论文检索

Adil Soubki, John Murzaku, Arash Yousefi Jordehi, Peter Zeng, Magdalena Markowska, Seyed Abolghasem Mirroshandel, Owen Rambow

Evaluating the theory of mind (ToM) capabilities of language models (LMs) has recently received a great deal of attention. However, many existing benchmarks rely on synthetic data, which risks misaligning the resulting experiments with human behavior. We introduce the first ToM dataset based on naturally occurring spoken dialogs, Common-ToM, and show that LMs struggle to demonstrate ToM. We then show that integrating a simple, explicit representation of beliefs improves LM performance on Common-ToM.

Zixia Jia, Mengmeng Wang, Baichen Tong, Song-Chun Zhu, Zilong Zheng

Recent advances in Large Language Models (LLMs) have shown inspiring achievements in constructing autonomous agents that rely onlanguage descriptions as inputs. However, it remains unclear how well LLMs can function as few-shot or zero-shot embodied agents in dynamic interactive environments. To address this gap, we introduce LangSuit·E, a versatile and simulation-free testbed featuring 6 representative embodied tasks in textual embodied worlds. Compared with previous LLM-based testbeds, LangSuit·E (i) offers adaptability to diverse environments without multiple simulation engines, (ii) evaluates agents’ capacity to develop “internalized world knowledge” with embodied observations, and (iii) allows easy customization of communication and action strategies. To address the embodiment challenge, we devise a novel chain-of-thought (CoT) schema, EmMem, which summarizes embodied states w.r.t. history information. Comprehensive benchmark results illustrate challenges and insights of embodied planning. LangSuit·E represents a significant step toward building embodied generalists in the context of language models.

Noah Wang, Z.Y. Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang 等

The advent of Large Language Models (LLMs) has paved the way for complex tasks such as role-playing, which enhances user interactions by enabling models to imitate various characters. However, the closed-source nature of state-of-the-art LLMs and their general-purpose training limit role-playing optimization. In this paper, we introduce RoleLLM, a framework to benchmark, elicit, and enhance role-playing abilities in LLMs. RoleLLM comprises four stages: (1) Role Profile Construction for 100 roles; (2) Context-Based Instruction Generation (Context-Instruct) for role-specific knowledge extraction; (3) Role Prompting using GPT (RoleGPT) for speaking style imitation; and (4) Role-Conditioned Instruction Tuning (RoCIT) for fine-tuning open-source models along with role customization. By Context-Instruct and RoleGPT, we create RoleBench, the first systematic and fine-grained character-level benchmark dataset for role-playing with 168,093 samples. Moreover, RoCIT on RoleBench yields RoleLLaMA (English) and RoleGLM (Chinese), significantly enhancing role-playing abilities and even achieving comparable results with RoleGPT (using GPT-4).

Jaap Jumelet, Willem Zuidema, Arabella Sinclair

We explore which linguistic factors—at the sentence and token level—play an important role in influencing language model predictions, and investigate whether these are reflective of results found in humans and human corpora (Gries and Kootstra, 2017). We make use of the structural priming paradigm—where recent exposure to a structure facilitates processing of the same structure—to investigate where priming effects manifest, and what factors predict them. We find these effects can be explained via the inverse frequency effect found in human priming, where rarer elements within a prime increase priming effects, as well as lexical dependence between prime and target. Our results provide an important piece in the puzzle of understanding how properties within their context affect structural prediction in language models.

Rongwu Xu, Zehan Qi, Wei Xu

Large language models (LLMs) showcase impressive reasoning capabilities when coupled with Chain-of-Thought (CoT) prompting. However, the robustness of this approach warrants further investigation. In this paper, we introduce a novel scenario termed preemptive answers, where the LLM obtains an answer before engaging in reasoning. This situation can arise inadvertently or induced by malicious users by prompt injection attacks. Experiments reveal that preemptive answers significantly impair the model’s reasoning capability across various CoT methods and a broad spectrum of datasets. To bolster the robustness of reasoning, we propose two measures aimed at mitigating this issue to some extent.

