EMNLP 2024 Papers

Data-generation based zero-shot learning, although effective in training Small Task-specific Models (STMs) via synthetic datasets generated by Pre-trained Language Models (PLMs), is often limited by the low quality of such synthetic datasets. Previous solutions have primarily focused on single PLM settings, where synthetic datasets are typically restricted to specific sub-spaces and often deviate from real-world distributions, leading to severe distribution bias. To mitigate such bias, we propose FuseGen, a novel data-generation based zero-shot learning framework that introduces a new criteria for subset selection from synthetic datasets via utilizing multiple PLMs and trained STMs. The chosen subset provides in-context feedback to each PLM, enhancing dataset quality through iterative data generation. Trained STMs are then used for sample re-weighting as well, further improving data quality. Extensive experiments across diverse tasks demonstrate that FuseGen substantially outperforms existing methods, highly effective in boosting STM performance in a PLM-agnostic way. The code is available at https://github.com/LindaLydia/FuseGen.

Model fusion research aims to aggregate the knowledge of multiple individual models to enhance performance by combining their weights. In this work, we study the inverse problem: investigating whether model fusion can be used to reduce unwanted knowledge. We investigate the effects of model fusion in three scenarios: the learning of shortcuts, social biases, and memorization of training data in fine-tuned language models. Through experiments covering classification and generation tasks, our analysis highlights that shared knowledge among models is enhanced during model fusion, while unshared knowledge is usually forgotten. Based on this observation, we demonstrate the potential of model fusion as a debiasing tool and showcase its efficacy in addressing privacy concerns associated with language models.

Test suites assess natural language processing models’ performance on specific functionalities: cases of interest involving model robustness, fairness, or particular linguistic capabilities. This paper introduces specification instructions: text descriptions specifying fine-grained task-specific behaviors. For each functionality in a suite, we generate an instruction that describes it. We combine the specification instructions to create specification-augmented prompts, which we feed to language models pre-trained on natural instruction data.We conduct experiments to measure how optimizing for some functionalities may negatively impact functionalities that are not covered by the specification set. Our analyses across four tasks and models of diverse sizes and families show that smaller models struggle to follow specification instructions. However, larger models (> 3B params.) can benefit from specifications and—surprisingly—even generalize certain desirable behaviors across functionalities.

We explore multi-step reasoning in vision-language models (VLMs). The problem is challenging, as reasoning data consisting of multiple steps of visual and language processing are barely available. To overcome the challenge, we first introduce a least-to-most visual reasoning paradigm, which interleaves steps of decomposing a question into sub-questions and invoking external tools for resolving sub-questions. Based on the paradigm, we further propose a novel data synthesis approach that can automatically create questions and multi-step reasoning paths for an image in a bottom-up manner. Our approach divides the complex synthesis task into a few simple sub-tasks, and (almost entirely) relies on open-sourced models to accomplish the sub-tasks. Therefore, the entire synthesis process is reproducible and cost-efficient, and the synthesized data is quality guaranteed. With the approach, we construct 50k visual reasoning examples. Then, we develop a visual reasoner through supervised fine-tuning, which is capable of generally enhancing the reasoning abilities of a wide range of existing VLMs in a plug-and-play fashion. Extensive experiments indicate that the visual reasoner can consistently and significantly improve four VLMs on four VQA benchmarks.

Multi-label text classification (MLTC) is an important task in the field of natural language processing. Most existing models rely on high-quality text representations provided by pre-trained language models (PLMs). They hence face the challenge of input length limitation caused by PLMs, when dealing with long texts. In light of this, we introduce a comprehensive approach to multi-label long text classification. We propose a text segmentation algorithm, which guarantees to produce the optimal segmentation, to address the issue of input length limitation caused by PLMs. We incorporate external knowledge, labels’ co-occurrence relations, and attention mechanisms in representation learning to enhance both text and label representations. Our method’s effectiveness is validated through extensive experiments on various MLTC datasets, unraveling the intricate correlations between texts and labels.

