EMNLP 2024 Papers

Various types of learning rate (LR) schedulers are being used for training or fine tuning of Large Language Models today. In practice, several mid-flight changes are required in the LR schedule either manually, or with careful choices around warmup steps, peak LR, type of decay and restarts. To study this further, we consider the effect of switching the learning rate at a predetermined time during training, which we refer to as “SkipLR”. We model SGD as a stochastic gradient flow and show that when starting from the same initial parameters, switching the learning rate causes the loss curves to contract towards each other. We demonstrate this theoretically for some simple cases, and empirically on large language models. Our analysis provides insight into how learning rate schedules affect the training dynamics, and could inform the design of new schedules to accelerate convergence.

The emergence of specialized large language models (LLMs) has shown promise in addressing complex tasks in materials science. Many LLMs, however, often struggle with the distinct complexities of materials science tasks, such as computational challenges, and rely heavily on outdated implicit knowledge, leading to inaccuracies and hallucinations. To address these challenges, we introduce HoneyComb, the first LLM-based agent system specifically designed for materials science. HoneyComb leverages a reliable, high-quality materials science knowledge base (MatSciKB) and a sophisticated tool hub (ToolHub) tailored specifically for materials science to enhance its reasoning and computational capabilities. MatSciKB is a curated, structured knowledge collection based on reliable literature, while ToolHub employs an Inductive Tool Construction method to generate, decompose, and refine API tools for materials science. Additionally, HoneyComb leverages a retriever module that adaptively selects the appropriate knowledge source or tools for specific tasks, thereby ensuring accuracy and relevance. Our results demonstrate that HoneyComb significantly outperforms baseline models across various tasks in materials science, effectively bridging the gap between current LLM capabilities and the specialized needs of this domain. Furthermore, our adaptable framework can be easily extended to other scientific domains, highlighting its potential for broad applicability in advancing scientific research and applications.

Interleaved text-and-image generation has been an intriguing research direction, where the models are required to generate both images and text pieces in an arbitrary order. Despite the emerging advancements in interleaved generation, the progress in its evaluation still significantly lags behind. Existing evaluation benchmarks do not support arbitrarily interleaved images and text for both inputs and outputs, and they only cover a limited number of domains and use cases. Also, current works predominantly use similarity-based metrics which fall short in assessing the quality in open-ended scenarios. To this end, we introduce InterleavedBench, the first benchmark carefully curated for the evaluation of interleaved text-and-image generation. InterleavedBench features a rich array of tasks to cover diverse real-world use cases. In addition, we present InterleavedEval, a strong reference-free metric powered by GPT-4o to deliver accurate and explainable evaluation. We carefully define five essential evaluation aspects for InterleavedEval, including text quality, perceptual quality, image coherence, text-image coherence, and helpfulness, to ensure a comprehensive and fine-grained assessment. Through extensive experiments and rigorous human evaluation, we show that our benchmark and metric can effectively evaluate the existing models with a strong correlation with human judgments surpassing previous reference-based metrics. We also provide substantial findings and insights to foster future research in interleaved generation and its evaluation.

Automated red teaming is an effective method for identifying misaligned behaviors in large language models (LLMs). Existing approaches, however, often focus primarily on improving attack success rates while overlooking the need for comprehensive test case coverage. Additionally, most of these methods are limited to single-turn red teaming, failing to capture the multi-turn dynamics of real-world human-machine interactions. To overcome these limitations, we propose **HARM** (**H**olistic **A**utomated **R**ed tea**M**ing), which scales up the diversity of test cases using a top-down approach based on an extensible, fine-grained risk taxonomy. Our method also leverages a novel fine-tuning strategy and reinforcement learning techniques to facilitate multi-turn adversarial probing in a human-like manner. Experimental results demonstrate that our framework enables a more systematic understanding of model vulnerabilities and offers more targeted guidance for the alignment process.

Generative AI has demonstrated unprecedented creativity in the field of computer vision, yet such phenomena have not been observed in natural language processing. In particular, large language models (LLMs) can hardly produce written works at the level of human experts due to the extremely high complexity of literature writing. In this paper, we present HoLLMwood, an automated framework for unleashing the creativity of LLMs and exploring their potential in screenwriting, which is a highly demanding task. Mimicking the human creative process, we assign LLMs to different roles involved in the real-world scenario. In addition to the common practice of treating LLMs as Writer, we also apply LLMs as Editor, who is responsible for providing feedback and revision advice to Writer. Besides, to enrich the characters and deepen the plots, we introduce a role-playing mechanism and adopt LLMs as Actors that can communicate and interact with each other. Evaluations on automatically generated screenplays show that HoLLMwood substantially outperforms strong baselines in terms of coherence, relevance, interestingness and overall quality.

