论文检索

Xiaolong Wang, Yile Wang, Yuanchi Zhang, Fuwen Luo, Peng Li, Maosong Sun, Yang Liu

Large Language Models (LLMs) have achieved remarkable performance in objective tasks such as open-domain question answering and mathematical reasoning, which can often be solved through recalling learned factual knowledge or chain-of-thought style reasoning. However, we find that the performance of LLMs in subjective tasks is still unsatisfactory, such as metaphor recognition, dark humor detection, etc. Compared to objective tasks, subjective tasks focus more on interpretation or emotional response rather than a universally accepted reasoning pathway. Based on the characteristics of the tasks and the strong dialogue-generation capabilities of LLMs, we propose RiC (Reasoning in Conversation), a method that focuses on solving subjective tasks through dialogue simulation. The motivation of RiC is to mine useful contextual information by simulating dialogues instead of supplying chain-of-thought style rationales, thereby offering potential useful knowledge behind dialogues for giving the final answers. We evaluate both API-based and open-source LLMs including GPT-4, ChatGPT, and OpenChat across twelve tasks. Experimental results show that RiC can yield significant improvement compared with various baselines.

Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, Alexander Turner

We introduce Contrastive Activation Addition (CAA), a method for steering language models by modifying their activations during forward passes. CAA computes “steering vectors” by averaging the difference in residual stream activations between pairs of positive and negative examples of a particular behavior, such as factual versus hallucinatory responses. During inference, these steering vectors are added at all token positions after the user’s prompt with either a positive or negative coefficient, allowing precise control over the degree of the targeted behavior. We evaluate CAA’s effectiveness on Llama 2 Chat using multiple-choice behavioral question datasets and open-ended generation tasks. We demonstrate that CAA significantly alters model behavior, is effective over and on top of traditional methods like finetuning and system prompt design, and minimally reduces capabilities. Moreover, we gain deeper insights into CAA’s mechanisms by employing various activation space interpretation methods. CAA accurately steers model outputs and sheds light on how high-level concepts are represented in Large Language Models (LLMs).

Roshan Sharma, Suwon Shon, Mark Lindsey, Hira Dhamyal, Bhiksha Raj

Reference summaries for abstractive speech summarization require human annotation, which can be performed by listening to an audio recording or by reading textual transcripts of the recording. In this paper, we examine whether summaries based on annotators listening to the recordings differ from those based on annotators reading transcripts. Using existing intrinsic evaluation based on human evaluation, automatic metrics, LLM-based evaluation, and a retrieval-based reference-free method, we find that summaries are indeed different based on the source modality, and that speech-based summaries are more factually consistent and information-selective than transcript-based summaries. Transcript-based summaries are impacted by recognition errors in the source, and expert-written summaries are more informative and reliable. We make all the collected data and analysis code public to facilitate the reproduction of our work and advance research in this area.

Junda Wu, Tong Yu, Xiang Chen, Haoliang Wang, Ryan Rossi, Sungchul Kim, Anup Rao, Julian McAuley

Large language models (LLMs) often require task-relevant knowledge to augment their internal knowledge through prompts. However, simply injecting external knowledge into prompts does not guarantee that LLMs can identify and use relevant information in the prompts to conduct chain-of-thought reasoning, especially when the LLM’s internal knowledge is derived from biased information on the pretraining data. In this paper, we propose a novel causal view to formally explain the internal knowledge bias of LLMs via a Structural Causal Model (SCM). We review the chain-of-thought (CoT) prompting from a causal perspective and discover that the biased information from pretrained models can impair LLMs’ reasoning abilities. When the CoT reasoning paths are misled by irrelevant information from prompts and are logically incorrect, simply editing factual information is insufficient to reach the correct answer. To estimate the confounding effect on CoT reasoning in LLMs, we use external knowledge as an instrumental variable. We further introduce CoT as a mediator to conduct front-door adjustment and generate logically correct CoTs where the spurious correlation between LLMs’ pretrained knowledge and task queries is reduced. With extensive experiments, we validate that our approach enables more accurate CoT reasoning and enhances LLM generation on knowledge-intensive tasks.

