In the rapidly advancing field of artificial intelligence, the concept of ‘Red-Teaming’ or ‘Jailbreaking’ large language models (LLMs) has emerged as a crucial area of study. This approach is especially significant in terms of assessing and enhancing the safety and robustness of these models. This paper investigates the intricate consequences of such modifications through model editing, uncovering a complex relationship between enhancing model accuracy and preserving its ethical integrity. Our in-depth analysis reveals a striking paradox: while injecting accurate information is crucial for model reliability, it can paradoxically destabilize the model’s foundational framework, resulting in unpredictable and potentially unsafe behaviors. Additionally, we propose a benchmark dataset NicheHazardQA to investigate this unsafe behavior both within the same and cross topical domain. This aspect of our research sheds light on how the edits, impact the model’s safety metrics and guardrails. Our findings show that model editing serves as a cost-effective tool for topical red-teaming by methodically applying targeted edits and evaluating the resultant model behavior.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
In conversational AI, personalizing dialogues with persona profiles and contextual understanding is essential. Despite large language models’ (LLMs) improved response coherence, effective persona integration remains a challenge. In this work, we first study two common approaches for personalizing LLMs: textual prompting and direct fine-tuning. We observed that textual prompting often struggles to yield responses that are similar to the ground truths in datasets, while direct fine-tuning tends to produce repetitive or overly generic replies. To alleviate those issues, we propose **S**elective **P**rompt **T**uning (SPT), which softly prompts LLMs for personalized conversations in a selective way. Concretely, SPT initializes a set of soft prompts and uses a trainable dense retriever to adaptively select suitable soft prompts for LLMs according to different input contexts, where the prompt retriever is dynamically updated through feedback from the LLMs. Additionally, we propose context-prompt contrastive learning and prompt fusion learning to encourage the SPT to enhance the diversity of personalized conversations. Experiments on the CONVAI2 dataset demonstrate that SPT significantly enhances response diversity by up to 90%, along with improvements in other critical performance indicators. Those results highlight the efficacy of SPT in fostering engaging and personalized dialogue generation. The SPT model code is [publicly available](https://github.com/hqsiswiliam/SPT) for further exploration.
Instruction tuning is critical to large language models (LLMs) for achieving better instruction following and task adaptation capabilities but its success heavily relies on the training data quality. Many recent methods focus on improving the data quality but often overlook the compatibility of the data with the student model being finetuned. This paper introduces Selective Reflection-Tuning, a novel paradigm that synergizes a teacher LLM’s reflection and introspection for improving existing data quality with the data selection capability of the student LLM, to automatically refine existing instruction-tuning data. This teacher-student collaboration produces high-quality and student-compatible instruction-response pairs, resulting in sample-efficient instruction tuning and LLMs of superior performance. Selective Reflection-Tuning is a data augmentation and synthesis that generally improves LLM finetuning and self-improvement without collecting brand-new data. We apply our method to Alpaca and WizardLM data and achieve much stronger and top-tier 7B and 13B LLMs.
Label-aware Hard Negative Sampling Strategies with Momentum Contrastive Learning for Implicit Hate Speech Detection
PDF ↗Detecting implicit hate speech that is not directly hateful remains a challenge. Recent research has attempted to detect implicit hate speech by applying contrastive learning to pre-trained language models such as BERT and RoBERTa, but the proposed models still do not have a significant advantage over cross-entropy loss-based learning. We found that contrastive learning based on randomly sampled batch data does not encourage the model to learn hard negative samples. In this work, we propose Label-aware Hard Negative sampling strategies (LAHN) that encourage the model to learn detailed features from hard negative samples, instead of naive negative samples in random batch, using momentum-integrated contrastive learning. LAHN outperforms the existing models for implicit hate speech detection both in- and cross-datasets. The code is available at https://github.com/Hanyang-HCC-Lab/LAHN
Can LLMs Speak For Diverse People? Tuning LLMs via Debate to Generate Controllable Controversial Statements
PDF ↗Making LLMs speak for different, especially minority groups of people, and generate statements supporting their diverse or even controversial perspectives is critical to creating an inclusive environment. However, existing LLMs lack sufficient controllability to the stance of their generated content, which often contains inconsistent, neutral, or biased statements. In this paper, we improve the controllability of LLMs in generating statements supporting an argument the user defined in the prompt. We find that multi-round debates between two LLMs with opposite stances generate higher-quality and more salient statements for each, which are important training data to improve the controllability of LLMs. Motivated by this, we develop a novel debate & tuning (“DEBATUNE”) pipeline finetuning LLMs to generate the statements obtained via debate. To examine DEBATUNE, we curate the largest dataset of debate topics so far, which covers 710 controversial topics and corresponding arguments for each topic. Evaluations by the GPT-4 judge with a novel controversy controllability metric show that LLMs’ capability of generating diverse perspectives is significantly improved by DEBATUNE. Moreover, such controllability can be generalized to unseen topics, generating high-quality statements supporting controversial arguments.
