EMNLP 2024 论文列表

Calibrating language models (LMs) aligns their generation confidence with the actual likelihood of answer correctness, which can inform users about LMs’ reliability and mitigate hallucinated content. However, prior calibration methods, such as self-consistency-based and logit-based approaches, are either limited in inference-time efficiency or fall short of providing informative signals. Moreover, simply filtering out low-confidence responses reduces the LM’s helpfulness when the answers are correct. Therefore, effectively using calibration techniques to enhance an LM’s factuality remains an unsolved challenge. In this paper, we first propose an activation-based calibration method, ActCab, which trains a linear layer on top of the LM’s last-layer activations that can better capture the representations of knowledge. Built on top of ActCab, we further propose CoDec, a confidence-guided decoding strategy to elicit truthful answers with high confidence from LMs. By evaluating on five popular QA benchmarks, ActCab achieves superior calibration performance than all competitive baselines, e.g., by reducing the average expected calibration error (ECE) score by up to 39%. Further experiments on CoDec show consistent improvements in several LMs’ factuality on challenging QA datasets, such as TruthfulQA, highlighting the value of confidence signals in enhancing the factuality.

In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities across various domains. Within the training pipeline of LLMs, the Reinforcement Learning with Human Feedback (RLHF) phase is crucial for aligning LLMs with human preferences and values. Label smoothing, a technique that replaces hard labels with soft labels, emerges as promising techniques to enhance RLHF training. Despite the benefits, the choice of label smoothing parameters often relies on heuristic approaches and lack theoretical understanding. This paper addresses the challenge of selecting the label smoothing parameter in a principled manner. We introduce Confidence Aware Label Smoothing (CALS), a method that iteratively updates the label smoothing parameter based on preference labels and model forecasts. Our theoretical analysis characterizes the optimal label smoothing parameter, demonstrates its dependence on the confidence level, and reveals its influence on training dynamics and equilibrium. Empirical evaluations on state-of-the-art alignment tasks show that CALS achieves competitive performance, highlighting its potential for improving alignment.

Large language models (LLMs) often struggle with processing extensive input contexts, which can lead to redundant, inaccurate, or incoherent summaries. Recent methods have used unstructured memory to incrementally process these contexts, but they still suffer from information overload due to the volume of unstructured data handled. In our study, we introduce structured knowledge representations (GU_json), which significantly improve summarization performance by 40% and 14% across two public datasets. Most notably, we propose the Chain-of-Key strategy (CoK_json) that dynamically updates or augments these representations with new information, rather than recreating the structured memory for each new source. This method further enhances performance by 7% and 4% on the datasets.

Large language models (LLMs) have demonstrated prominent reasoning capabilities in recommendation tasks by transforming them into text-generation tasks. However, existing approaches either disregard or ineffectively model the user-item high-order interactions. To this end, this paper presents an enhanced LLM-based recommender (ELMRec). We enhance whole-word embeddings to substantially enhance LLMs’ interpretation of graph-constructed interactions for recommendations, without requiring graph pre-training. This finding may inspire endeavors to incorporate rich knowledge graphs into LLM-based recommenders via whole-word embedding. We also found that LLMs often recommend items based on users’ earlier interactions rather than recent ones, and present a reranking solution. Our ELMRec outperforms state-of-the-art (SOTA) methods, especially achieving a 124.3% to 293.7% improvement over SOTA LLM-based methods in direct recommendations. Our code is available online.

Black-box large language models (LLMs) are increasingly deployed in various environments, making it essential for these models to effectively convey their confidence and uncertainty, especially in high-stakes settings. However, these models often exhibit overconfidence, leading to potential risks and misjudgments. Existing techniques for eliciting and calibrating LLM confidence have primarily focused on general reasoning datasets, yielding only modest improvements. Accurate calibration is crucial for informed decision-making and preventing adverse outcomes but remains challenging due to the complexity and variability of tasks these models perform. In this work, we investigate the miscalibration behavior of black-box LLMs within the healthcare setting. We propose a novel method, Atypical Presentations Recalibration, which leverages atypical presentations to adjust the model’s confidence estimates. Our approach significantly improves calibration, reducing calibration errors by approximately 60% on three medical question answering datasets and outperforming existing methods such as vanilla verbalized confidence, CoT verbalized confidence and others. Additionally, we provide an in-depth analysis of the role of atypicality within the recalibration framework.

