribution ### Overview The image depicts a line chart titled "Neuron Activation Distribution," visualizing neuron activation levels across a range of neuron indexes. The chart shows fluctuations in activation values, with a cyan-colored line representing the data series. The y-axis measures neuron activations (ranging from -30 to 30), while the x-axis represents neuron indexes (0 to 4000). The line exhibits variability, with peaks and troughs distributed across the index range. ### Components/Axes - **Title**: "Neuron Activation Distribution" (centered at the top). - **Y-Axis**: Labeled "Neuron Activations," scaled from -30 to 30 in increments of 10. - **X-Axis**: Labeled "Neuron Indexes," scaled from 0 to 4000 in increments of 1000. - **Legend**: A single cyan-colored line labeled "Neuron Activations" (positioned near the top-right, though partially cropped). - **Line**: Cyan-colored, plotted across the entire x-axis range. ### Detailed Analysis - **Neuron Activations**: - The line fluctuates around a mean of approximately 0, with values ranging from -30 to +30. - Notable peaks: - A sharp spike to ~25 at neuron index ~500. - A trough to ~-25 at neuron index ~2500. - Most activations cluster between -10 and +10, with occasional outliers. - **Distribution Pattern**: - The line shows no clear upward or downward trend, suggesting no systematic bias in activation levels. - Variability is distributed across the index range, with no dominant region of high or low activity. ### Key Observations 1. **Outliers**: Two prominent outliers at indices ~500 (peak) and ~2500 (trough), deviating significantly from the general range. 2. **Symmetry**: The distribution appears roughly symmetric around 0, though the trough at ~2500 is slightly deeper than the peak at ~500. 3. **Noise**: Minor fluctuations (e.g., ±5 units) are frequent, indicating high variability in neuron activations. ### Interpretation The data suggests that neuron activations are highly variable across the indexed population, with no consistent directional trend. The outliers at indices ~500 and ~2500 may represent neurons with heightened sensitivity or unique connectivity patterns, potentially critical for specific neural processing tasks. The overall symmetry around 0 implies a balanced system, but the presence of outliers highlights the importance of individual neuron behavior in broader neural dynamics. This distribution could reflect responses to external stimuli, internal noise, or adaptive mechanisms in neural networks. </details> <details> <summary>x13.png Details</summary>  ### Visual Description ## Line Chart: Neuron Activation Distribution ### Overview The image depicts a line chart titled "Neuron Activation Distribution," visualizing neuron activation levels across a range of neuron indexes. The y-axis represents neuron activation values (ranging from -20 to 30), while the x-axis represents neuron indexes (0 to 4000). The chart shows a cyan-colored line with significant fluctuations, indicating variability in neuron activation across the dataset. ### Components/Axes - **Title**: "Neuron Activation Distribution" (centered at the top). - **Y-Axis**: Labeled "Neuron Activations," with tick marks at intervals of 10 (-20, -10, 0, 10, 20, 30). - **X-Axis**: Labeled "Neuron Indexes," with tick marks at intervals of 1000 (0, 1000, 2000, 3000, 4000). - **Legend**: Positioned at the top-right, labeled "Neuron Activations" with a cyan color swatch. - **Data Line**: A single cyan line traversing the chart, with no additional data series. ### Detailed Analysis - **Y-Axis Values**: The neuron activation values oscillate between approximately -20 and +30, with no clear baseline (mean or median) explicitly