论文检索

Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, Yiqun Liu

Hallucinations in large language models (LLMs) refer to the phenomenon of LLMs producing responses that are coherent yet factually inaccurate. This issue undermines the effectiveness of LLMs in practical applications, necessitating research into detecting and mitigating hallucinations of LLMs. Previous studies have mainly concentrated on post-processing techniques for hallucination detection, which tend to be computationally intensive and limited in effectiveness due to their separation from the LLM’s inference process. To overcome these limitations, we introduce MIND, an unsupervised training framework that leverages the internal states of LLMs for real-time hallucination detection without requiring manual annotations. Additionally, we present HELM, a new benchmark for evaluating hallucination detection across multiple LLMs, featuring diverse LLM outputs and the internal states of LLMs during their inference process. Our experiments demonstrate that MIND outperforms existing state-of-the-art methods in hallucination detection.

Navreet Kaur, Monojit Choudhury, Danish Pruthi

As corporations rush to integrate large language models (LLMs) it is critical that they provide factually accurate information, that is robust to any presuppositions that a user may express. In this work, we introduce UPHILL, a dataset consisting of health-related queries with varying degrees of presuppositions. Using UPHILL, we evaluate the factual accuracy and consistency of InstructGPT, ChatGPT, GPT-4 and Bing Copilot models. We find that while model responses rarely contradict true health claims (posed as questions), all investigated models fail to challenge false claims. Alarmingly, responses from these models agree with 23-32% of the existing false claims, and 49-55% with novel fabricated claims. As we increase the extent of presupposition in input queries, responses from all models except Bing Copilot agree with the claim considerably more often, regardless of its veracity. Given the moderate factual accuracy, and the inability of models to challenge false assumptions, our work calls for a careful assessment of current LLMs for use in high-stakes scenarios.

Alessandro Scirè, Karim Ghonim, Roberto Navigli

Recent advancements in text summarization, particularly with the advent of Large Language Models (LLMs), have shown remarkable performance. However, a notable challenge persists as a substantial number of automatically-generated summaries exhibit factual inconsistencies, such as hallucinations. In response to this issue, various approaches for the evaluation of consistency for summarization have emerged. Yet, these newly-introduced metrics face several limitations, including lack of interpretability, focus on short document summaries (e.g., news articles), and computational impracticality, especially for LLM-based metrics. To address these shortcomings, we propose Factuality Evaluation of summarization based on Natural language Inference and Claim Extraction (FENICE), a more interpretable and efficient factuality-oriented metric. FENICE leverages an NLI-based alignment between information in the source document and a set of atomic facts, referred to as claims, extracted from the summary. Our metric sets a new state of the art on AGGREFACT, the de-facto benchmark for factuality evaluation. Moreover, we extend our evaluation to a more challenging setting by conducting a human annotation process of long-form summarization. In the hope of fostering research in summarization factuality evaluation, we release the code of our metric and our factuality annotations of long-form summarization at https://github.com/Babelscape/FENICE.

Jingtao Guo, Chunxia Zhang, Lingxi Li, Xiaojun Xue, Zhendong Niu

Knowledge graph completion (KGC) task is to infer the missing knowledge in the knowledge graph based on known factual triples. However, present KGC approaches still face the following two challenges. Those methods perform simple linear update on relation representation, and only local neighborhood information is aggregated, which makes it difficult to capture logic semantic between relations and global topological context information. To tackle the above challenges, we propose a unified joint approach with Topological Context learning and Rule Augmentation (TCRA) for KGC. The TCRA framework consists of an entity topological context learning mechanism based on dual-branch hierarchical graph attention network, and a relation rule context learning mechanism based on Rule-Transformer and rule-to-relation aggregator. The former mechanism encodes the topological structure features of entities, aggregates the local neighborhood topological context information of entities on the three levels (entity, relation and triple), and build clusters of global head or tail entities related to the same relation. It can capture the local and global topological context information of entities related to the same relation. The latter mechanism introduces chain-like Horn rules as the context information of relations, and encodes the logical semantic of relations to enrich the relation representation. Experimental performances on three benchmark datasets FB15k-237, WN18RR and Kinship indicate the effectiveness and superiority of our proposed approach. The codes are publicly available.