Fine-grained image classification, especially in zero-/few-shot scenarios, poses a considerable challenge for vision-language models (VLMs) like CLIP, which often struggle to differentiate between semantically similar classes due to insufficient supervision for fine-grained tasks. On the other hand, Large Vision Language Models (LVLMs) have demonstrated remarkable capabilities in tasks like Visual Question Answering (VQA) but remain underexplored in the context of fine-grained image classification. This paper presents CascadeVLM, a novel framework that harnesses the complementary strengths of both CLIP-like and LVLMs VLMs to tackle these challenges. Using granular knowledge effectively in LVLMs and integrating a cascading approach, CascadeVLM dynamically allocates samples using an entropy threshold, balancing computational efficiency with classification accuracy. Experiments on multiple fine-grained datasets, particularly the Stanford Cars dataset, show that CascadeVLM outperforms existing models, achieving 92% accuracy. Our results highlight the potential of combining VLM and LVLM for robust, efficient and interpretable fine-grained image classification, offering new insights into their synergy.

Emotion-Cause Pair Extraction in Conversations (ECPEC) aims to identify emotion utterances and their corresponding cause utterances in unannotated conversations, this task that has garnered increasing attention recently. Previous methods often apply Emotion-Cause Pair Extraction (ECPE) task models, treating the entire conversation as a whole for contextual interaction. However, statistical analysis shows that the number of emotion-cause pairs in ECPEC conversation data far exceeds that in ECPE datasets, leading to interference among multiple events within a conversation and causing noise to propagate between different events. To address this issue, we propose a novel CEnter eveNT-guided framEwoRk (CENTER). This model introduces a Center Event Detection task to construct a center event-aware graph that captures the unique representations of different event regions. Additionally, mimicking human reasoning processes, we build a center event reasoning graph and use graph neural network to facilitate the flow of information between utterance pairs, thereby uncovering the relationships between emotions and their causes. Experimental results demonstrate that our approach achieves state-of-the-art performance across three benchmark datasets.

Due to the high complexity of Discourse Dependency Parsing (DDP) tasks, their existing annotation resources are relatively scarce compared to other NLP tasks, and different DDP tasks also have significant differences in annotation schema. These issues have led to the dilemma of low resources for DDP tasks. Thanks to the powerful capabilities of Large Language Models (LLMs) in cross-task learning, we can use LLMs to model dependency parsing under different annotation schema in an unified manner, in order to alleviate the dilemma of low resources for DDP tasks. However, enabling LLMs to deeply comprehend dependency parsing tasks is a challenge that remains underexplored. Inspired by the application of code-based methods in complex tasks, we propose a code-based unified dependency parsing method. We treat the process of dependency parsing as a search process of dependency paths and use code to represent this search process. Furthermore, we use a curriculum-learning based instruction tuning strategy for joint training of multiple dependency parsing tasks. The experimental results show that our proposed code-based DDP system has achieved good performance on two Chinese DDP tasks (especially significant improvement on the DDP task with relatively less training data).

Research in natural language processing (NLP) for Computational Social Science (CSS) heavily relies on data from social media platforms. This data plays a crucial role in the development of models for analysing socio-linguistic phenomena within online communities. In this work, we conduct an in-depth examination of 20 datasets extensively used in NLP for CSS to comprehensively examine data quality. Our analysis reveals that social media datasets exhibit varying levels of data duplication. Consequently, this gives rise to challenges like label inconsistencies and data leakage, compromising the reliability of models. Our findings also suggest that data duplication has an impact on the current claims of state-of-the-art performance, potentially leading to an overestimation of model effectiveness in real-world scenarios. Finally, we propose new protocols and best practices for improving dataset development from social media data and its usage.

Byzantine-resistant aggregations detect poisonous clients and discard them to ensure that the global model is not poisoned or attacked by malicious clients. However, these aggregations are mainly conducted on the parameter space, and the parameter distances cannot reflect the data distribution divergences between clients. Therefore, existing Byzantine-resistant aggregations cannot defend against backdoor injection by malicious attackers in federated natural language tasks. In this paper, we propose the client embedding for malicious client detection to enhance Byzantine-resistant aggregations. The distances between client embeddings are required to reflect the data distribution divergences of the corresponding clients. Experimental results validate the effectiveness of the proposed client embeddings.

