The evaluation of natural language generation (NLG) tasks is a significant and longstanding research area. With the recent emergence of powerful large language models (LLMs), some studies have turned to LLM-based automatic evaluation methods, which demonstrate great potential to become a new evaluation paradigm following traditional string-based and model-based metrics. However, despite the improved performance of existing methods, they still possess some deficiencies, such as dependency on references and limited evaluation flexibility. Therefore, in this paper, we meticulously construct a large-scale NLG evaluation corpus **NLG-Eval** with annotations from both human and GPT-4 to alleviate the lack of relevant data in this field. Furthermore, we propose **Themis**, an LLM dedicated to NLG evaluation, which has been trained with our designed multi-perspective consistency verification and rating-oriented preference alignment methods. Themis can conduct flexible and interpretable evaluations without references, and it exhibits superior evaluation performance on various NLG tasks, simultaneously generalizing well to unseen tasks and surpassing other evaluation models, including GPT-4.
论文检索
输入标题、作者或关键词，从 2,271 篇学术成果中精准定位
Mixture-of-Experts (MoE) has gained increasing popularity as a promising framework for scaling up large language models (LLMs). However, training MoE from scratch in a large-scale setting still suffers from data-hungry and instability problems. Motivated by this limit, we investigate building MoE models from existing dense large language models. Specifically, based on the well-known LLaMA-2 7B model, we obtain an MoE model by: (1) Expert Construction, which partitions the parameters of original Feed-Forward Networks (FFNs) into multiple experts; (2) Continual pre-training, which further trains the transformed MoE model and additional gate networks. In this paper, we comprehensively explore different methods for expert construction and various data sampling strategies for continual pre-training. After these stages, our LLaMA-MoE models could maintain language abilities and route the input tokens to specific experts with part of the parameters activated. Empirically, by training 200B tokens, LLaMA-MoE-3.5B models significantly outperform dense models that contain similar activation parameters.
Can LLMs replace Neil deGrasse Tyson? Evaluating the Reliability of LLMs as Science Communicators
PDF ↗Large Language Models (LLMs) and AI assistants driven by these models are experiencing exponential growth in usage among both expert and amateur users. In this work, we focus on evaluating the reliability of current LLMs as science communicators. Unlike existing benchmarks, our approach emphasizes assessing these models on scientific question-answering tasks that require a nuanced understanding and awareness of answerability. We introduce a novel dataset, SCiPS-QA, comprising 742 Yes/No queries embedded in complex scientific concepts, along with a benchmarking suite that evaluates LLMs for correctness and consistency across various criteria. We benchmark three proprietary LLMs from the OpenAI GPT family and 13 open-access LLMs from the Meta Llama-2, Llama-3, and Mistral families. While most open-access models significantly underperform compared to GPT-4 Turbo, our experiments identify Llama-3-70B as a strong competitor, often surpassing GPT-4 Turbo in various evaluation aspects. We also find that even the GPT models exhibit a general incompetence in reliably verifying LLM responses. Moreover, we observe an alarming trend where human evaluators are deceived by incorrect responses from GPT-4 Turbo.
Pretrained language models (PLMs) display impressive performances and have captured the attention of the NLP community.Establishing best practices in pretraining has, therefore, become a major focus of NLP research, especially since insights gained from monolingual English models may not necessarily apply to more complex multilingual models.One significant caveat of the current state of the art is that different works are rarely comparable: they often discuss different parameter counts, training data, and evaluation methodology.This paper proposes a comparison of multilingual pretraining objectives in a controlled methodological environment. We ensure that training data and model architectures are comparable, and discuss the downstream performances across 6 languages that we observe in probing and fine-tuning scenarios.We make two key observations: (1) the architecture dictates which pretraining objective is optimal; (2) multilingual translation is a very effective pretraining objective under the right conditions.We make our code, data, and model weights available at https://github.com/Helsinki-NLP/lm-vs-mt.
The prolific use of Large Language Models (LLMs) as an alternate knowledge base requires them to be factually consistent, necessitating both correctness and consistency traits for paraphrased queries. Recently, significant attempts have been made to benchmark datasets and metrics to evaluate LLMs for these traits. However, structural simplicity (subject-relation-object) and contemporary association in their query formulation limit the broader definition of factuality and consistency. In this study, we introduce TeCFaP, a novel Temporally Consistent Factuality Probe task to expand the consistent factuality probe in the temporal dimension. To this end, we propose TEMP-COFAC, a high-quality dataset of prefix-style English query paraphrases. Subsequently, we extend the definitions of existing metrics to represent consistent factuality across temporal dimension. We experiment with a diverse set of LLMs and find most of them performing poorly on TeCFaP. Next, we propose a novel solution CoTSeLF (Consistent-Time-Sensitive Learning Framework) combining multi-task instruction tuning (MT-IT) with consistent-time-sensitive reinforcement learning (CTSRL) to improve temporally consistent factuality in LLMs. Our experiments demonstrate the efficacy of CoTSeLF over several baselines.