Zhiyuan Fan, Zhihong Chen, Benyou Wang

Despite the success of data augmentation in improving CLIP model, existing methods that utilize LLM or SAM to enrich the information in captions still suffer from several limitations, including insufficient detail and excessive hallucinations, ultimately resulting in compromised alignment and masking the true potential of dense information. This can lead to erroneous conclusions about CLIP’s ability to handle rich data, impeding the development of more effective models. To address the limitations of existing methods, we introduce a novel pipeline that generates highly detailed, factually accurate captions for images, which facilitates in-depth analysis of the potential for dense information in multimodal alignment. Contrary to previous findings, our investigation revealed that lengthening captions boosts performance across diverse benchmarks, even surpassing the effectiveness of meticulously crafted hard negative samples. Building on these insights, DELIP is introduced, demonstrably enhancing both foundational multimodal alignment and compositional reasoning abilities. Finally, we explore strategies to expand the context window of the text encoder, unlocking the potential of richer data for CLIP and paving the way for advancements in leveraging dense information for multimodal alignment.

Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liangyan Gui, Yu-Xiong Wang, Yiming Yang 等

Large Multimodal Models (LMM) are built across modalities and the misalignment between two modalities can result in “hallucination”, generating textual outputs that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the vision-language alignment, where human annotators are asked to compare two responses and pinpoint the more hallucinated one, and the vision-language model is trained to maximize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image-text pairs to improve the general capabilities of our model. To evaluate the proposed approach in real-world scenarios, we develop a new evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaVA-Bench dataset with the 96% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improvement of 60% on MMHAL-BENCH over other baselines.

Aswin Rrv, Nemika Tyagi, Md Nayem Uddin, Neeraj Varshney, Chitta Baral

This study explores the sycophantic tendencies of Large Language Models (LLMs), where these models tend to provide answers that match what users want to hear, even if they are not entirely correct. The motivation behind this exploration stems from the common behavior observed in individuals searching the internet for facts with partial or misleading knowledge. Similar to using web search engines, users may recall fragments of misleading keywords and submit them to an LLM, hoping for a comprehensive response. Our empirical analysis of several LLMs shows the potential danger of these models amplifying misinformation when presented with misleading keywords. Additionally, we thoroughly assess four existing hallucination mitigation strategies to reduce LLMs sycophantic behavior. Our experiments demonstrate the effectiveness of these strategies for generating factually correct statements. Furthermore, our analyses delve into knowledge-probing experiments on factual keywords and different categories of sycophancy mitigation.

Xuming Hu, Xiaochuan Li, Junzhe Chen, Yinghui Li, Yangning Li, Xiaoguang Li, Yasheng Wang, Qun Liu, Lijie Wen, Philip Yu 等

Generative search engines have the potential to transform how people seek information online, but generated responses from existing large language models (LLMs)-backed generative search engines may not always be accurate. Nonetheless, retrieval-augmented generation exacerbates safety concerns, since adversaries may successfully evade the entire system by subtly manipulating the most vulnerable part of a claim. To this end, we propose evaluating the robustness of generative search engines in the realistic and high-risk setting, where adversaries have only black-box system access and seek to deceive the model into returning incorrect responses. Through a comprehensive human evaluation of various generative search engines, such as Bing Chat, PerplexityAI, and YouChat across diverse queries, we demonstrate the effectiveness of adversarial factual questions in inducing incorrect responses. Moreover, retrieval-augmented generation exhibits a higher susceptibility to factual errors compared to LLMs without retrieval. These findings highlight the potential security risks of these systems and emphasize the need for rigorous evaluation before deployment. The dataset and code will be publicly available.

Jinwen He, Yujia Gong, Zijin Lin, Cheng’an Wei, Yue Zhao, Kai Chen

Large Language Models (LLMs) have revolutionized various domains with extensive knowledge and creative capabilities. However, a critical issue with LLMs is their tendency to produce outputs that diverge from factual reality. This phenomenon is particularly concerning in sensitive applications such as medical consultation and legal advice, where accuracy is paramount. Inspired by human lie detectors using physiological responses, we introduce the LLM Factoscope, a novel Siamese network-based model that leverages the inner states of LLMs for factual detection. Our investigation reveals distinguishable patterns in LLMs’ inner states when generating factual versus non-factual content. We demonstrate its effectiveness across various architectures, achieving over 96% accuracy on our custom-collected factual detection dataset. Our work opens a new avenue for utilizing LLMs’ inner states for factual detection and encourages further exploration into LLMs’ inner workings for enhanced reliability and transparency.