marked. - **X-Axis Values**: Neuron indexes span from 0 to 4000, with no gaps or missing data points. - **Line Behavior**: The cyan line exhibits irregular spikes and troughs. Notable peaks occur near neuron indexes 500 (activation ~25), 1500 (activation ~20), and 3500 (activation ~15). Troughs are observed near indexes 1000 (activation ~-15) and 3000 (activation ~-10). The majority of the line hovers near 0, with smaller fluctuations (±5) in between larger deviations. ### Key Observations 1. **High Variability**: Neuron activations show significant variability, with some neurons exhibiting extreme positive or negative activations. 2. **Outliers**: Peaks at indexes 500, 1500, and 3500 stand out as local maxima, while troughs at 1000 and 3000 are notable minima. 3. **Baseline Proximity**: Most neuron activations cluster near 0, suggesting a baseline activation level with intermittent deviations. ### Interpretation The chart suggests a heterogeneous distribution of neuron activation, where individual neurons exhibit diverse responses. The large spikes and troughs may indicate neurons with specialized or heightened sensitivity to specific stimuli, while the baseline clustering implies a majority of neurons maintain moderate activation levels. The lack of a consistent trend (e.g., gradual increase/decrease) implies no overarching pattern in the dataset, potentially reflecting biological noise or contextual variability in neural activity. The absence of a labeled dataset or experimental conditions limits conclusions about causality or functional significance. </details> Figure 6: Activation distributions of four selected tokens in LLaMA-7B. Appendix G Limitations and future work By exploring the internal states of LLM and introducing an EigenScore metric, the hallucination detection performance has been significantly improved. However, there are several limitations of this study. One critical limitation is that the proposed method relies on the internal information of the LLMs, therefore cannot be applied to the black-box models (OpenAI, 2023), where users can not access the hidden states. Additionally, our proposal is a sampling-based approach, necessitating the generation of multiple outputs, which introduces additional inference overhead for accurate hallucination detection. Furthermore, another limitation is that we only focus on hallucination detection in this study and have not yet mitigate the hallucination with our proposal. In the future work, we hope to reduce the inference cost of our proposal and leverage the EigenScore to mitigate hallucinations. We believe the proposed EigenScore is a strong baseline for detecting hallucination and we invite further researchers to utilize and enhance our proposal. Appendix H Cases Study H.1 Hallucination Detection Cases We show several cases with LLaMA-7B and NQ dataset. The number of generation is set to $N=10$ . We determine the optimal detection threshold by maximizing the G-Mean value, which is defined as $\textbf{G-Mean}=\sqrt{TPR*(1-FPR)}$ . The hallucination detection thresholds for different metrics are Perplexity: 0.535; LN-Entropy: 0.153; LexicalSimilarity 0.489; SelfCheckGPT: 0.168; EigenScore: -1.74. For LexicalSimilarity, a score larger than the threshold indicates non-hallucination. For other metrics, a score smaller than the thresholds indicates non-hallucination. Green indicates non-hallucination answer and Red indicates hallucination answer generated by LLM. ✓ (✗) indicates that the hallucination is (not) correctly identified by the metric. Question: the german princes who chose the holy roman empire were called GTAns: prince-electors LLMAns: electors BatchGenerations: [’electors’, ’electors’, ’electors’, ’electors’, ’electors’, ’electors’, ’electors’, ’electors’, ’electors’, ’electors’] Perplexity: 0.361 ✓ LN-Entropy: 0.027 ✓ LexicalSimilarity: 1.0 ✓ SentBERTScore: 0.0 ✓ EigenScore: -2.63 ✓ EigenValue: [4.87719579e+00 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: where is fe best absorbed in the body GTAns: in the duodenum LLMAns: in the small intestine BatchGenerations: [’in the liver’, ’small intestine’, ’in the intestines’, ’the small intestine’, ’the small intestine’, ’in the liver’, ’small intestine’, ’fat’, ’in the small intestine’, ’fatty tissues’] Perplexity: 0.641 ✓ LN-Entropy: 0.213 ✓ LexicalSimilarity: 0.357 ✓ SentBERTScore: 0.258 ✓ EigenScore: -1.40 ✓ EigenValue: [3.71561676e+00 4.34496729e-01 3.77751922e-01 1.75326593e-01 9.92596975e-02 4.20723353e-02 2.49385766e-02 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: who did the united states win its independence from GTAns: the British Empire LLMAns: britain BatchGenerations: [’britain’, ’england’, ’great britain’, ’great britain’, ’england’, ’england’, ’england’, ’england’, ’great britain’, ’great britain’] Perplexity: 0.598 ✗ LN-Entropy: 0.266 ✗ LexicalSimilarity: 0.415 ✗ SentBERTScore: 0.397 ✗ EigenScore: -2.23 ✓ EigenValue: [4.46843402e+00 2.82423429e-01 3.88702191e-02 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: who won the most stanley cups in history GTAns: Montreal Canadiens LLMAns: the detroit red wings BatchGenerations: [’the detroit red wings’, ’the detroit red wings’, ’the detroit red wings’, ’the detroit red wings’, ’the detroit red wings’, ’the detroit red wings’, ’the detroit red wings’, ’the detroit red wings’, ’the detroit red wings’, ’the detroit red wings’] Perplexity: 0.366 ✗ LN-Entropy: 0.025 ✗ LexicalSimilarity: 1.0 ✗ SentBERTScore: 0.0 ✗ EigenScore: -2.63 ✗ EigenValue: [5.23534401e+00 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: what is the second book in the alchemyst series GTAns: The Magician LLMAns: the alchemyst: the secret of the immortal Nicholas flamel BatchGenerations: [’the magician in the middle’, ”the magician’s nephew”, ’the magician’, ’the alchemyst’, ’the magician’, ’the alchemyst’, ’the magician in the middle’, ’the magician in amsterdam’, ’the magician in the middle’, ’the secret of the near witch’] Perplexity: 0.459 ✗ LN-Entropy: 0.207 ✓ LexicalSimilarity: 0.499 ✗ SentBERTScore: 0.431 ✓ EigenScore: -1.41 ✓ EigenValue: [2.09138499e+00 6.95605781e-01 3.85931973e-01 3.40671669e-01 2.15372994e-01 1.77304781e-01 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: a good that can be used in place of another good GTAns: A substitute good LLMAns: substitute BatchGenerations: [’a substitute’, ’a substitute’, ’substitute’, ’a substitute’, ’substitute’, ’substitute’, ’substitute’, ’substitute’, ’substitute’, ’substitute’] Perplexity: 0.614 ✗ LN-Entropy: 0.194 ✗ LexicalSimilarity: 0.844 ✓ SentBERTScore: 0.076 ✓ EigenScore: -2.40 ✓ EigenValue: [5.11365925e+00 1.75884104e-01 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: who said i will not go quietly into the night GTAns: Welsh poet Dylan Thomas LLMAns: douglas macarthur BatchGenerations: [’douglas macarthur’, ’elvis presley’, ’douglas macarthur’, ’douglas macarthur’, ’douglas macarthur’, ’robin williams’, ’robert f. kennedy’, ’robin williams’, ’robert f kennedy’, ’Dwight Eisenhower’] Perplexity: 0.502 ✗ LN-Entropy: 0.121 ✗ LexicalSimilarity: 0.178 ✓ SentBERTScore: 0.337 ✓ EigenScore: -1.61 ✓ EigenValue: [3.32824135e+00 5.87944819e-01 3.70390066e-01 1.70849836e-01 1.17707239e-01 5.17925563e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: when did they stop saying mass in latin GTAns: from 1967 onwards LLMAns: 1969 BatchGenerations: [’in 1969’, ’1969’, ’around 1969’, ’in 1965’, ’in 1969’, ’1969’, ’In 1969, the Roman Missal was changed from Latin to the vernacular.’, ’1969’, ’1969 when the second vatican council was held’, ’1969’] Perplexity: 0.631 ✓ LN-Entropy: 0.152 ✗ LexicalSimilarity: 0.540 ✗ SentBERTScore: 0.273 ✓ EigenScore: -1.59 ✓ EigenValue: [3.80408192e+00 4.83987672e-01 3.03207580e-01 8.80366008e-02 6.59790286e-02 3.26742841e-02 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: where does the highway of heroes start and end GTAns: Glen Miller Road in Trenton LLMAns: it starts in ottawa and ends in toronto BatchGenerations: [’start in ottawa and end in toronto’, ’it starts in toronto and ends in ottawa’, ”it starts at the queen’s park gate at queen’s park and ends at toronto’s coronation park”, ”it starts in toronto and ends at queen’s park”, ’starts in Toronto and ends in Trenton’, ’it starts at the CNE and ends at the cenotaph in downtown toronto’, ’it starts in downtown toronto and ends in ottawa’, ’it starts in the city of Toronto and ends in the city of Hamilton’, ’it starts in toronto and ends in ottawa’, ’it starts in Toronto and ends in Ottawa’] Perplexity: 0.657 ✓ LN-Entropy: 0.260 ✓ LexicalSimilarity: 0.594 ✗ SentBERTScore: 0.149 ✗ EigenScore: -1.05 ✓ EigenValue: [3.31983018e+00 3.98560810e-01 2.17094299e-01 2.06965709e-01 1.53575354e-01 1.27925588e-01 7.82365136e-02 3.28158137e-02 1.01995086e-02 1.00000000e-03] Question: which body part(s) occupy the greatest portion of the primary motor cortex GTAns: face LLMAns: the hand BatchGenerations: [’the hand’, ’the hand’, ’the fingers’, ’the hand’, ’the hand’, ’the hand and the foot’, ’the hand’, ’the hand and the foot’, ’the hand and face’, ’arms and legs’] Perplexity: 1.41 ✓ LN-Entropy: 0.300 ✓ LexicalSimilarity: 0.568 ✗ SentBERTScore: 0.163 ✗ EigenScore: -1.69 ✓ EigenValue: [3.76273036e+00 6.16284067e-01 1.96541049e-01 1.73505005e-01 1.28407153e-01 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: who said have you no sense of decency GTAns: Joseph Nye Welch LLMAns: Joseph Nye Welch BatchGenerations: [’Joseph N. Welch’, ’Joseph N. Welch’, ’joe stalin’, ’joseph mccarthy’, ’Joseph N. Welch’, ’Joseph N. Welch’, ’Joseph Nye Welch’, ’joseph mccarthy’, ’joe mccarthy’, ’joseph mccarthy’] Perplexity: 0.666 ✗ LN-Entropy: 0.212 ✗ LexicalSimilarity: 0.437 ✗ SentBERTScore: 0.391 ✗ EigenScore: -1.85 ✓ EigenValue: [3.63114083e+00 8.11672323e-01 2.00385898e-01 3.19140618e-02 1.74251264e-02 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] H.2 Model generations with many and few Outliers To demonstrate the relationship between the number of extreme features and model outputs, we provide several examples with many/few extreme features. The results show that when there are many extreme features, the model tends to generate consistent hallucination outputs across multiple generations. Instead, when there are few extreme features, the model generates diverse hallucination outputs which can be spotted by different hallucination detection metrics. Question: who sang on great gig in the sky GTAns: Clare Torry LLMAns: freddie mercury AvgNumOutliers: 15 BatchGenerations: [’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’] Perplexity: 0.263 ✗ LN-Entropy: 0.028 ✗ LexicalSimilarity: 1.0 ✗ SentBERTScore: 0.0 ✗ EigenScore: -2.63 ✗ EigenValue: [4.65740187e+00 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: what are the top five wine producing states GTAns: Washington LLMAns: California AvgNumOutliers: 13 BatchGenerations: [’ California,’, ’ California,’, ’ california,’, ’ California,’, ’ California,’, ’ California,’, ’ California,’, ’ california,’, ’ California,’, ’ California,’] Perplexity: 0.368 ✗ LN-Entropy: 