For subjective tasks such as hate detection, where people perceive hate differently, the Large Language Model’s (LLM) ability to represent diverse groups is unclear. By including additional context in prompts, we comprehensively analyze LLM’s sensitivity to geographical priming, persona attributes, and numerical information to assess how well the needs of various groups are reflected. Our findings on two LLMs, five languages, and six datasets reveal that mimicking persona-based attributes leads to annotation variability. Meanwhile, incorporating geographical signals leads to better regional alignment. We also find that the LLMs are sensitive to numerical anchors, indicating the ability to leverage community-based flagging efforts and exposure to adversaries. Our work provides preliminary guidelines and highlights the nuances of applying LLMs in culturally sensitive cases.
Large language models (LLMs) have revolutionized the domain of natural language processing because of their excellent performance on various tasks. Despite their impressive capabilities, LLMs also have the potential to generate texts that pose risks of misuse. Consequently, detecting LLM-generated text has become increasingly important.Previous LLM-generated text detection methods use semantic features, which are stored in the last layer. This leads to methods that overfit the training set domain and exhibit shortcomings in generalization. Therefore, We argue that utilizing intrinsic features rather than semantic features for detection results in better performance.In this work, we design Text Fluoroscopy, a black-box method with better generalizability for detecting LLM-generated text by mining the intrinsic features of the text to be detected. Our method captures the text’s intrinsic features by identifying the layer with the largest distribution difference from the last and first layers when projected to the vocabulary space.Our method achieves 7.36% and 2.84% average improvement in detection performance compared to the baselines in detecting texts from different domains generated by GPT-4 and Claude3, respectively.
Socio-demographic prompting is a commonly employed approach to study cultural biases in LLMs as well as for aligning models to certain cultures. In this paper, we systematically probe four LLMs (Llama 3, Mistral v0.2, GPT-3.5 Turbo and GPT4) with prompts that are conditioned on culturally sensitive and non-sensitive cues, on datasets that are supposed to be culturally sensitive (EtiCor and CALI) or neutral (MMLU and ETHICS). We observe that all models except GPT4 show significant variations in their responses on both kinds of datasets for both kinds of prompts, casting doubt on the robustness of the culturally-conditioned prompting as a method for eliciting cultural bias in models that are not sufficiently stable with respect to arbitrary prompting cues. Further, we also show that some of the supposedly culturally neutral datasets have a non-trivial fraction of culturally sensitive questions/tasks.
Emotion Support Conversation (ESC) is a crucial application, which aims to reduce human stress, offer emotional guidance, and ultimately enhance human mental and physical well-being. With the advancement of Large Language Models (LLMs), many researchers have employed LLMs as the ESC models. However, the evaluation of these LLM-based ESCs remains uncertain. In detail, we first re-organize 2,801 role-playing cards from seven existing datasets to define the roles of the role-playing agent. Second, we train a specific role-playing model called ESC-Role which behaves more like a confused person than GPT-4. Third, through ESC-Role and organized role cards, we systematically conduct experiments using 14 LLMs as the ESC models, including general AI-assistant LLMs (e.g., ChatGPT) and ESC-oriented LLMs (e.g., ExTES-Llama). We conduct comprehensive human annotations on interactive multi-turn dialogues of different ESC models. The results show that ESC-oriented LLMs exhibit superior ESC abilities compared to general AI-assistant LLMs, but there is still a gap behind human performance. Moreover, to automate the scoring process for future ESC models, we developed ESC-RANK, which trained on the annotated data, achieving a scoring performance surpassing 35 points of GPT-4.
We present a survey of more than 90 recent papers that aim to study cultural representation and inclusion in large language models (LLMs). We observe that none of the studies explicitly define “culture, which is a complex, multifaceted concept; instead, they probe the models on some specially designed datasets which represent certain aspects of “culture”. We call these aspects the proxies of culture, and organize them across two dimensions of demographic and semantic proxies. We also categorize the probing methods employed. Our analysis indicates that only certain aspects of “culture,” such as values and objectives, have been studied, leaving several other interesting and important facets, especially the multitude of semantic domains (Thompson et al., 2020) and aboutness (Hershcovich et al., 2022), unexplored. Two other crucial gaps are the lack of robustness of probing techniques and situated studies on the impact of cultural mis- and under-representation in LLM-based applications.