Afra Feyza Akyürek, Ekin Akyürek, Leshem Choshen, Derry Wijaya, Jacob Andreas

While language models (LMs) can sometimes generate factually correct text and estimate truth values of individual claims, these generally do not reflect a globally coherent, manipulable model of the world. As a consequence, current LMs also generate incorrect or nonsensical content, and are difficult to edit and bring up to date. We present a method called Deductive Closure Training (DCT) that uses LMs themselves to identify implications of (and contradictions within) the text that they generate, yielding an efficient self-supervised procedure for improving LM factuality. Given a collection of seed documents, DCT prompts LMs to generate additional text implied by these documents, reason globally about the correctness of this generated text, and finally fine-tune on text inferred to be correct. Given seed documents from a trusted source, DCT provides a tool for supervised model updating; if seed documents are sampled from the LM itself, DCT enables fully unsupervised fine-tuning for improved coherence and accuracy. Across the CREAK, MQuAKE, and Reversal Curse datasets, supervised DCT improves LM fact verification and text generation accuracy by 3-26%; on CREAK, fully unsupervised DCT improves verification accuracy by 12%. These results show that LMs’ reasoning capabilities during inference can be leveraged during training to improve their reliability.

Ekaterina Fadeeva, Aleksandr Rubashevskii, Artem Shelmanov, Sergey Petrakov, Haonan Li, Hamdy Mubarak, Evgenii Tsymbalov, Gleb Kuzmin, Alexander Panchenko, Timothy Baldwin 等

Large language models (LLMs) are notorious for hallucinating, i.e., producing erroneous claims in their output. Such hallucinations can be dangerous, as occasional factual inaccuracies in the generated text might be obscured by the rest of the output being generally factually correct, making it extremely hard for the users to spot them. Current services that leverage LLMs usually do not provide any means for detecting unreliable generations. Here, we aim to bridge this gap. In particular, we propose a novel fact-checking and hallucination detection pipeline based on token-level uncertainty quantification. Uncertainty scores leverage information encapsulated in the output of a neural network or its layers to detect unreliable predictions, and we show that they can be used to fact-check the atomic claims in the LLM output. Moreover, we present a novel token-level uncertainty quantification method that removes the impact of uncertainty about what claim to generate on the current step and what surface form to use. Our method Claim Conditioned Probability (CCP) measures only the uncertainty of a particular claim value expressed by the model. Experiments on the task of biography generation demonstrate strong improvements for CCP compared to the baselines for seven different LLMs and four languages. Human evaluation reveals that the fact-checking pipeline based on uncertainty quantification is competitive with a fact-checking tool that leverages external knowledge.

Zengkui Sun, Yijin Liu, Jiaan Wang, Fandong Meng, Jinan Xu, Yufeng Chen, Jie Zhou

Recently, Knowledge Editing has received increasing attention, since it could update the specific knowledge from outdated ones in pretrained models without re-training. However, as pointed out by recent studies, existing related methods tend to merely memorize the superficial word composition of the edited knowledge, rather than truly learning and absorbing it. Consequently, on the reasoning questions, we discover that existing methods struggle to utilize the edited knowledge to reason the new answer, and tend to retain outdated responses, which are generated by the original models utilizing original knowledge. Nevertheless, the outdated responses are unexpected for the correct answers to reasoning questions, which we named as the outdated issue. To alleviate this issue, in this paper, we propose a simple yet effective decoding strategy, i.e., outDated ISsue aware deCOding (DISCO), to enhance the performance of edited models on reasoning questions. Specifically, we capture the difference in the probability distribution between the original and edited models. Further, we amplify the difference of the token prediction in the edited model to alleviate the outdated issue, and thus enhance the model performance w.r.t the edited knowledge. Experimental results suggest that applying DISCO could enhance edited models to reason, e.g., on reasoning questions, DISCO outperforms the prior SOTA method by 12.99 F1 scores, and reduces the ratio of the outdated issue to 5.78% on the zsRE dataset.