Da Ma, Lu Chen, Pengyu Wang, Hongshen Xu, Hanqi Li, Liangtai Sun, Su Zhu, Shuai Fan, Kai Yu

Large language models (LLMs) have demonstrated proficiency across various natural language processing (NLP) tasks but often require additional training, such as continual pre-training and supervised fine-tuning. However, the costs associated with this, primarily due to their large parameter count, remain high. This paper proposes leveraging sparsity in pre-trained LLMs to expedite this training process. By observing sparsity in activated neurons during forward iterations, we identify the potential for computational speed-ups by excluding inactive neurons. We address associated challenges by extending existing neuron importance evaluation metrics and introducing a ladder omission rate scheduler. Our experiments on Llama-2 demonstrate that Sparsity-Accelerated Training (SAT) achieves comparable or superior performance to standard training while significantly accelerating the process. Specifically, SAT achieves a 45% throughput improvement in continual pre-training and saves 38% training time in supervised fine-tuning. It offers a simple, hardware-agnostic, and easily deployable framework for additional LLM training.

Yintao Tai, Xiyang Liao, Alessandro Suglia, Antonio Vergari

Recent work showed the possibility of building open-vocabulary large language models (LLMs) that directly operate on pixel representations. These models are implemented as autoencoders that reconstruct masked patches of rendered text.However, these pixel-based LLMs are limited to discriminative tasks (e.g., classification) and, similar to BERT, cannot be used to generate text.Therefore, they cannot be used for generative tasks such as free-form question answering. In this work, we introduce PIXAR, the first pixel-based autoregressive LLM that performs text generation. Consisting of only a decoder, PIXAR can perform free-form generative tasks while keeping the number of parameters on par with previous encoder-decoder models.Furthermore, we highlight the challenges of generating text as non-noisy images and show this is due to using a maximum likelihood objective. To overcome this problem, we propose an adversarial pretraining stage that improves the readability and accuracy of PIXAR by 8.1 on LAMBADA and 8.5 on bAbI— making it comparable to GPT-2 on text generation tasks.This paves the way to build open-vocabulary LLMs that operate on perceptual input only and calls into question the necessity of the usual symbolic input representation, i.e., text as (sub)tokens.

Jiayu Song, Jenny Chim, Adam Tsakalidis, Julia Ive, Dana Atzil-Slonim, Maria Liakata

We introduce a hybrid abstractive summarisation approach combining hierarchical VAEs with LLMs to produce clinically meaningful summaries from social media user timelines, appropriate for mental health monitoring. The summaries combine two different narrative points of view: (a) clinical insights in third person, generated by feeding into an LLM clinical expert-guided prompts, and importantly, (b) a temporally sensitive abstractive summary of the user’s timeline in first person, generated by a novel hierarchical variational autoencoder, TH-VAE. We assess the generated summaries via automatic evaluation against expert summaries and via human evaluation with clinical experts, showing that timeline summarisation by TH-VAE results in more factual and logically coherent summaries rich in clinical utility and superior to LLM-only approaches in capturing changes over time.

Matteo Gabburo, Nicolaas Paul Jedema, Siddhant Garg, Leonardo F. R. Ribeiro, Alessandro Moschitti

In this paper, we investigate which questions are challenging for retrieval-based Question Answering (QA). We (i) propose retrieval complexity (RC), a novel metric conditioned on the completeness of retrieved documents, which measures the difficulty of answering questions, and (ii) propose an unsupervised pipeline to measure RC given an arbitrary retrieval system.Our proposed pipeline measures RC more accurately than alternative estimators, including LLMs, on six challenging QA benchmarks. Further investigation reveals that RC scores strongly correlate with both QA performance and expert judgment across five of the six studied benchmarks, indicating that RC is an effective measure of question difficulty.Subsequent categorization of high-RC questions shows that they span a broad set of question shapes, including multi-hop, compositional, and temporal QA, indicating that RC scores can categorize a new subset of complex questions. Our system can also have a major impact on retrieval-based systems by helping to identify more challenging questions on existing datasets.