Zihao Lin, Mohammad Beigi, Hongxuan Li, Yufan Zhou, Yuxiang Zhang, Qifan Wang, Wenpeng Yin, Lifu Huang

Memory Editing (ME) has emerged as an efficient method to modify erroneous facts or inject new facts into Large Language Models (LLMs). Two mainstream ME methods exist: parameter-modifying ME and parameter-preserving ME (integrating extra modules while preserving original parameters). Regrettably, previous studies on ME evaluation have two critical limitations: (i) evaluating LLMs with single edit only, neglecting the need for continuous editing, and (ii) evaluations focusing solely on basic factual triples, overlooking broader LLM capabilities like logical reasoning and reading understanding. This study addresses these limitations with contributions threefold: (i) We explore how ME affects a wide range of fundamental capabilities of LLMs under sequential editing. Experimental results reveal an intriguing phenomenon: Most parameter-modifying ME consistently degrade performance across all tasks after a few sequential edits. In contrast, parameter-preserving ME effectively maintains LLMs’ fundamental capabilities but struggles to accurately recall edited knowledge presented in a different format. (ii) We extend our evaluation to different editing settings, such as layers to edit, model size, instruction tuning, etc. Experimental findings indicate several strategies that can potentially mitigate the adverse effects of ME. (iii) We further explain why parameter-modifying damages LLMs from three dimensions: parameter changes after editing, language modeling capability, and the in-context learning capability. Our in-depth study advocates more careful use of ME in real-world scenarios.

Xiaokang Zhang, Zijun Yao, Jing Zhang, Kaifeng Yun, Jifan Yu, Juanzi Li, Jie Tang

This paper proposes PiNose, which trains a probing model on offline self-consistency checking results, thereby circumventing the need for human-annotated data and achieving transferability across diverse data distributions. As the consistency check process is offline, PiNose reduces the computational burden of generating multiple responses by online consistency verification. Additionally, it examines various aspects of internal states prior to response decoding, contributing to more effective detection of factual inaccuracies. Experiment results on both factuality detection and question answering benchmarks show that PiNose achieves surpassing results than existing factuality detection methods.

Rami Aly, Zhiqiang Tang, Samson Tan, George Karypis

Large Language Models (LLMs) frequently hallucinate, impeding their reliability in mission-critical situations. One approach to address this issue is to provide citations to relevant sources alongside generated content, enhancing the verifiability of generations. However, citing passages accurately in answers remains a substantial challenge. This paper proposes a weakly-supervised fine-tuning method leveraging factual consistency models (FCMs). Our approach alternates between generating texts with citations and supervised fine-tuning with FCM-filtered citation data. Focused learning is integrated into the objective, directing the fine-tuning process to emphasise the factual unit tokens, as measured by an FCM. Results on the ALCE few-shot citation benchmark with various instruction-tuned LLMs demonstrate superior performance compared to in-context learning, vanilla supervised fine-tuning, and state-of-the-art methods, with an average improvement of 34.1, 15.5, and 10.5 citation F_1 points, respectively. Moreover, in a domain transfer setting we show that the obtained citation generation ability robustly transfers to unseen datasets. Notably, our citation improvements contribute to the lowest factual error rate across baselines.

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

Temporal knowledge graph completion aims to infer the missing facts in temporal knowledge graphs. Current approaches usually embed factual knowledge into continuous vector space and apply geometric operations to learn potential patterns in temporal knowledge graphs. However, these methods only adopt a single operation, which may have limitations in capturing the complex temporal dynamics present in temporal knowledge graphs. Therefore, we propose a simple but effective method, i.e. TCompoundE, which is specially designed with two geometric operations, including time-specific and relation-specific operations. We provide mathematical proofs to demonstrate the ability of TCompoundE to encode various relation patterns. Experimental results show that our proposed model significantly outperforms existing temporal knowledge graph embedding models. Our code is available at https://github.com/nk-ruiying/TCompoundE.