Ante Wang, Linfeng Song, Baolin Peng, Lifeng Jin, Ye Tian, Haitao Mi, Jinsong Su, Dong Yu

This work studies mitigating fact-conflicting hallucinations for large language model (LLM) at inference time.Particularly, we propose a self-endorsement framework that leverages the fine-grained fact-level comparisons across multiple sampled responses.Compared with prior ensemble methods (e.g., self-consistency) that perform response-level selection, our approach can better alleviate hallucinations for knowledge-intensive tasks.Our approach can broadly benefit smaller and open-source LLMs as it mainly conducts simple content-based comparisons.Experiments on Biographies show that our method can effectively improve the factuality of generations with simple and intuitive prompts across different scales of LLMs.Besides, comprehensive analyses on TriviaQA and GSM8K demonstrate the potential of self-endorsement for broader application.

Aditi Mishra, Sajjadur Rahman, Kushan Mitra, Hannah Kim, Estevam Hruschka

Large language models (LLMs) are proficient at generating fluent text with minimal task-specific supervision. However, their ability to generate rationales for knowledge-intensive tasks (KITs) remains under-explored. Generating rationales for KIT solutions, such as commonsense multiple-choice QA, requires external knowledge to support predictions and refute alternate options. In this work, we consider the task of generating retrieval-augmented rationalization of KIT model predictions via external knowledge guidance within a few-shot setting. Surprisingly, crowd-workers preferred LLM-generated rationales over existing crowd-sourced rationales, generated in a similar knowledge-guided setting, on aspects such as factuality, sufficiency, and convincingness. However, fine-grained evaluation of such rationales highlights the need for further improvements in conciseness, novelty, and domain invariance. Additionally, through an expert-sourced study evaluating the reliability of the rationales, we demonstrate that humans’ trust in LLM-generated rationales erodes when communicated faithfully, i.e., without taking model prediction accuracy into account. We find that even instrumenting simple guardrails can be effective for reliable rationalization.

Zhunheng Wang, Xiaoyi Liu, Mengting Hu, Rui Ying, Ming Jiang, Jianfeng Wu, Yalan Xie, Hang Gao, Renhong Cheng

The demand for understanding and expressing emotions in the field of natural language processing is growing rapidly. Knowledge graphs, as an important form of knowledge representation, have been widely utilized in various emotion-related tasks. However, existing knowledge graphs mainly focus on the representation and reasoning of general factual knowledge, while there are still significant deficiencies in the understanding and reasoning of emotional knowledge. In this work, we construct a comprehensive and accurate emotional commonsense knowledge graph, ECoK. We integrate cutting-edge theories from multiple disciplines such as psychology, cognitive science, and linguistics, and combine techniques such as large language models and natural language processing. By mining a large amount of text, dialogue, and sentiment analysis data, we construct rich emotional knowledge and establish the knowledge generation model COMET-ECoK. Experimental results show that ECoK contains high-quality emotional reasoning knowledge, and the performance of our knowledge generation model surpasses GPT-4-Turbo, which can help downstream tasks better understand and reason about emotions. Our data and code is available from https://github.com/ZornWang/ECoK.

Marcio Fonseca, Shay B. Cohen

Although large language models (LLMs) exhibit remarkable capacity to leverage in-context demonstrations, it is still unclear to what extent they can learn new facts or concept definitions via prompts. To address this question, we examine the capacity of instruction-tuned LLMs to follow in-context concept annotation guidelines for zero-shot sentence labeling tasks. We design guidelines that present different types of factual and counterfactual concept definitions, which are used as prompts for zero-shot sentence classification tasks. Our results show that although concept definitions consistently help in task performance, only the larger models (with 70B parameters or more) have limited ability to work under counterfactual contexts. Importantly, only proprietary models such as GPT-3.5 can recognize nonsensical guidelines, which we hypothesize is due to more sophisticated alignment methods. Finally, we find that Falcon-180B-chat is outperformed by Llama-2-70B-chat is most cases, which indicates that increasing model scale does not guarantee better adherence to guidelines. Altogether, our simple evaluation method reveals significant gaps in concept understanding between the most capable open-source language models and the leading proprietary APIs.