Parameter-Efficient Fine-Tuning (PEFT) on small Pre-trained Language Models (PLMs) has emerged as a promising approach to enhance their multi-tasking capabilities. Prevalent methods simultaneously train additional modules (i.e., one task-shared module and multiple task-specific modules) for adapting PLMs to downstream tasks. However, their adaptability to new tasks is constrained, as the task-specific modules independently adapt to each task, overlooking the potential for knowledge transfer across tasks. In this paper, we propose a novel multi-task learning framework, Inspirational Pointer (IP), that enables the transfer of prior knowledge across tasks through human language intervention. Specifically, we attach task descriptions to the input samples, which are then mapped to corresponding task embeddings. Based on those embeddings, we adapt PLMs for downstream tasks. Similar tasks share akin descriptions, allowing new task samples close to similar trained tasks in the task embedding space, hitting the memory about trained tasks of the model. Our experiments on the T5 model demonstrate performance improvements of our method in multi-task learning and few-shot transfer learning. Further, we implemented the IP in decoder-only models including GPT2 and large language models (LLMs), and the results show that IP enhances the capabilities of decoder-only models.

Despite the advances in large language models (LLMs), how they use their knowledge for reasoning is not yet well understood.In this study, we propose a method that deconstructs complex real-world questions into a graph, representing each question as a node with predecessors of background knowledge needed to solve the question. We develop the DepthQA dataset, deconstructing questions into three depths: (i) recalling conceptual knowledge, (ii) applying procedural knowledge, and (iii) analyzing strategic knowledge. Based on a hierarchical graph, we quantify forward discrepancy, a discrepancy in LLM performance on simpler sub-problems versus complex questions. We also measure backward discrepancy where LLMs answer complex questions but struggle with simpler ones. Our analysis shows that smaller models exhibit more discrepancies than larger models. Distinct patterns of discrepancies are observed across model capacity and possibility of training data memorization. Additionally, guiding models from simpler to complex questions through multi-turn interactions improves performance across model sizes, highlighting the importance of structured intermediate steps in knowledge reasoning. This work enhances our understanding of LLM reasoning and suggests ways to improve their problem-solving abilities.

Do LLMs have political leanings and are LLMs able to shift our political views? This paper explores these questions in the context of the 2024 U.S. presidential election. Through a voting simulation, we demonstrate 18 open-weight and closed-source LLMs’ political preference for Biden over Trump. We show how Biden-leaning becomes more pronounced in instruction-tuned and reinforced models compared to their base versions by analyzing their responses to political questions related to the two nominees. We further explore the potential impact of LLMs on voter choice by recruiting 935 U.S. registered voters. Participants interacted with LLMs (Claude-3, Llama-3, and GPT-4) over five exchanges. Intriguingly, although LLMs were not asked to persuade users to support Biden, about 20% of Trump supporters reduced their support for Trump after LLM interaction. This result is noteworthy given that many studies on the persuasiveness of political campaigns have shown minimal effects in presidential elections. Many users also expressed a desire for further interaction with LLMs on political subjects. Further research on how LLMs affect users’ political views is required, as their use becomes more widespread.

Full-parameter fine-tuning (FPFT) has become the go-to choice for adapting language models (LMs) to downstream tasks due to its excellent performance. As LMs grow in size, fine-tuning the full parameters of LMs requires a prohibitively large amount of GPU memory. Existing approaches utilize zeroth-order optimizer to conserve GPU memory, which potentially compromises the performance of LMs as non-zero order optimizers tend to converge more readily on most downstream tasks. We propose a novel, memory-efficient, optimizer-independent, end-to-end hierarchical fine-tuning strategy, HiFT, which only updates a subset of parameters at each training step. HiFT significantly reduces the amount of gradients and optimizer state parameters residing in GPU memory at the same time, thereby reducing GPU memory usage. Our results demonstrate that: (1) HiFT achieves comparable performance with parameter-efficient fine-tuning and standard FPFT. (2) Results on six models show that HiFT reduces the number of trainable parameters by about 89.18% on average compared to FPFT. (3) HiFT supports FPFT of 7B models for 24G GPU memory devices under mixed precision without using any memory saving techniques. (4) HiFT supports various optimizers including AdamW, AdaGrad, SGD, etc. The source code link is https://github.com/misonsky/HiFT.