Mathematical question answering over long-form documents is challenging across domains like finance or Wikipedia due to the abundance of candidate arguments within evidence, which complicates recognizing proper arguments for mathematical reasoning and poses hard to learning. In this paper, we propose an approach for training a generator to improve argument recognition. Our method enhances the probabilities of proper arguments in a reasoning program generation so that the arguments comprising the ground truth have higher weights. The proposed approach consists of an argument aggregator to model the probabilities in each candidate generation and an argument set loss to compute the cross-entropy between that probability and the candidates’ existence in the ground truth in terms of the argument set. In our experiments, we show performance improvements of 3.62% and 3.98% in execution accuracy and program accuracy, respectively, over the existing FinQANet model based on a financial mathematical QA dataset. Also, we observed that the similarity of argument sets between the generated program and the ground truth improved by about 2.9%, indicating a mitigation of the misrecognition problem.

Direct Preference Optimization (DPO) is an effective technique that leverages pairwise preference data (one chosen and rejected response per prompt) to align LLMs to human preferences. In practice, multiple responses could exist for a given prompt with varying quality relative to each other. We propose to utilize these responses to create multiple preference pairs for a given prompt. Our work focuses on aligning LLMs by systematically curating multiple preference pairs and presenting them in a meaningful manner facilitating curriculum learning to enhance the prominent DPO technique. We order multiple preference pairs from easy to hard, according to various criteria thus emulating curriculum learning. Our method, which is referred to as Curri-DPO consistently shows increased performance gains on MTbench, Vicuna bench, WizardLM, highlighting its effectiveness over standard DPO setting that utilizes single preference pair. More specifically, Curri-DPO achieves a score of 7.43 on MTbench with Zephyr-7B, outperforming majority of existing LLMs with similar parameter size. Curri-DPO also achieves the highest win rates on Vicuna, WizardLM, and UltraFeedback test sets (90.7%, 87.1%, and 87.9% respectively) in our experiments, with notable gains of up to 7.5% when compared to standard DPO. We release the preference pairs used in alignment at: https://huggingface.co/datasets/ServiceNow-AI/Curriculum_DPO_preferences.

Large language models (LLMs), trained on vast amounts of internet data, have developed a broad understanding of the world, enhancing the decision-making capabilities of embodied agents. This success is largely due to the comprehensive and in-depth domain knowledge within their training datasets. However, the extent of this knowledge can vary across different domains, and existing methods often assume that LLMs have a complete understanding of their environment, overlooking potential gaps in their grasp of actual world dynamics. To address this gap, we introduce Discover, Verify, and Evolve (DiVE), a framework that discovers world dynamics from a small number of demonstrations, verifies the correctness of these dynamics, and evolves new, advanced dynamics tailored to the current situation. Through extensive evaluations, we analyze the impact of each component on performance and compare the automatically generated dynamics from with human-annotated world dynamics. Our results demonstrate that LLMs guided by can make better decisions, achieving rewards comparable to human players in the Crafter environment.

Recent advancements in Vision-Language (VL) research have sparked new benchmarks for complex visual reasoning, challenging models’ advanced reasoning ability. Traditional Vision-Language models (VLMs) perform well in visual perception tasks while struggling with complex reasoning scenarios. Conversely, Large Language Models (LLMs) demonstrate robust text reasoning capabilities; however, they lack visual acuity. To bridge this gap, we propose **C**omplex **V**isual **R**easoning **L**arge **L**anguage **M**odels (**CVR-LLM**), capitalizing on VLMs’ visual perception proficiency and LLMs’ extensive reasoning capability. Unlike recent multimodal large language models (MLLMs) that require a projection layer, our approach transforms images into detailed, context-aware descriptions using an iterative self-refinement loop and leverages LLMs’ text knowledge for accurate predictions without extra training. We also introduce a novel multi-modal in-context learning (ICL) methodology to enhance LLMs’ contextual understanding and reasoning. Additionally, we introduce Chain-of-Comparison (CoC), a step-by-step comparison technique enabling contrasting various aspects of predictions. Our CVR-LLM presents the first comprehensive study across a wide array of complex visual reasoning tasks and achieves SOTA performance among all.