We describe an approach for aligning an LLM based dialogue agent for long-term social dialogue, where there is only a single global score given by the user at the end of the session. In this paper, we propose the usage of denser naturally-occurring multimodal communicative signals as local implicit feedback to improve the turn-level utterance generation. Therefore, our approach (dubbed GELI) learns a local, turn-level reward model by decomposing the human-provided Global Explicit (GE) session level reward, using Local Implicit (LI) multimodal reward signals to crossmodally shape the reward decomposition step. This decomposed reward model is then used as part of the RLHF pipeline to improve an LLM-based dialog agent. We run quantitative and qualitative human studies on two large-scale datasets to evaluate the performance of our GELI approach, and find that it shows consistent improvements across various conversational metrics compared to baseline methods.
Scientific reasoning poses an excessive challenge for even the most advanced Large Language Models (LLMs). To make this task more practical and solvable for LLMs, we introduce a new task setting named tool-augmented scientific reasoning. This setting supplements LLMs with scalable toolsets, and shifts the focus from pursuing an omniscient problem solver to a proficient tool-user. To facilitate the research of such setting, we construct a tool-augmented training corpus named MathFunc which encompasses over 30,000 samples and roughly 6,000 tools. Building on MathFunc, we develop SciAgent to retrieve, understand and, if necessary, use tools for scientific problem solving. Additionally, we craft a benchmark, SciToolBench, spanning five scientific domains to evaluate LLMs’ abilities with tool assistance. Extensive experiments on SciToolBench confirm the effectiveness of SciAgent. Notably, SciAgent-Llama3-8B surpasses other LLMs with the comparable size by more than 8.0% in absolute accuracy. Furthermore, SciAgent-DeepMath-7B shows much superior performance than ChatGPT.
Mitigating the Language Mismatch and Repetition Issues in LLM-based Machine Translation via Model Editing
PDF ↗Large Language Models (LLMs) have recently revolutionized the NLP field, while they still fall short in some specific down-stream tasks. In the work, we focus on utilizing LLMs to perform machine translation, where we observe that two patterns of errors frequently occur and drastically affect the translation quality: language mismatch and repetition. The work sets out to explore the potential for mitigating these two issues by leveraging model editing methods, e.g., by locating Feed-Forward Network (FFN) neurons or something that are responsible for the errors and deactivating them in the inference time.We find that directly applying such methods either limited effect on the targeted errors or has significant negative side-effect on the general translation quality, indicating that the located components may also be crucial for ensuring machine translation with LLMs on the rails.To this end, we propose to refine the located components by fetching the intersection of the locating results under different language settings, filtering out the aforementioned information that is irrelevant to targeted errors. The experiment results empirically demonstrate that our methods can effectively reduce the language mismatch and repetition ratios and meanwhile enhance or keep the general translation quality in most cases.
Images Speak Louder than Words: Understanding and Mitigating Bias in Vision-Language Model from a Causal Mediation Perspective
PDF ↗Vision-language models (VLMs) pre-trained on extensive datasets can inadvertently learn biases by correlating gender information with specific objects or scenarios. Current methods, which focus on modifying inputs and monitoring changes in the model’s output probability scores, often struggle to comprehensively understand bias from the perspective of model components. We propose a framework that incorporates causal mediation analysis to measure and map the pathways of bias generation and propagation within VLMs. Our framework is applicable to a wide range of vision-language and multimodal tasks. In this work, we apply it to the object detection task and implement it on the GLIP model. This approach allows us to identify the direct effects of interventions on model bias and the indirect effects of interventions on bias mediated through different model components. Our results show that image features are the primary contributors to bias, with significantly higher impacts than text features, specifically accounting for 32.57% and 12.63% of the bias in the MSCOCO and PASCAL-SENTENCE datasets, respectively. Notably, the image encoder’s contribution surpasses that of the text encoder and the deep fusion encoder. Further experimentation confirms that contributions from both language and vision modalities are aligned and non-conflicting. Consequently, focusing on blurring gender representations within the image encoder which contributes most to the model bias, reduces bias efficiently by 22.03% and 9.04% in the MSCOCO and PASCAL-SENTENCE datasets, respectively, with minimal performance loss or increased computational demands.
Jailbreak attacks enable malicious queries to evade detection by LLMs. Existing attacks focus on meticulously constructing prompts to disguise harmful intentions. However, the incorporation of sophisticated disguising prompts may incur the challenge of “intention shift”. Intention shift occurs when the additional semantics within the prompt distract the LLMs, causing the responses to deviate significantly from the original harmful intentions. In this paper, we propose a novel component, “bait”, to alleviate the effects of intention shift. Bait comprises an initial response to the harmful query, prompting LLMs to rectify or supplement the knowledge within the bait. By furnishing rich semantics relevant to the query, the bait helps LLMs focus on the original intention. To conceal the harmful content within the bait, we further propose a novel attack paradigm, BaitAttack. BaitAttack adaptively generates necessary components to persuade targeted LLMs that they are engaging with a legitimate inquiry in a safe context. Our proposal is evaluated on a popular dataset, demonstrating state-of-the-art attack performance and an exceptional capability for mitigating intention shift. The implementation of BaitAttack is accessible at: https://anonymous.4open.science/r/BaitAttack-D1F5.