LLMs can now perform a variety of complex writing tasks. They also excel in answering questions pertaining to natural language inference and commonsense reasoning. Composing these questions is itself a skilled writing task, so in this paper we consider LLMs as authors of commonsense assessment items. We prompt LLMs to generate items in the style of a prominent benchmark for commonsense reasoning, the Choice of Plausible Alternatives (COPA). We examine the outcome according to analyses facilitated by the LLMs and human annotation. We find that LLMs that succeed in answering the original COPA benchmark are also more successful in authoring their own items.

Large language models (LLMs) have achieved significant performance gains using advanced prompting techniques over various tasks. However, the increasing length of prompts leads to high computational costs and often obscures crucial information. Prompt compression has been proposed to alleviate these issues, but it faces challenges in (i) capturing the global context and (ii) training the compressor effectively. To tackle these challenges, we introduce a novel prompt compression method, namely Reading To Compressing (R2C), utilizing the Fusion-in-Decoder (FiD) architecture to identify the important information in the prompt. Specifically, the cross-attention scores of the FiD are used to discern essential chunks and sentences from the prompt. R2C effectively captures the global context without compromising semantic consistency while detouring the necessity of pseudo-labels for training the compressor. Empirical results show that R2C retains key contexts, enhancing the LLM performance by 6% in out-of-domain evaluations while reducing the prompt length by 80%.

We present RICHES, a novel approach that interleaves retrieval with sequence generation tasks. RICHES offers an alternative to conventional RAG systems by eliminating the need for separate retriever and generator. It retrieves documents by directly decoding their contents, constrained on the corpus. Unifying retrieval with generation allows us to adapt to diverse new tasks via prompting alone. RICHES can work with any Instruction-tuned model, without additional training. It provides attributed evidence, supports multi-hop retrievals and interleaves thoughts to plan on what to retrieve next, all within a single decoding pass of the LLM. We demonstrate the strong performance of RICHES across ODQA tasks including attributed and multi-hop QA.

Self-anthropomorphism in robots manifests itself through their display of human-like characteristics in dialogue, such as expressing preferences and emotions. Our study systematically analyzes self-anthropomorphic expression within various dialogue datasets, outlining the contrasts between self-anthropomorphic and non-self-anthropomorphic responses in dialogue systems. We show significant differences in these two types of responses and propose transitioning from one type to the other. We also introduce Pix2Persona, a novel dataset aimed at developing ethical and engaging AI systems in various embodiments. This dataset preserves the original dialogues from existing corpora and enhances them with paired responses: self-anthropomorphic and non-self-anthropomorphic for each original bot response. Our work not only uncovers a new category of bot responses that were previously under-explored but also lays the groundwork for future studies about dynamically adjusting self-anthropomorphism levels in AI systems to align with ethical standards and user expectations.

Despite recent advancements in vision-language models, their performance remains suboptimal on images from non-western cultures due to underrepresentation in training datasets. Various benchmarks have been proposed to test models’ cultural inclusivity. Still, they have limited coverage of cultures and do not adequately assess cultural diversity across universal and culture-specific local concepts. To address these limitations, we introduce the GlobalRG benchmark, comprising two challenging tasks: retrieval across universals and cultural visual grounding. The former task entails retrieving culturally diverse images for universal concepts from 50 countries, while the latter aims at grounding culture-specific concepts within images from 15 countries. Our evaluation across a wide range of models reveals that the performance varies significantly across cultures – underscoring the necessity for enhancing multicultural understanding in vision-language models.

The rapid development of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) has exposed vulnerabilities to various adversarial attacks. This paper provides a comprehensive overview of jailbreaking research targeting both LLMs and MLLMs, highlighting recent advancements in evaluation benchmarks, attack techniques and defense strategies. Compared to the more advanced state of unimodal jailbreaking, multimodal domain remains underexplored. We summarize the limitations and potential research directions of multimodal jailbreaking, aiming to inspire future research and further enhance the robustness and security of MLLMs.