AI-mediated communication enables users to communicate more quickly and efficiently. Various systems have been proposed such as smart reply and AI-assisted writing. Yet, the heterogeneity of the forms of inputs and architectures often renders it challenging to combine insights from user behaviour in one system to improve performance in another. In this work, we consider the case where the user does not select any of the suggested replies from a smart reply system, and how this can be used as one-shot implicit negative feedback to enhance the accuracy of an AI writing model. We introduce Nifty, an approach that uses classifier guidance to controllably integrate implicit user feedback into the text generation process. Empirically, we find up to 34% improvement in Rouge-L, 89% improvement in generating the correct intent, and an 86% win-rate according to human evaluators compared to a vanilla AI writing system on the MultiWOZ and Schema-Guided Dialog datasets. The code is available at https://github.com/BenjaminTowle/NIFTY.

Hallucinated translations pose significant threats and safety concerns when it comes to practical deployment of machine translation systems. Previous research works have identified that detectors exhibit complementary performance — different detectors excel at detecting different types of hallucinations. In this paper, we propose to address the limitations of individual detectors by combining them and introducing a straightforward method for aggregating multiple detectors. Our results demonstrate the efficacy of our aggregated detector, providing a promising step towards evermore reliable machine translation systems.

Modern large language models (LLMs) like ChatGPT have shown remarkable performance on general language tasks but still struggle on complex reasoning tasks, which drives the research on cognitive behaviors of LLMs to explore human-like problem-solving strategies. Along this direction, one representative strategy is self-reflection, which asks an LLM to refine the solution with the feedback generated by itself iteratively. However, our study shows that such reflection-style methods suffer from the Degeneration-of-Thought (DoT) problem: once the LLM has established confidence in its solutions, it is unable to generate novel thoughts later through reflection even if its initial stance is incorrect. To address the DoT problem, we propose a Multi-Agent Debate (MAD) framework, in which multiple agents express their arguments in the state of “tit for tat” and a judge manages the debate process to obtain a final solution. Clearly, our MAD framework encourages divergent thinking in LLMs which would be helpful for tasks that require deep levels of contemplation. Experiment results on two challenging datasets, commonsense machine translation and counter-intuitive arithmetic reasoning, demonstrate the effectiveness of our MAD framework. Extensive analyses suggest that the adaptive break of debate and the modest level of “tit for tat” state are required for MAD to obtain good performance. Moreover, we find that LLMs might not be a fair judge if different LLMs are used for agents.

To better interpret the intrinsic mechanism of large language models (LLMs), recent studies focus on monosemanticity on its basic units. A monosemantic neuron is dedicated to a single and specific concept, which forms a one-to-one correlation between neurons and concepts. Despite extensive research in monosemanticity probing, it remains unclear whether monosemanticity is beneficial or harmful to model capacity. To explore this question, we revisit monosemanticity from the feature decorrelation perspective and advocate for its encouragement. We experimentally observe that the current conclusion by (CITATION), which suggests that decreasing monosemanticity enhances model performance, does not hold when the model changes. Instead, we demonstrate that monosemanticity consistently exhibits a positive correlation with model capacity, in the preference alignment process. Consequently, we apply feature correlation as a proxy for monosemanticity and incorporate a feature decorrelation regularizer into the dynamic preference optimization process. The experiments show that our method not only enhances representation diversity and activation sparsity but also improves preference alignment performance.

Seeking answers effectively for long videos is essential to build video question answering (videoQA) systems. Previous methods adaptively select frames and regions from long videos to save computations. However, this fails to reason over the whole sequence of video, leading to sub-optimal performance. To address this problem, we introduce a state space layer (SSL) into multi-modal Transformer to efficiently integrate global semantics of the video, which mitigates the video information loss caused by frame and region selection modules. Our SSL includes a gating unit to enable controllability over the flow of global semantics into visual representations. To further enhance the controllability, we introduce a cross-modal compositional congruence objective to encourage global semantics aligned with the question. To rigorously evaluate long-form videoQA capacity, we construct two new benchmarks Ego-QA and MAD-QA featuring videos of considerably long length, i.e. 17.5 minutes and 1.9 hours, respectively. Extensive experiments demonstrate the superiority of our framework on these new as well as existing datasets.