Rhetorical Role Labeling (RRL) of legal documents is pivotal for various downstream tasks such as summarization, semantic case search and argument mining. Existing approaches often overlook the varying difficulty levels inherent in legal document discourse styles and rhetorical roles. In this work, we propose HiCuLR, a hierarchical curriculum learning framework for RRL. It nests two curricula: Rhetorical Role-level Curriculum (RC) on the outer layer and Document-level Curriculum (DC) on the inner layer. DC categorizes documents based on their difficulty, utilizing metrics like deviation from a standard discourse structure and exposes the model to them in an easy-to-difficult fashion. RC progressively strengthens the model to discern coarse-to-fine-grained distinctions between rhetorical roles. Our experiments on four RRL datasets demonstrate the efficacy of HiCuLR, highlighting the complementary nature of DC and RC.

Foundation models (FMs) adapt surprisingly well to downstream tasks with fine-tuning. However, their colossal parameter space prohibits their training on resource-constrained edge-devices. For federated fine-tuning, we need to consider the smaller FMs of few billion parameters at most, namely on-device FMs (ODFMs), which can be deployed on-device. Federated fine-tuning of ODFMs has unique challenges non-present in standard fine-tuning: i) ODFMs poorly generalize to downstream tasks due to their limited sizes making proper fine-tuning imperative to their performance, and ii) devices have limited and heterogeneous system capabilities and data that can deter the performance of fine-tuning.Tackling these challenges, we propose HetLoRA, a feasible and effective federated fine-tuning method for ODFMs that leverages the system and data heterogeneity at the edge. HetLoRA allows heterogeneous LoRA ranks across clients for their individual system resources, and efficiently aggregates and distributes these LoRA modules in a data-aware manner by applying rank self-pruning locally and sparsity-weighted aggregation at the server. It combines the advantages of high and low-rank LoRAs, achieving improved convergence speed and final performance compared to homogeneous LoRA. Furthermore, HetLoRA has enhanced computation and communication efficiency compared to full fine-tuning making it more feasible for the edge.

As generative AI progresses, collaboration be-tween doctors and AI scientists is leading to thedevelopment of personalized models to stream-line healthcare tasks and improve productivity.Summarizing doctor-patient dialogues has be-come important, helping doctors understandconversations faster and improving patient care.While previous research has mostly focused ontext data, incorporating visual cues from pa-tient interactions allows doctors to gain deeperinsights into medical conditions. Most of thisresearch has centered on English datasets, butreal-world conversations often mix languagesfor better communication. To address the lackof resources for multimodal summarization ofcode-mixed dialogues in healthcare, we devel-oped the MCDH dataset. Additionally, we cre-ated HealthAlignSumm, a new model that in-tegrates visual components with the BART ar-chitecture. This represents a key advancementin multimodal fusion, applied within both theencoder and decoder of the BART model. Ourwork is the first to use alignment techniques,including state-of-the-art algorithms like DirectPreference Optimization, on encoder-decodermodels with synthetic datasets for multimodalsummarization. Through extensive experi-ments, we demonstrated the superior perfor-mance of HealthAlignSumm across severalmetrics validated by both automated assess-ments and human evaluations. The datasetMCDH and our proposed model HealthAlign-Summ will be available in this GitHub accounthttps://github.com/AkashGhosh/HealthAlignSumm-Utilizing-Alignment-for-Multimodal-Summarization-of-Code-Mixed-Healthcare-DialoguesDisclaimer: This work involves medical im-agery based on the subject matter of the topic.

Hate speech detection is a prevalent research field, yet it remains underexplored at the level of word meaning. This is significant, as terms used to convey hate often involve non-standard or novel usages which might be overlooked by commonly leveraged LMs trained on general language use. In this paper, we introduce the Hateful Word in Context Classification (HateWiC) task and present a dataset of ~4000 WiC-instances, each labeled by three annotators. Our analyses and computational exploration focus on the interplay between the subjective nature (context-dependent connotations) and the descriptive nature (as described in dictionary definitions) of hateful word senses. HateWiC annotations confirm that hatefulness of a word in context does not always derive from the sense definition alone. We explore the prediction of both majority and individual annotator labels, and we experiment with modeling context- and sense-based inputs. Our findings indicate that including definitions proves effective overall, yet not in cases where hateful connotations vary. Conversely, including annotator demographics becomes more important for mitigating performance drop in subjective hate prediction.

The widespread use of social media necessitates reliable and efficient detection of offensive content to mitigate harmful effects. Although sophisticated models perform well on individual datasets, they often fail to generalize due to varying definitions and labeling of “offensive content.” In this paper, we introduce HateCOT, an English dataset with over 52,000 samples from diverse sources, featuring explanations generated by GPT-3.5Turbo and curated by humans. We demonstrate that pretraining on HateCOT significantly enhances the performance of open-source Large Language Models on three benchmark datasets for offensive content detection in both zero-shot and few-shot settings, despite differences in domain and task. Additionally, HateCOT facilitates effective K-shot fine-tuning of LLMs with limited data and improves the quality of their explanations, as confirmed by our human evaluation.