0.075 ✗ LexicalSimilarity: 1.0 ✗ SentBERTScore: 0.054 ✗ EigenScore: -2.42 ✗ EigenValue: [5.30709315e+00 1.13222379e-01 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: how many seasons of rules of engagement is there GTAns: 7 LLMAns: 4 seasons AvgNumOutliers: 2 BatchGenerations: [’3 seasons’, ’4 seasons’, ’4 seasons’, ’6 seasons’, ’7 seasons’, ’3 (2007-2009)’, ’3 (2007-2009)’, ’4 seasons’, ’three’, ’11 seasons’] Perplexity: 0.996 ✓ LN-Entropy: 0.292 ✓ LexicalSimilarity: 0.307 ✓ SentBERTScore: 0.285 ✓ EigenScore: -1.60 ✓ EigenValue: [3.58548815e+00 5.87838054e-01 2.28057934e-01 1.36461300e-01 3.49712302e-02 1.11346059e-02 3.82259086e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: where did the first persian gulf war take place GTAns: Israel LLMAns: kuwait AvgNumOutliers: 3 BatchGenerations: [’Iraq’, ’Iraq and Kuwait’, ’Iraq’, ’kuwait’, ’kuwait’, ’in the middle east’, ’in iraq’, ’kuwait’, ’iraq’, ’kuwait’] Perplexity: 0.546 ✓ LN-Entropy: 0.281 ✓ LexicalSimilarity: 0.339 ✓ SentBERTScore: 0.224 ✓ EigenScore: -1.62 ✓ EigenValue: [3.59463352e+00 4.23782982e-01 2.57087067e-01 1.41513403e-01 6.20790226e-02 1.75980481e-02 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] H.3 Impact of Feature Clipping The texts in yellow represents model generations after applying feature clipping. The results show that after feature clipping, the overconfident generations can be appropriately suppressed, and some self-consistent hallucinations are finally identified. Question: what are the top five wine producing states GTAns: Washington LLMAns: California BatchGenerations: [’ California,’, ’ California,’, ’ california,’, ’ California,’, ’ California,’, ’ California,’, ’ California,’, ’ california,’, ’ California,’, ’ California,’] Perplexity: 0.368 ✗ LN-Entropy: 0.075 ✗ LexicalSimilarity: 1.0 ✗ SentBERTScore: 0.054 ✗ EigenScore: -2.42 ✗ EigenValue: [5.30709315e+00 1.13222379e-01 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] BatchGenerations: [’california’, ’california’, ’Washington’, ’california’, ’new york’, ’california’, ’washington’, ’california’, ’new york’, ’michigan’] EigenScore: -1.32 ✓ EigenValue: [3.23392755e+00 8.41049340e-01 2.52322804e-01 1.33473529e-01 7.19449437e-02 6.12184197e-02 1.02734249e-02 5.33703500e-03 3.09878029e-03 1.00000000e-03] Question: who sang on great gig in the sky GTAns: Clare Torry LLMAns: freddie mercury AvgNumOutliers: 15 BatchGenerations: [’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’] Perplexity: 0.263 ✗ LN-Entropy: 0.028 ✗ LexicalSimilarity: 1.0 ✗ SentBERTScore: 0.0 ✗ EigenScore: -2.63 ✗ EigenValue: [4.65740187e+00 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] BatchGenerations: [’claire torry’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’, ’freddie mercury’] EigenScore: -2.38 ✗ EigenValue: [4.38745800e+00 3.14982649e-01 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] Question: who are you in assassin’s creed 4 GTAns: third-person perspective LLMAns: Edward Kenway BatchGenerations: [’Edward Kenway’, ’Edward Kenway’, ’Edward Kenway’, ’Edward Kenway’, ’Edward Kenway’, ’Edward Kenway’, ’Edward Kenway’, ’Edward Kenway’, ’Edward Kenway’, ’Edward Kenway’] Perplexity: 0.264 ✗ LN-Entropy: 0.002 ✗ LexicalSimilarity: 1.0 ✗ SentBERTScore: 0.0 ✗ EigenScore: -2.67 ✗ EigenValue: [2.10973201e+00 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03] BatchGenerations: [’Edward Kenway’, ’Edward Kenway’, ’Connor’, ’Edward Kenway’, ’connor’, ’Connor’, ’alexander hamilton’, ’Edward Kenway’, ’ezio’, ’connor’] EigenScore: -1.68 ✓ EigenValue: [3.47825477e+00 7.48127381e-01 3.24792650e-01 2.17182636e-01 8.15050807e-02 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03 1.00000000e-03]