Temporal Knowledge Graph (TKG) forecasting aims to predict future facts based on given histories. Most recent graph-based models excel at capturing structural information within TKGs but lack semantic comprehension abilities. Nowadays, with the surge of LLMs, the LLM-based TKG prediction model has emerged. However, the existing LLM-based model exhibits three shortcomings: (1) It only focuses on the first-order history for prediction while ignoring high-order historical information, resulting in the provided information for LLMs being extremely limited. (2) LLMs struggle with optimal reasoning performance under heavy historical information loads. (3) For TKG prediction, the temporal reasoning capability of LLM alone is limited. To address the first two challenges, we propose Chain-of-History (CoH) reasoning which explores high-order histories step-by-step, achieving effective utilization of high-order historical information for LLMs on TKG prediction. To address the third issue, we design CoH as a plug-and-play module to enhance the performance of graph-based models for TKG prediction. Extensive experiments on three datasets and backbones demonstrate the effectiveness of CoH.
Social bias is shaped by the accumulation of social perceptions towards targets across various demographic identities. To fully understand such social bias in large language models (LLMs), it is essential to consider the composite of social perceptions from diverse perspectives among identities. Previous studies have either evaluated biases in LLMs by indirectly assessing the presence of sentiments towards demographic identities in the generated text or measuring the degree of alignment with given stereotypes. These methods have limitations in directly quantifying social biases at the level of distinct perspectives among identities. In this paper, we aim to investigate how social perceptions from various viewpoints contribute to the development of social bias in LLMs. To this end, we propose a novel strategy to intuitively quantify these social perceptions and suggest metrics that can evaluate the social biases within LLMs by aggregating diverse social perceptions. The experimental results show the quantitative demonstration of the social attitude in LLMs by examining social perception. The analysis we conducted shows that our proposed metrics capture the multi-dimensional aspects of social bias, enabling a fine-grained and comprehensive investigation of bias in LLMs.
Leveraging external knowledge is crucial for achieving high performance in knowledge-intensive tasks, such as question answering. The retrieve-and-read approach is widely adopted for integrating external knowledge into a language model. However, this approach suffers from increased computational cost and latency due to the long context length, which grows proportionally with the number of retrieved knowledge. Furthermore, existing retrieval-augmented models typically retrieve information from a single type of knowledge source, limiting their scalability to diverse knowledge sources with varying structures. In this work, we introduce an efficient memory-augmented transformer called MATTER, designed to retrieve relevant knowledge from multiple heterogeneous knowledge sources. Specifically, our model retrieves and reads from both unstructured sources (paragraphs) and semi-structured sources (QA pairs) in the form of fixed-length neural memories. We demonstrate that our model outperforms existing efficient retrieval-augmented models on popular QA benchmarks in terms of both accuracy and speed. Furthermore, MATTER achieves competitive results compared to conventional read-and-retrieve models while having 100x throughput during inference.