Microblog content (e.g., Tweets) is noisy due to its informal use of language and its lack of contextual information within each post. To tackle these challenges, state-of-the-art microblog classification models rely on pre-training language models (LMs). However, pre-training dedicated LMs is resource-intensive and not suitable for small labs. Supervised contrastive learning (SCL) has shown its effectiveness with small, available resources. In this work, we examine the effectiveness of fine-tuning transformer-based language models, regularized with a SCL loss for English microblog classification. Despite its simplicity, the evaluation on two English microblog classification benchmarks (TweetEval and Tweet Topic Classification) shows an improvement over baseline models. The result shows that, across all subtasks, our proposed method has a performance gain of up to 11.9 percentage points. All our models are open source.
Large language models (LLMs) exhibit enhanced reasoning at larger scales, driving efforts to distill these capabilities into smaller models via teacher-student learning.Previous works simply fine-tune student models on teachers’ generated Chain-of-Thoughts (CoTs) data. Although these methods enhance in-domain (IND) reasoning performance, they struggle to generalize to out-of-domain (OOD) tasks.We believe that the widespread spurious correlations between questions and answers may lead the model to preset a specific answer which restricts the diversity and generalizability of its reasoning process.In this paper, we propose Cascading Decomposed CoTs Distillation (CasCoD) to address these issues by decomposing the traditional single-step learning process into two cascaded learning steps. Specifically, by restructuring the training objectives—removing the answer from outputs and concatenating the question with the rationale as input—CasCoD’s two-step learning process ensures that students focus on learning rationales without interference from the preset answers, thus improving reasoning generalizability. Extensive experiments demonstrate the effectiveness of CasCoD on both IND and OOD benchmark reasoning datasets
The quality is a crucial issue for crowd annotations. Answer aggregation is an important type of solution. The aggregated answers estimated from multiple crowd answers to the same instance are the eventually collected annotations, rather than the individual crowd answers themselves. Recently, the capability of Large Language Models (LLMs) on data annotation tasks has attracted interest from researchers. Most of the existing studies mainly focus on the average performance of individual crowd workers; several recent works studied the scenarios of aggregation on categorical labels and LLMs used as label creators. However, the scenario of aggregation on text answers and the role of LLMs as aggregators are not yet well-studied. In this paper, we investigate the capability of LLMs as aggregators in the scenario of close-ended crowd text answer aggregation. We propose a human-LLM hybrid text answer aggregation method with a Creator-Aggregator Multi-Stage (CAMS) crowdsourcing framework. We make the experiments based on public crowdsourcing datasets. The results show the effectiveness of our approach based on the collaboration of crowd workers and LLMs.
ERVQA: A Dataset to Benchmark the Readiness of Large Vision Language Models in Hospital Environments
PDF ↗The global shortage of healthcare workers has demanded the development of smart healthcare assistants, which can help monitor and alert healthcare workers when necessary. We examine the healthcare knowledge of existing Large Vision Language Models (LVLMs) via the Visual Question Answering (VQA) task in hospital settings through expert annotated open-ended questions. We introduce the Emergency Room Visual Question Answering (ERVQA) dataset, consisting of <image, question, answer> triplets covering diverse emergency room scenarios, a seminal benchmark for LVLMs. By developing a detailed error taxonomy and analyzing answer trends, we reveal the nuanced nature of the task. We benchmark state-of-the-art open-source and closed LVLMs using traditional and adapted VQA metrics: Entailment Score and CLIPScore Confidence. Analyzing errors across models, we infer trends based on properties like decoder type, model size, and in-context examples. Our findings suggest the ERVQA dataset presents a highly complex task, highlighting the need for specialized, domain-specific solutions.
Prior work has explored the possibility of using the semantic information obtained from embedding representations to quantify social stereotypes, leveraging techniques such as word embeddings combined with a list of traits (Garg et al., 2018; Charlesworth et al., 2022) or semantic axes (An et al., 2018; Lucy et al., 2022). However, these approaches have struggled to fully capture the variability in stereotypes across different conceptual domains for the same social group (e.g., black in science, health, and art), in part because the identity of a word and the associations formed during pre-training can dominate its contextual representation (Field and Tsvetkov, 2019). This study explores the ability to recover stereotypes from the contexts surrounding targeted entities by utilizing state-of-the-art text embedding models and adaptive semantic axes enhanced by large language models (LLMs). Our results indicate that the proposed pipeline not only surpasses token-based methods in capturing in-domain framing but also effectively tracks stereotypes over time and along domain-specific semantic axes for in-domain texts. Our research highlights the potential of employing text embedding models to achieve a deeper understanding of nuanced social stereotypes.