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.

Automating the measurement of hallucinations in LLM generated responses is a challenging task as it requires careful investigation of each factual claim in a response. In this paper, we introduce HalluMeasure, a new LLM-based hallucination detection mechanism that decomposes an LLM response into atomic claims, and evaluates each atomic claim against the provided reference context. The model uses a step-by-step reasoning process called Chain-of-Thought and can identify 3 major categories of hallucinations (e.g., contradiction) as well as 10 more specific subtypes (e.g., overgeneralization) which help to identify reasons behind the hallucination errors. Specifically, we explore four different configurations for HalluMeasure’s classifier: with and without CoT prompting, and using a single classifier call to classify all claims versus separate calls for each claim. The best-performing configuration (with CoT and separate calls for each claim) demonstrates significant improvements in detecting hallucinations, achieving a 10-point increase in F1 score on our TechNewsSumm dataset, and a 3-point increase in AUC ROC on the SummEval dataset, compared to three baseline models (RefChecker, AlignScore, and Vectara HHEM). We further show reasonable accuracy on detecting 10 novel error subtypes of hallucinations (where even humans struggle in classification) derived from linguistic analysis of the errors made by the LLMs.

Heart sound auscultation holds significant importance in the diagnosis of congenital heart disease. However, existing methods for Heart Sound Diagnosis (HSD) tasks are predominantly limited to a few fixed categories, framing the HSD task as a rigid classification problem that does not fully align with medical practice and offers only limited information to physicians. Besides, such methods do not utilize echocardiography reports, the gold standard in the diagnosis of related diseases. To tackle this challenge, we introduce HSDreport, a new benchmark for HSD, which mandates the direct utilization of heart sounds obtained from auscultation to predict echocardiography reports. This benchmark aims to merge the convenience of auscultation with the comprehensive nature of echocardiography reports. First, we collect a new dataset for this benchmark, comprising 2,275 heart sound samples along with their corresponding reports. Subsequently, we develop a knowledge-aware query-based transformer to handle this task. The intent is to leverage the capabilities of medically pre-trained models and the internal knowledge of large language models (LLMs) to address the task’s inherent complexity and variability, thereby enhancing the robustness and scientific validity of the method. Furthermore, our experimental results indicate that our method significantly outperforms traditional HSD approaches and existing multimodal LLMs in detecting key abnormalities in heart sounds.

Large Vision-Language Models (LVLMs) have shown remarkable performance on many visual-language tasks. However, these models still suffer from multimodal hallucination, which means the generation of objects or content that violates the images. Many existing work detects hallucination by directly judging whether an object exists in an image, overlooking the association between the object and semantics. To address this issue, we propose Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding (HELPD). This framework incorporates hallucination feedback at both object and sentence semantic levels. Remarkably, even with a marginal degree of training, this approach can alleviate over 15% of hallucination. Simultaneously, HELPD penalizes the output logits according to the image attention window to avoid being overly affected by generated text. HELPD can be seamlessly integrated with any LVLMs. Our experiments demonstrate that the proposed framework yields favorable results across multiple hallucination benchmarks. It effectively mitigates hallucination for different LVLMs and concurrently improves their text generation quality.

Empathy serves as a cornerstone in enabling prosocial behaviors, and can be evoked through sharing of personal experiences in stories. While empathy is influenced by narrative content, intuitively, people respond to the way a story is told as well, through narrative style. Yet the relationship between empathy and narrative style is not fully understood. In this work, we empirically examine and quantify this relationship between style and empathy using LLMs and large-scale crowdsourcing studies. We introduce a novel, theory-based taxonomy, HEART (Human Empathy and Narrative Taxonomy) that delineates elements of narrative style that can lead to empathy with the narrator of a story. We establish the performance of LLMs in extracting narrative elements from HEART, showing that prompting with our taxonomy leads to reasonable, human-level annotations beyond what prior lexicon-based methods can do. To show empirical use of our taxonomy, we collect a dataset of empathy judgments of stories via a large-scale crowdsourcing study with N=2,624 participants. We show that narrative elements extracted via LLMs, in particular, vividness of emotions and plot volume, can elucidate the pathways by which narrative style cultivates empathy towards personal stories. Our work suggests that such models can be used for narrative analyses that lead to human-centered social and behavioral insights.