Junyi Li, Jie Chen, Ruiyang Ren, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, Ji-Rong Wen

In the era of large language models (LLMs), hallucination (the tendency to generate factually incorrect content) poses great challenges to trustworthy and reliable deployment of LLMs in real-world applications. To tackle the hallucination, three key questions should be well studied: how to detect hallucinations (detection), why do LLMs hallucinate (source), and what can be done to mitigate them (mitigation). To address these challenges, this work presents a systematic empirical study on LLM hallucinations, focused on the three aspects of hallucination detection, source and mitigation. Specially, we construct a new hallucination benchmark HaluEval 2.0, and design a simple yet effective detection method for LLM hallucinations. Furthermore, we zoom into the different training or utilization stages of LLMs and extensively analyze the potential factors that lead to the LLM hallucinations. Finally, we implement and examine a series of widely used techniques to mitigate the hallucinations in LLMs. Our work has led to several important findings to understand the hallucination origin and mitigate the hallucinations in LLMs.

Nigel Fernandez, Alexander Scarlatos, Andrew Lan

Automated teaching assistants and chatbots have significant potential to reduce the workload of human instructors, especially for logistics-related question answering, which is important to students yet repetitive for instructors. However, due to privacy concerns, there is a lack of publicly available datasets. We introduce SyllabusQA, an open-source dataset with 63 real course syllabi covering 36 majors, containing 5,078 open-ended course logistics-related question-answer pairs that are diverse in both question types and answer formats. Since many logistics-related questions contain critical information like the date of an exam, it is important to evaluate the factuality of answers. We benchmark several strong baselines on this task, from large language model prompting to retrieval-augmented generation. We introduce Fact-QA, an LLM-based (GPT-4) evaluation metric to evaluate the factuality of predicted answers. We find that despite performing close to humans on traditional metrics of textual similarity, there remains a significant gap between automated approaches and humans in terms of fact precision.

Yakir Yehuda, Itzik Malkiel, Oren Barkan, Jonathan Weill, Royi Ronen, Noam Koenigstein

Despite the many advances of Large Language Models (LLMs) and their unprecedented rapid evolution, their impact and integration into every facet of our daily lives is limited due to various reasons. One critical factor hindering their widespread adoption is the occurrence of hallucinations, where LLMs invent answers that sound realistic, yet drift away from factual truth. In this paper, we present a novel method for detecting hallucinations in large language models, which tackles a critical issue in the adoption of these models in various real-world scenarios. Through extensive evaluations across multiple datasets and LLMs, including Llama-2, we study the hallucination levels of various recent LLMs and demonstrate the effectiveness of our method to automatically detect them. Notably, we observe up to 87% hallucinations for Llama-2 in a specific experiment, where our method achieves a Balanced Accuracy of 81%, all without relying on external knowledge.

Tianjie Ju, Yijin Chen, Xinwei Yuan, Zhuosheng Zhang, Wei Du, Yubin Zheng, Gongshen Liu

Recent work has showcased the powerful capability of large language models (LLMs) in recalling knowledge and reasoning. However, the reliability of LLMs in combining these two capabilities into reasoning through multi-hop facts has not been widely explored. This paper systematically investigates the possibilities for LLMs to utilize shortcuts based on direct connections between the initial and terminal entities of multi-hop knowledge. We first explore the existence of factual shortcuts through Knowledge Neurons, revealing that: (i) the strength of factual shortcuts is highly correlated with the frequency of co-occurrence of initial and terminal entities in the pre-training corpora; (ii) few-shot prompting leverage more shortcuts in answering multi-hop questions compared to chain-of-thought prompting. Then, we analyze the risks posed by factual shortcuts from the perspective of multi-hop knowledge editing. Analysis shows that approximately 20% of the failures are attributed to shortcuts, and the initial and terminal entities in these failure instances usually have higher co-occurrences in the pre-training corpus. Finally, we propose erasing shortcut neurons to mitigate the associated risks and find that this approach significantly reduces failures in multiple-hop knowledge editing caused by shortcuts. Code is publicly available at https://github.com/Jometeorie/MultiHopShortcuts.