Amin Abolghasemi, Zhaochun Ren, Arian Askari, Mohammad Aliannejadi, Maarten de Rijke, Suzan Verberne

An important unexplored aspect in previous work on user satisfaction estimation for Task-Oriented Dialogue (TOD) systems is their evaluation in terms of robustness for the identification of user dissatisfaction: current benchmarks for user satisfaction estimation in TOD systems are highly skewed towards dialogues for which the user is satisfied. The effect of having a more balanced set of satisfaction labels on performance is unknown. However, balancing the data with more dissatisfactory dialogue samples requires further data collection and human annotation, which is costly and time-consuming. In this work, we leverage large language models (LLMs) and unlock their ability to generate satisfaction-aware counterfactual dialogues to augment the set of original dialogues of a test collection. We gather human annotations to ensure the reliability of the generated samples. We evaluate two open-source LLMs as user satisfaction estimators on our augmented collection against state-of-the-art fine-tuned models. Our experiments show that when used as few-shot user satisfaction estimators, open-source LLMs show higher robustness to the increase in the number of dissatisfaction labels in the test collection than the fine-tuned state-of-the-art models. Our results shed light on the need for data augmentation approaches for user satisfaction estimation in TOD systems. We release our aligned counterfactual dialogues, which are curated by human annotation, to facilitate further research on this topic.

Armineh Nourbakhsh, Sameena Shah, Carolyn Rose

The field of multimodal document understanding has produced a suite of models that have achieved stellar performance across several tasks, even coming close to human performance on certain benchmarks. Nevertheless, the application of these models to real-world enterprise datasets remains constrained by a number of limitations. In this position paper, we discuss these limitations in the context of three key aspects of research: dataset curation, model development, and evaluation on downstream tasks. By analyzing 14 datasets and 7 SotA models, we identify major gaps in their utility in the context of a real-world scenario. We demonstrate how each limitation impedes the widespread use of SotA models in enterprise settings, and present a set of research challenges that are motivated by these limitations. Lastly, we propose a research agenda that is aimed at driving the field towards higher impact in enterprise applications.

Jixiang Hong, Quan Tu, Changyu Chen, Gao Xing, Ji Zhang, Rui Yan

Language models trained on large-scale corpus often generate harmful responses that are harmful and contrary to human values. A prevalent approach for human alignment is reinforcement learning from human feedback (RLHF), utilizing algorithms such as proximal policy optimization (PPO). However, these methods are often characterized by complexity, instability, and substantial resource consumption. Considering that existing large language models (LLMs) like ChatGPT are already relatively well-aligned and cost-friendly, researchers propose to align the language model with human preferences from AI feedback. Nevertheless, the common practices, that unidirectionally distill the responses, are constrained by the inherent capability of LLMs. To address it, we introduce CycleAlign, a framework that distills alignment capabilities from the parameter-invisible LLMs (black-box) to the parameter-visible models (white-box) in an iterative manner. CycleAlign iteratively improves both the white-box and black-box models by integrating static and dynamic in-context learning and a belief alignment method.Empirical results illustrate that the model fine-tuned by CycleAlign remarkably exceeds existing methods, and achieves the state-of-the-art performance in alignment with human value.

Jingcong Liang, Rong Ye, Meng Han, Ruofei Lai, Xinyu Zhang, Xuanjing Huang, Zhongyu Wei

How can we construct an automated debate judge to evaluate an extensive, vibrant, multi-turn debate? This task is challenging, as judging a debate involves grappling with lengthy texts, intricate argument relationships, and multi-dimensional assessments.At the same time, current research mainly focuses on short dialogues, rarely touching upon the evaluation of an entire debate.In this paper, by leveraging Large Language Models (LLMs), we propose Debatrix, which makes the analysis and assessment of multi-turn debates more aligned with majority preferences. Specifically, Debatrix features a vertical, iterative chronological analysis and a horizontal, multi-dimensional evaluation collaboration.To align with real-world debate scenarios, we introduced the PanelBench benchmark, comparing our system’s performance to actual debate outcomes.The findings indicate a notable enhancement over directly using LLMs for debate evaluation.Source code and benchmark data are available at https://github.com/ljcleo/debatrix.