Self-correction techniques have recently emerged as a promising framework to improve the quality of responses generated by large language models (LLMs). Few-shot prompted LLMs act as critics to produce feedback for an input, which is further fed to a refiner (also an LLM) to produce an output. However, these critique-refine steps require multiple expensive LLM calls. To circumvent this large inference cost, we borrow inspiration from prior work on knowledge distillation and propose the use of critique distillation to train critic models. These are smaller sequence-to-sequence models that are trained on input-critique pairs generated by an LLM. We focus on the problem of text simplification for three Indian languages: Hindi, Bengali and Marathi. This task is a good fit for self-correction style techniques. It also hasn’t been systematically explored for Indian languages before. We train two separate critics that focus on lexical and structure complexity, and show that it is surprisingly more effective than using an LLM directly as a critic in both 0-shot and few-shot settings. We also show the benefits of training multilingual critics, as opposed to monolingual critics. Extensive human evaluations show that on average, raters find 80% of DIMSIM’s output to be simple and easy to read.
Unveiling the Spectrum of Data Contamination in Language Model: A Survey from Detection to Remediation
PDF ↗Data contamination has garnered increased attention in the era of Large language models (LLMs) due to the reliance on extensive internet-derived training corpora. The issue of training corpus overlap with evaluation benchmarks—referred to as contamination—has been the focus of significant recent research. This body of work aims to identify contamination, understand its impacts, and explore mitigation strategies from diverse perspectives. However, comprehensive studies that provide a clear pathway from foundational concepts to advanced insights are lacking in this nascent field. Therefore, we present the first survey in the field of data contamination. We begin by examining the effects of data contamination across various stages and forms. We then provide a detailed analysis of current contamination detection methods, categorizing them to highlight their focus, assumptions, strengths, and limitations. We also discuss mitigation strategies, offering a clear guide for future research. This survey serves as a succinct overview of the most recent advancements in data contamination research, providing a straightforward guide for the benefit of future research endeavors.
Exploring Reasoning Biases in Large Language Models Through Syllogism: Insights from the NeuBAROCO Dataset
PDF ↗This paper explores the question of how accurately current large language models can perform logical reasoning in natural language, with an emphasis on whether these models exhibit reasoning biases similar to humans. Specifically, our study focuses on syllogistic reasoning, a form of deductive reasoning extensively studied in cognitive science as a natural form of human reasoning. We present a syllogism dataset called NeuBAROCO, which consists of syllogistic reasoning problems in English and Japanese. This dataset was originally designed for psychological experiments to assess human reasoning capabilities using various forms of syllogisms. Our experiments with leading large language models indicate that these models exhibit reasoning biases similar to humans, along with other error tendencies. Notably, there is significant room for improvement in reasoning problems where the relationship between premises and hypotheses is neither entailment nor contradiction. We also present experimental results and in-depth analysis using a new Chain-of-Thought prompting method, which asks LLMs to translate syllogisms into abstract logical expressions and then explain their reasoning process. Our analysis using this method suggests that the primary limitations of LLMs lie in the reasoning process itself rather than the interpretation of syllogisms.
A Large Collection of Model-generated Contradictory Responses for Consistency-aware Dialogue Systems
PDF ↗Mitigating the generation of contradictory responses poses a substantial challenge in dialogue response generation. The quality and quantity of available contradictory response data play a vital role in suppressing these contradictions, offering two significant benefits. First, having access to large contradiction data enables a comprehensive examination of their characteristics. Second, data-driven methods to mitigate contradictions may be enhanced with large-scale contradiction data for training. Nevertheless, no attempt has been made to build an extensive collection of model-generated contradictory responses. In this paper, we build a large dataset of response generation models’ contradictions for the first time. Then, we acquire valuable insights into the characteristics of model-generated contradictions through an extensive analysis of the collected responses. Lastly, we also demonstrate how this dataset substantially enhances the performance of data-driven contradiction suppression methods.
Although many large language models (LLMs) have been trained to refuse harmful requests, they are still vulnerable to jailbreaking attacks which rewrite the original prompt to conceal its harmful intent. In this paper, we propose a new method for defending LLMs against jailbreaking attacks by “backtranslation”. Specifically, given an initial response generated by the target LLM from an input prompt, our backtranslation prompts a language model to infer an input prompt that can lead to the response. The inferred prompt is called the backtranslated prompt which tends to reveal the actual intent of the original prompt, since it is generated based on the LLM’s response and not directly manipulated by the attacker. We then run the target LLM again on the backtranslated prompt, and we refuse the original prompt if the model refuses the backtranslated prompt. We explain that the proposed defense provides several benefits on its effectiveness and efficiency. We empirically demonstrate that our defense significantly outperforms the baselines, in the cases that are hard for the baselines, and our defense also has little impact on the generation quality for benign input prompts. Our implementation is based on our library for LLM jailbreaking defense algorithms at https://github.com/YihanWang617/llm-jailbreaking-defense, and the code for reproducing our experiments is available at https://github.com/YihanWang617/LLM-Jailbreaking-Defense-Backtranslation.