Syrine Krichene, Francesco Piccinno, Fangyu Liu, Julian Eisenschlos

Chart-to-summary generation can help explore data, communicate insights, and help the visually impaired people. Multi-modal generative models have been used to produce fluent summaries, but they can suffer from factual and perceptual errors. In this work we present CHATS-CRITIC, a reference-free chart summarization metric for scoring faithfulness. CHATS-CRITIC is composed of an image-to-text model to recover the table from a chart, and a tabular entailment model applied to score the summary sentence by sentence. We find that CHATS-CRITIC evaluates the summary quality according to human ratings better than reference-based metrics, either learned or n-gram based, and can be further used to fix candidate summaries by removing not supported sentences. We then introduce CHATS-PI, a chart-to-summary pipeline that leverages CHATS-CRITIC during inference to fix and rank sampled candidates from any chart-summarization model. We evaluate CHATS-PI and CHATS-CRITIC using human raters, establishing state-of-the-art results on two popular chart-to-summary datasets.

Sebastian Joseph, Lily Chen, Jan Trienes, Hannah Göke, Monika Coers, Wei Xu, Byron Wallace, Junyi Jessy Li

Plain language summarization with LLMs can be useful for improving textual accessibility of technical content. But how factual are these summaries in a high-stakes domain like medicine? This paper presents FactPICO, a factuality benchmark for plain language summarization of medical texts describing randomized controlled trials (RCTs), which are the basis of evidence-based medicine and can directly inform patient treatment. FactPICO consists of 345 plain language summaries of RCT abstracts generated from three LLMs (i.e., GPT-4, Llama-2, and Alpaca), with fine-grained evaluation and natural language rationales from experts. We assess the factuality of critical elements of RCTs in those summaries: Populations, Interventions, Comparators, Outcomes (PICO), as well as the reported findings concerning these. We also evaluate the correctness of the extra information (e.g., explanations) added by LLMs. Using FactPICO, we benchmark a range of existing factuality metrics, including the newly devised ones based on LLMs. We find that plain language summarization of medical evidence is still challenging, especially when balancing between simplicity and factuality, and that existing metrics correlate poorly with expert judgments on the instance level.

Francesco Ortu, Zhijing Jin, Diego Doimo, Mrinmaya Sachan, Alberto Cazzaniga, Bernhard Schölkopf

Interpretability research aims to bridge the gap between the empirical success and our scientific understanding of the inner workings of large language models (LLMs). However, most existing research in this area focused on analyzing a single mechanism, such as how models copy or recall factual knowledge. In this work, we propose the formulation of competition of mechanisms, which instead of individual mechanisms focuses on the interplay of multiple mechanisms, and traces how one of them becomes dominant in the final prediction. We uncover how and where the competition of mechanisms happens within LLMs using two interpretability methods, logit inspection and attention modification. Our findings show traces of the mechanisms and their competition across various model components, and reveal attention positions that effectively control the strength of certain mechanisms.

Parag Jain, Andreea Marzoca, Francesco Piccinno

We consider the task of generating structured representations of text using large language models (LLMs). We focus on tables and mind maps as representative modalities. Tables are more organized way of representing data, while mind maps provide a visually dynamic and flexible approach, particularly suitable for sparse content. Despite the effectiveness of LLMs on different tasks, we show that current models struggle with generating structured outputs. In response, we present effective prompting strategies for both of these tasks. We introduce a taxonomy of problems around factuality, global and local structure, common to both modalities and propose a set of critiques to tackle these issues resulting in an absolute improvement in accuracy of +37pp (79%) for mind maps and +15pp (78%) for tables. To evaluate semantic coverage of generated structured representations we propose Auto-QA, and we verify the adequacy of Auto-QA using SQuAD dataset. We further evaluate the usefulness of structured representations via a text comprehension user study. The results show a significant reduction in comprehension time compared to text when using table (42.9%) and mind map (31.9%), without loss in accuracy.

Giovanni Monea, Maxime Peyrard, Martin Josifoski, Vishrav Chaudhary, Jason Eisner, Emre Kiciman, Hamid Palangi, Barun Patra, Robert West