Zhiqiang Zhong, Kuangyu Zhou, Davide Mottin

As Machine Learning (ML) models grow in size and demand higher-quality training data, the expenses associated with re-training and fine-tuning these models are escalating rapidly. Inspired by recent impressive achievements of Large Language Models (LLMs) in different fields, this paper delves into the question: can LLMs efficiently improve an ML’s performance at a minimal cost? We show that, through our proposed training-free framework LLMCorr, an LLM can work as a post-hoc corrector to propose corrections for the predictions of an arbitrary ML model. In particular, we form a contextual knowledge database by incorporating the dataset’s label information and the ML model’s predictions on the validation dataset. Leveraging the in-context learning capability of LLMs, we ask the LLM to summarise the instances in which the ML model makes mistakes and the correlation between primary predictions and true labels. Following this, the LLM can transfer its acquired knowledge to suggest corrections for the ML model’s predictions. Our experimental results on text analysis and the challenging molecular predictions show that LLMCorr improves the performance of a number of models by up to 39%.

Zhu Liu, Cunliang Kong, Ying Liu, Maosong Sun

Large language models have achieved remarkable success in general language understanding tasks. However, as a family of generative methods with the objective of next token prediction, the semantic evolution with the depth of these models are not fully explored, unlike their predecessors, such as BERT-like architectures. In this paper, we specifically investigate the bottom-up evolution of lexical semantics for a popular LLM, namely Llama2, by probing its hidden states at the end of each layer using a contextualized word identification task. Our experiments show that the representations in lower layers encode lexical semantics, while the higher layers, with weaker semantic induction, are responsible for prediction. This is in contrast to models with discriminative objectives, such as mask language modeling, where the higher layers obtain better lexical semantics. The conclusion is further supported by the monotonic increase in performance via the hidden states for the last meaningless symbols, such as punctuation, in the prompting strategy. Our codes are available at https://github.com/RyanLiut/LLM_LexSem.

Akari Haga, Saku Sugawara, Akiyo Fukatsu, Miyu Oba, Hiroki Ouchi, Taro Watanabe, Yohei Oseki

The imitation of the children’s language acquisition process has been explored to make language models (LMs) more efficient.In particular, errors caused by children’s regularization (so-called overregularization, e.g., using wroted for the past tense of write) have been widely studied to reveal the mechanisms of language acquisition. Existing research has analyzed regularization in language acquisition only by modeling word inflection directly, which is unnatural in light of human language acquisition. In this paper, we hypothesize that language models that imitate the errors children make during language acquisition have a learning process more similar to humans. To verify this hypothesis, we analyzed the learning curve and error preferences of verb inflections in small-scale LMs using acceptability judgments. We analyze the differences in results by model architecture, data, and tokenization. Our model shows child-like U-shaped learning curves clearly for certain verbs, but the preferences for types of overgeneralization did not fully match the observations in children.