ProgGen: Generating Named Entity Recognition Datasets Step-by-step with Self-Reflexive Large Language Models
PDF ↗Although Large Language Models (LLMs) exhibit remarkable adaptability across domains, these models often fall short in structured knowledge extraction tasks such as named entity recognition (NER). This paper explores an innovative, cost-efficient strategy to harness LLMs with modest NER capabilities for producing superior NER datasets. Our approach diverges from the basic class-conditional prompts by instructing LLMs to self-reflect on the specific domain, thereby generating domain-relevant attributes (such as category and emotions for movie reviews), which are utilized for creating attribute-rich training data. Furthermore, we preemptively generate entity terms and then develop NER context data around these entities, effectively bypassing the LLMs’ challenges with complex structures. Our experiments across both general and niche domains reveal significant performance enhancements over conventional data generation methods while being more cost-effective than existing alternatives.
Improving Multi-hop Logical Reasoning in Knowledge Graphs with Context-Aware Query Representation Learning
PDF ↗Multi-hop logical reasoning on knowledge graphs is a pivotal task in natural language processing, with numerous approaches aiming to answer First-Order Logic (FOL) queries. Recent geometry (e.g., box, cone) and probability (e.g., beta distribution)-based methodologies have effectively addressed complex FOL queries. However, a common challenge across these methods lies in determining accurate geometric bounds or probability parameters for these queries. The challenge arises because existing methods rely on linear sequential operations within their computation graphs, overlooking the logical structure of the query and the relation-induced information that can be gleaned from the relations of the query, which we call the context of the query. To address the problem, we propose a model-agnostic methodology that enhances the effectiveness of existing multi-hop logical reasoning approaches by fully integrating the context of the FOL query graph. Our approach distinctively discerns (1) the structural context inherent to the query structure and (2) the relation-induced context unique to each node in the query graph as delineated in the corresponding knowledge graph. This dual-context paradigm helps nodes within a query graph attain refined internal representations throughout the multi-hop reasoning steps. Through experiments on two datasets, our method consistently enhances the three multi-hop reasoning foundation models, achieving performance improvements of up to 19.5%. Our codes are available at https://github.com/kjh9503/caqr.
Wavelet transforms, a powerful mathematical tool, have been widely used in different domains, including Signal and Image processing, to unravel intricate patterns, enhance data representation, and extract meaningful features from data. Tangible results from their application suggest that Wavelet transforms can be applied to NLP capturing a variety of linguistic and semantic properties.In this paper, we empirically leverage the application of Discrete Wavelet Transforms (DWT) to word and sentence embeddings. We aim to showcase the capabilities of DWT in analyzing embedding representations at different levels of resolution and compressing them while maintaining their overall quality.We assess the effectiveness of DWT embeddings on semantic similarity tasks to show how DWT can be used to consolidate important semantic information in an embedding vector. We show the efficacy of the proposed paradigm using different embedding models, including large language models, on downstream tasks. Our results show that DWT can reduce the dimensionality of embeddings by 50-93% with almost no change in performance for semantic similarity tasks, while achieving superior accuracy in most downstream tasks. Our findings pave the way for applying DWT to improve NLP applications.