Haochen Liu, Song Wang, Yaochen Zhu, Yushun Dong, Jundong Li

Large Language Models (LLMs) have shown unprecedented performance in various real-world applications. However, they are known to generate factually inaccurate outputs, a.k.a. the hallucination problem. In recent years, incorporating external knowledge extracted from Knowledge Graphs (KGs) has become a promising strategy to improve the factual accuracy of LLM-generated outputs. Nevertheless, most existing explorations rely on LLMs themselves to perform KG knowledge extraction, which is highly inflexible as LLMs can only provide binary judgment on whether a certain knowledge (e.g., a knowledge path in KG) should be used. In addition, LLMs tend to pick only knowledge with direct semantic relationship with the input text, while potentially useful knowledge with indirect semantics can be ignored. In this work, we propose a principled framework KELP with three stages to handle the above problems. Specifically, KELP is able to achieve finer granularity of flexible knowledge extraction by generating scores for knowledge paths with input texts via latent semantic matching. Meanwhile, knowledge paths with indirect semantic relationships with the input text can also be considered via trained encoding between the selected paths in KG and the input text. Experiments on real-world datasets validate the effectiveness of KELP.

Daniel Tamayo, Aitor Gonzalez-Agirre, Javier Hernando, Marta Villegas

Recent research has explored methods for updating and modifying factual knowledge in large language models, often focusing on specific multi-layer perceptron blocks. This study expands on this work by examining the effectiveness of existing knowledge editing methods across languages and delving into the role of attention mechanisms in this process. Drawing from the insights gained, we propose Mass-Editing Memory with Attention in Transformers (MEMAT), a method that achieves significant improvements in all metrics while requiring minimal parameter modifications. MEMAT delivers a remarkable 10% increase in magnitude metrics, benefits languages not included in the training data and also demonstrates a high degree of portability. Our code and data are at https://github.com/dtamayo-nlp/MEMAT.

Huazheng Wang, Haifeng Sun, Jingyu Wang, Qi Qi, Zixuan Xia, Menghao Zhang, Jianxin Liao

Language Models (LMs) acquire factual knowledge during pre-training and store it in the parameters, which can be valuable for downstream tasks. As world evolves, some facts may be incorrectly induced or become obsolete over time. Various model editing methods have been proposed to modify specific examples in LMs. However, existing training-based methods still suffer from sub-optimal locality, where irrelevant neighborhood examples can be adversely influenced. Model’s gradients are still struggling to identify the appropriate direction when updating the parameters. To address this issue, we find that directing the hidden state of the edit example towards spaces where semantics are sparse tends to help preserve the semantics of irrelevant neighborhood examples. Based on this hypothesis, we propose a novel metric, named SSS, to evaluate the degree of sparsity around a sentence embedding in the semantic space without any human or machine annotation. Subsequently, we incorporate SSS into the original loss function of the existing training-based methods to enhance locality. Experiments conducted on two datasets across various models demonstrate that SSS is effective in improving both locality and reasoning capability.

Fan Gao, Hang Jiang, Rui Yang, Qingcheng Zeng, Jinghui Lu, Moritz Blum, Tianwei She, Yuang Jiang, Irene Li

Educational materials such as survey articles in specialized fields like computer science traditionally require tremendous expert inputs and are therefore expensive to create and update. Recently, Large Language Models (LLMs) have achieved significant success across various general tasks. However, their effectiveness and limitations in the education domain are yet to be fully explored. In this work, we examine the proficiency of LLMs in generating succinct survey articles specific to the niche field of NLP in computer science, focusing on a curated list of 99 topics. Automated benchmarks reveal that GPT-4 surpasses its predecessors, inluding GPT-3.5, PaLM2, and LLaMa2 by margins ranging from 2% to 20% in comparison to the established ground truth. We compare both human and GPT-based evaluation scores and provide in-depth analysis. While our findings suggest that GPT-created surveys are more contemporary and accessible than human-authored ones, certain limitations were observed. Notably, GPT-4, despite often delivering outstanding content, occasionally exhibited lapses like missing details or factual errors. At last, we compared the rating behavior between humans and GPT-4 and found systematic bias in using GPT evaluation.