Yanda Li, Chi Zhang, Gang Yu, Wanqi Yang, Zhibin Wang, Bin Fu, Guosheng Lin, Chunhua Shen, Ling Chen, Yunchao Wei

The remarkable multimodal capabilities demonstrated by OpenAI’s GPT-4 have sparked significant interest in the development of multimodal Large Language Models (LLMs). A primary research objective of such models is to align visual and textual modalities effectively while comprehending human instructions.Current methodologies often rely on annotations derived from benchmark datasets to construct image-dialogue datasets for training purposes, akin to instruction tuning in LLMs. However, these datasets often exhibit domain bias, potentially constraining the generative capabilities of the models. In an effort to mitigate these limitations, we propose a novel data collection methodology that synchronously synthesizes images and dialogues for visual instruction tuning. This approach harnesses the power of generative models, marrying the abilities of ChatGPT and text-to-image generative models to yield a diverse and controllable dataset with varied image content. This not only provides greater flexibility compared to existing methodologies but also significantly enhances several model capabilities. Our research includes comprehensive experiments conducted on various datasets using the open-source LLAVA model as a testbed for our proposed pipeline. Our results underscore marked enhancements across more than ten commonly assessed capabilities.

Meishan Zhang, Hao Fei, Bin Wang, Shengqiong Wu, Yixin Cao, Fei Li, Min Zhang

In the field of information extraction (IE), tasks across a wide range of modalities and their combinations have been traditionally studied in isolation, leaving a gap in deeply recognizing and analyzing cross-modal information. To address this, this work for the first time introduces the concept of grounded Multimodal Universal Information Extraction (MUIE), providing a unified task framework to analyze any IE tasks over various modalities, along with their fine-grained groundings. To tackle MUIE, we tailor a multimodal large language model (MLLM), Reamo, capable of extracting and grounding information from all modalities, i.e., recognizing everything from all modalities at once. Reamo is updated via varied tuning strategies, equipping it with powerful capabilities for information recognition and fine-grained multimodal grounding. To address the absence of a suitable benchmark for grounded MUIE, we curate a high-quality, diverse, and challenging test set, which encompasses IE tasks across 9 common modality combinations with the corresponding multimodal groundings. The extensive comparison of Reamo with existing MLLMs integrated into pipeline approaches demonstrates its advantages across all evaluation dimensions, establishing a strong benchmark for the follow-up research. Our resources are publicly released at https://haofei.vip/MUIE.

Huajian Zhang, Laura Perez-Beltrachini

Synthetically created Cross-Lingual Summarisation (CLS) datasets are prone to include document-summary pairs where the reference summary is unfaithful to the corresponding document as it contains content not supported by the document (i.e., hallucinated content). This low data quality misleads model learning and obscures evaluation results. Automatic ways to assess hallucinations and improve training have been proposed for monolingual summarisation, predominantly in English. For CLS, we propose to use off-the-shelf cross-lingual Natural Language Inference (X-NLI) to evaluate faithfulness of reference and model generated summaries. Then, we study training approaches that are aware of faithfulness issues in the training data and propose an approach that uses unlikelihood loss to teach a model about unfaithful summary sequences. Our results show that it is possible to train CLS models that yield more faithful and at the same time informative summaries.

Sheza Munir, Wassay Sajjad, Mukeet Raza, Emaan Abbas, Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza

Deepfakes, particularly in the auditory domain, have become a significant threat, necessitating the development of robust countermeasures. This paper addresses the escalating challenges posed by deepfake attacks on Automatic Speaker Verification (ASV) systems. We present a novel Urdu deepfake audio dataset for deepfake detection, focusing on two spoofing attacks – Tacotron and VITS TTS. The dataset construction involves careful consideration of phonemic cover and balance and comparison with existing corpora like PRUS and PronouncUR. Evaluation with AASIST-L model shows EERs of 0.495 and 0.524 for VITS TTS and Tacotron-generated audios, respectively, with variability across speakers. Further, this research implements a detailed human evaluation, incorporating a user study to gauge whether people are able to discern deepfake audios from real (bonafide) audios. The ROC curve analysis shows an area under the curve (AUC) of 0.63, indicating that individuals demonstrate a limited ability to detect deepfakes (approximately 1 in 3 fake audio samples are regarded as real). Our work contributes a valuable resource for training deepfake detection models in low-resource languages like Urdu, addressing the critical gap in existing datasets. The dataset is publicly available at: https://github.com/CSALT-LUMS/urdu-deepfake-dataset.