Large language models (LLMs) have an impressive ability to draw on novel information supplied in their context. Yet the mechanisms underlying this contextual grounding remain unknown, especially in situations where contextual information contradicts factual knowledge stored in the parameters, which LLMs also excel at recalling. Favoring the contextual information is critical for retrieval-augmented generation methods, which enrich the context with up-to-date information, hoping that grounding can rectify outdated or noisy stored knowledge. We present a novel method to study grounding abilities using Fakepedia, a novel dataset of counterfactual texts constructed to clash with a model’s internal parametric knowledge. In this study, we introduce Fakepedia, a counterfactual dataset designed to evaluate grounding abilities when the internal parametric knowledge clashes with the contextual information. We benchmark various LLMs with Fakepedia and conduct a causal mediation analysis of LLM components when answering Fakepedia queries, based on our Masked Grouped Causal Tracing (MGCT) method. Through this analysis, we identify distinct computational patterns between grounded and ungrounded responses. We finally demonstrate that distinguishing grounded from ungrounded responses is achievable through computational analysis alone. Our results, together with existing findings about factual recall mechanisms, provide a coherent narrative of how grounding and factual recall mechanisms interact within LLMs.

Gal Yona, Roee Aharoni, Mor Geva

Factual questions typically can be answered correctly at different levels of granularity. For example, both “August 4, 1961” and “1961” are correct answers to the question “When was Barack Obama born?”. Standard question answering (QA) evaluation protocols, however, do not explicitly take this into account and compare a predicted answer against answers of a single granularity level. In this work, we propose GRANOLA QA, a novel evaluation setting where a predicted answer is evaluated in terms of accuracy and informativeness against a set of multi-granularity answers. We present a simple methodology for enriching existing datasets with multi-granularity answers, and create GRANOLA-EQ, a multi-granularity version of the EntityQuestions dataset. We evaluate a range of decoding methods on GRANOLA-EQ, including a new algorithm, called Decoding with Response Aggregation (DRAG), that is geared towards aligning the response granularity with the model’s uncertainty. Our experiments show that large language models with standard decoding tend to generate specific answers, which are often incorrect. In contrast, when evaluated on multi-granularity answers, DRAG yields a nearly 20 point increase in accuracy on average, which further increases for rare entities. Overall, this reveals that standard evaluation and decoding schemes may significantly underestimate the knowledge encapsulated in LMs.

Zhengbao Jiang, Zhiqing Sun, Weijia Shi, Pedro Rodriguez, Chunting Zhou, Graham Neubig, Xi Victoria Lin, Wen-tau Yih, Srinivasan Iyer

In order for large language model (LLM)-based assistants to effectively adapt to evolving information needs, it must be possible to update their factual knowledge through continued training on new data. The standard recipe for doing so involves continued pre-training on new documents followed by instruction-tuning on question-answer (QA) pairs. However, we find that LLMs trained with this recipe struggle to answer questions, even though the perplexity of documents is minimized. We found that QA pairs are generally straightforward, while documents are more complex, weaving many factual statements together in an intricate manner. Therefore, we hypothesize that it is beneficial to expose LLMs to QA pairs before continued pre-training on documents so that the process of encoding knowledge from complex documents takes into account how this knowledge is accessed through questions. Based on this, we propose pre-instruction-tuning (PIT), a method that instruction-tunes on questions prior to training on documents. This contrasts with standard instruction-tuning, which learns how to extract knowledge after training on documents. Extensive experiments and ablation studies demonstrate that pre-instruction-tuning significantly enhances the ability of LLMs to absorb knowledge from new documents, outperforming standard instruction-tuning by 17.8%.

Jianing Wang, Qiushi Sun, Xiang Li, Ming Gao

Recently, Chain-of-Thought (CoT) prompting has delivered success on complex reasoning tasks, which aims at designing a simple prompt like “Let’s think step by step” or multiple in-context exemplars with well-designed rationales to elicit Large Language Models (LLMs) to generate intermediate reasoning steps. However, the generated rationales often come with hallucinations, making unfactual and unfaithful reasoning chains. To mitigate this brittleness, we propose a novel Chain-of-Knowledge (CoK) prompting, where we aim at eliciting LLMs to generate explicit pieces of knowledge evidence in the form of structure triple. This is inspired by our human behaviors, i.e., we can draw a mind map or knowledge map as the reasoning evidence in the brain before answering a complex question. Benefiting from CoK, we additionally introduce an F2-Verification method to estimate the reliability of the reasoning chains in terms of factuality and faithfulness. For the unreliable response, the wrong evidence can be indicated to prompt the LLM to rethink. Extensive experiments demonstrate that our method can further improve the performance of commonsense, factual, symbolic, and arithmetic reasoning tasks.