Interpretability and analysis (IA) research is a growing subfield within NLP with the goal of developing a deeper understanding of the behavior or inner workings of NLP systems and methods. Despite growing interest in the subfield, a criticism of this work is that it lacks actionable insights and therefore has little impact on NLP. In this paper, we seek to quantify the impact of IA research on the broader field of NLP. We approach this with a mixed-methods analysis of: (1) a citation graph of 185K+ papers built from all papers published at ACL and EMNLP conferences from 2018 to 2023, and their references and citations, and (2) a survey of 138 members of the NLP community. Our quantitative results show that IA work is well-cited outside of IA, and central in the NLP citation graph. Through qualitative analysis of survey responses and manual annotation of 556 papers, we find that NLP researchers build on findings from IA work and perceive it as important for progress in NLP, multiple subfields, and rely on its findings and terminology for their own work. Many novel methods are proposed based on IA findings and highly influenced by them, but highly influential non-IA work cites IA findings without being driven by them. We end by summarizing what is missing in IA work today and provide a call to action, to pave the way for a more impactful future of IA research.

As artificial intelligence reasoning abilities gain prominence, generating reliable benchmarks becomes crucial. The Abstract and Reasoning Corpus (ARC) offers challenging problems yet unsolved by AI. While ARC effectively assesses reasoning, its generation-based evaluation overlooks other assessment aspects. Bloom’s Taxonomy suggests evaluating six cognitive stages: Remember, Understand, Apply, Analyze, Evaluate, and Create. To extend ARC’s focus beyond the Create stage, we developed MC-LARC, a multiple-choice format suitable for assessing stages like Understand and Apply in Large Language Models (LLMs). Our evaluation of ChatGPT4V’s analogical reasoning using MC-LARC confirmed that this format supports LLMs’ reasoning capabilities and facilitates evidence analysis. However, we observed LLMs using shortcuts in MC-LARC tasks. To address this, we propose a self-feedback framework where LLMs identify issues and generate improved options. MC-LARC is available at https://mc-larc.github.io/.

Large language models (LLMs) have enhanced the capacity of vision-language models to caption visual text. This generative approach to image caption enrichment further makes textual captions more descriptive, improving alignment with the visual context. However, while many studies focus on the benefits of generative caption enrichment (GCE), are there any negative side effects? We compare standard-format captions and recent GCE processes from the perspectives of gender bias and hallucination, showing that enriched captions suffer from increased gender bias and hallucination. Furthermore, models trained on these enriched captions amplify gender bias by an average of 30.9% and increase hallucination by 59.5%. This study serves as a caution against the trend of making captions more descriptive.

It is imperative for Large language models (LLMs) to follow instructions with elaborate requirements (i.e. Complex Instructions Following). Yet, it remains under-explored how to enhance the ability of LLMs to follow complex instructions with multiple constraints. To bridge the gap, we initially study what training data is effective in enhancing complex constraints following abilities. We found that training LLMs with instructions containing multiple constraints enhances their understanding of complex instructions, especially those with lower complexity levels. Additionally, we further propose methods addressing how to obtain and utilize the effective training data. Finally, we conduct extensive experiments to prove the effectiveness of our methods in terms of overall performance and training efficiency. We also demonstrate that our methods improve models’ ability to follow instructions generally and generalize effectively across out-of-domain, in domain, and adversarial settings, while maintaining general capabilities.