The present level of proliferation of fake, biased, and propagandistic content online has made it impossible to fact-check every single suspicious claim or article, either manually or automatically. An increasing number of scholars are focusing on a coarser granularity, aiming to profile entire news outlets, which allows fast identification of potential “fake news” by checking the reliability of their source. Source factuality is also an important element of systems for automatic fact-checking and “fake news” detection, as they need to assess the reliability of the evidence they retrieve online. Political bias detection, which in the Western political landscape is about predicting left-center-right bias, is an equally important topic, which has experienced a similar shift toward profiling entire news outlets. Moreover, there is a clear connection between the two, as highly biased media are less likely to be factual; yet, the two problems have been addressed separately. In this survey, we review the state of the art on media profiling for factuality and bias, arguing for the need to model them jointly. We also shed light on some of the major challenges for modeling bias and factuality jointly. We further discuss interesting recent advances in using different information sources and modalities, which go beyond the text of the articles the target news outlet has published. Finally, we discuss current challenges and outline future research directions.
We propose a simple method that applies a large language model (LLM) to large-scale retrieval in zero-shot scenarios. Our method, the Large language model as Retriever (LameR), is built upon no other neural models but an LLM in a retrieval-augmented retrieval fashion, while breaking brute-force combinations of retrievers with LLMs and lifting the performance of zero-shot retrieval to be very competitive on benchmark datasets. Essentially, we propose to augment a query with its potential answers by prompting LLMs with a composition of the query and the query’s in-domain candidates. The candidates, regardless of correct or wrong, are obtained by a vanilla retrieval procedure on the target collection. As a part of the prompts, they are likely to help LLM generate more precise answers by pattern imitation or candidate summarization. Even if all the candidates are wrong, the prompts at least make LLM aware of in-collection patterns and genres. Moreover, due to the low performance of a self-supervised retriever, the LLM-based query augmentation becomes less effective as the retriever bottlenecks the whole pipeline. Therefore, we propose to leverage a non-parametric lexicon-based method (e.g., BM25) as the retrieval module to capture query-document overlap in a literal fashion. As such, LameR makes the retrieval procedure transparent to the LLM, thus circumventing the bottleneck.
Healthcare Community Question Answering (CQA) forums offer an accessible platform for individuals seeking information on various healthcare-related topics. People find such platforms suitable for self-disclosure, seeking medical opinions, finding simplified explanations for their medical conditions, and answering others’ questions. However, answers on these forums are typically diverse and prone to off-topic discussions. It can be challenging for readers to sift through numerous answers and extract meaningful insights, making answer summarization a crucial task for CQA forums. While several efforts have been made to summarize the community answers, most of them are limited to the open domain and overlook the different perspectives offered by these answers. To address this problem, this paper proposes a novel task of perspective-specific answer summarization. We identify various perspectives, within healthcare-related responses and frame a perspective-driven abstractive summary covering all responses. To achieve this, we annotate 3167 CQA threads with 6193 perspective-aware summaries in our PUMA dataset. Further, we propose PLASMA, a prompt-driven controllable summarization model. To encapsulate the perspective-specific conditions, we design an energy-controlled loss function for the optimization. We also leverage the prefix tuner to learn the intricacies of the healthcare perspective summarization. Our evaluation against five baselines suggests the superior performance of PLASMA by a margin of ~1.5 - 21% improvement. We supplement our experiments with ablation and qualitative analysis.
In this paper, we introduce SCALE, a collaborative framework that connects a compact Specialized Translation Model (STM) and a general-purpose Large Language Model (LLM) as one unified translation engine. By introducing translation from STM into the triplet in-context demonstrations, SCALE unlocks refinement and pivoting ability of LLM, thus 1) mitigating language bias of LLMs and parallel data bias of STMs, 2) enhancing LLM speciality without sacrificing generality, and 3) facilitating continual learning in a LLM-tuning-free way.Our comprehensive experiments show that SCALE significantly outperforms both LLMs (GPT-4, GPT-3.5) and supervised models (NLLB, M2M) in either high-resource or challenging low-resource settings. Moreover SCALE shows great scalability by only updating the lightweight STM and witness consistent system improvement, an averaged 4 BLEURT score across 4 languages without tuning LLM. Interestingly, SCALE could also effectively exploit the existing language bias of LLMs by using an English-centric STM as a pivot to conduct translation between any language pairs, outperforming GPT-4 by an average of 6 COMET points across eight translation directions. Furthermore we provide an in-depth analysis of SCALE’s robustness, translation characteristics, latency costs and inherent language bias, providing solid foundation for future studies exploring the potential synergy between LLMs and more specialized models.