With the proliferation of large language models, Parameter Efficient Fine-Tuning (PEFT) method, which freeze pre-trained parameters and only fine-tune a few task-specific parameters, are playing an increasingly important role. However, previous work primarily applied uniform operations across all layers of the model, overlooking the fact that different layers in a transformer store different information. In the process of exploration, We find that there is a significant differences in fine-tuning strategies between different layers, and fine-tuning only a subset of layers can even achieve comparable performance. Based on this, we propose the Hybrid LoRA-Prefix Tuning(HLPT) method, which uses enhanced LoRA and Prefix-tuning methods with learnable adaptive mechanism separately for the bottom and top layers, and the Half Hybrid LoRA-Prefix Tuning(H2LPT) method, which goes a step further, reducing the parameter count to nearly half by omitting fine-tuning in the middle layers. Extensive experiments with large language models on various downstream tasks provide strong evidence for the potential of PEFT focusing on different layers’ interactions and the effectiveness of our methods. Furthermore, we validate the robustness of these methods and their advantages in speeding up training convergence, reducing inference time requirements.

The rapid growth of evaluation methodologies and datasets for large language models (LLMs) has created a pressing need for their unified integration. Meanwhile, concerns about data contamination and bias compromise the trustworthiness of evaluation findings, while the efficiency of evaluation processes remains a bottleneck due to the significant computational costs associated with LLM inference.In response to these challenges, we introduce FreeEval, a modular framework not only for conducting trustworthy and efficient automatic evaluations of LLMs but also serving as a platform to develop and validate new evaluation methodologies. FreeEval addresses key challenges through: (1) unified abstractions that simplify the integration of diverse evaluation methods, including dynamic evaluations requiring complex LLM interactions; (2) built-in meta-evaluation techniques such as data contamination detection and human evaluation to enhance result fairness; (3) a high-performance infrastructure with distributed computation and caching strategies for efficient large-scale evaluations; and (4) an interactive Visualizer for result analysis and interpretation to support innovation of evaluation techniques. We open-source all our code at https://github.com/WisdomShell/FreeEval and our demostration video, live demo, installation guides are available at: https://freeeval.zhuohao.me/.

Recently, LLMs have significantly improved code generation, making it increasingly accessible to users. As a result, LLM-powered code generation applications have sprung up, vastly boosting user productivity. This paper mainly explores how to improve the efficiency and experience of users in formatting the document. Specifically, we propose an automatic document formatting method, Text-to-Format, which is driven by various prompting strategies. Text-to-Format takes the user’s formatting instructions and then generates code that can be run in Microsoft Word to format the content in a document. Further, to evaluate automatic document formatting approaches and advance the document formatting task, we built an evaluation specification including a high-quality dataset DocFormEval data, a code runtime environment, and evaluation metrics. Extensive experimental results on data reveal that the prompting strategy’s effect positively correlates with how much knowledge it introduces related to document formatting task. We believe the constructed DocFormEval data and the exploration about Text-to-Format can help developers build more intelligent tools for automatic document formatting, especially in offline scenarios, where the data privacy is the top priority.

As large language models (LLMs) evolve, evaluating their output reliably becomes increasingly difficult due to the high cost of human evaluation. To address this, we introduce FLAMe, a family of Foundational Large Autorater Models. FLAMe is trained on a diverse set of over 100 quality assessment tasks, incorporating 5M+ human judgments curated from publicly released human evaluations. FLAMe outperforms models like GPT-4 and Claude-3 on various held-out tasks, and serves as a powerful starting point for fine-tuning, as shown in our reward model evaluation case study (FLAMe-RM). On Reward-Bench, FLAMe-RM-24B achieves 87.8% accuracy, surpassing GPT-4-0125 (85.9%) and GPT-4o (84.7%). Additionally, we introduce FLAMe-Opt-RM, an efficient tail-patch fine-tuning approach that offers competitive RewardBench performance using 25×fewer training datapoints. Our FLAMe variants outperform popular proprietary LLM-as-a-Judge models on 8 of 12 autorater benchmarks, covering 53 quality assessment tasks, including RewardBench and LLM-AggreFact. Finally, our analysis shows that FLAMe is significantly less biased than other LLM-as-a-Judge models on the CoBBLEr autorater bias benchmark.