EMNLP 2024 Papers

Finding appropriate prompts for the specific task has become an important issue as the usage of Large Language Models (LLM) have expanded. However, the variety of input-output formats complicate finding the prompts. Reinforcement Learning (RL) is a promising for prompt tuning due to its ability to incrementally produce better results through interaction with the environment. But its inherent training instability and environmental dependency make it difficult to use in practice. In this paper, we propose StablePrompt, a prompt tuning method based on RL. We formulate prompt tuning as RL problem between agent and target LLM, and introduce Adaptive Proximal Policy Optimization (APPO), an modified version of PPO for prompt tuning. APPO introduces an anchor model and updates it adaptively based on the training trajectory. Using this anchor model for the KL divergence term in PPO keeps the search space flexible and ensures training stability. We evaluate StablePrompt on various tasks, including text classification, question answering, and text generation. StablePrompt achieves State-of-The-Art performance across diverse tasks. We demonstrates that StablePrompt performs well across various types and sizes of LLMs. Furthermore, we present TTE-StablePrompt, an extension for generating input-dependent prompts. It outperforms StablePrompt in tasks that are hard to solve with a single prompt. This shows that StablePrompt is an extensible and stable RL framework for LLM.

Large language models have shown their ability to become effective few-shot learners with prompting, revoluting the paradigm of learning with data scarcity. However, this approach largely depends on the quality of prompt initialization and always exhibits large variability among different runs. Such property makes prompt tuning highly unreliable and vulnerable to poorly constructed prompts, which limits its extension to more real-world applications. To tackle this issue, we propose to treat the hard prompt and soft prompt as separate inputs to mitigate noise brought by the prompt initialization. Furthermore, we optimize soft prompts with contrastive learning for utilizing class-aware information in the training process to maintain model performance. Experimental results demonstrate that StablePT outperforms state-of-the-art methods by 6.97% in accuracy and reduces the standard deviation by 1.92 on average. Furthermore, extensive experiments underscore its robustness and stability across 8 datasets covering various tasks.

Stable pre-training is essential for achieving better-performing language models. However, tracking pre-training stability is impractical due to high computational costs. We study Token Embedding Variability as a simple proxy to estimate pre-training stability. We theoretically and empirically demonstrate that Multi-head Low-Rank Attention acts as a fundamental approach to reducing instability. This is supported by empirical findings on variants on GPT-2, demonstrating improved stability and lower perplexities, even at deeper layer counts.

The rapid advancement of generative AI has heightened environmental concerns, particularly regarding carbon emissions. Our framework, Sprout, addresses these challenges by reducing the carbon footprint of inference in large language models (LLMs). Sprout introduces “generation directives” to guide the autoregressive generation process, achieving a balance between ecological sustainability and high-quality outputs. By employing a strategic optimizer for directive assignment and a novel offline quality evaluator, Sprout reduces the carbon footprint of generative LLM inference by over 40% in real-world evaluations, using the Llama model and global electricity grid data. This work is crucial as the rising interest in inference time compute scaling laws amplifies environmental concerns, emphasizing the need for eco-friendly AI solutions.

Large language models (LLMs) have shown promise as automated evaluators for assessing the quality of answers generated by AI systems. However, LLM-based evaluators exhibit position bias, or inconsistency, when used to evaluate candidate answers in pairwise comparisons, favoring either the first or second answer regardless of content. To address this limitation, we propose PORTIA, an alignment-based system designed to mimic human comparison strategies to calibrate position bias in a lightweight yet effective manner. Specifically, PORTIA splits the answers into multiple segments, taking into account both length and semantics, and merges them back into a single prompt for evaluation by LLMs. Extensive experiments with six LLMs on 11,520 answer pairs demonstrate that PORTIA markedly enhances the consistency rates for all models and forms of comparison tested, achieving an average relative improvement of 47.46%. It also enables PORTIA-enhanced GPT-3.5 to achieve agreement rates with humans comparable to GPT-4 and elevates GPT-4’s consistency rate up to 98%. Subsequent human evaluations indicate that the PORTIA-enhanced GPT-3.5 model can even surpass standalone GPT-4 in terms of alignment with human evaluators, highlighting PORTIA’s ability to correct position bias, improve LLM consistency, and boost performance while keeping cost efficiency.

Current instruction-tuned language models are exclusively trained with textual preference data and thus may not be aligned to the unique requirements of other modalities, such as speech. To better align language models with the speech domain, we explore i) prompting strategies based on radio-industry best practices and ii) preference learning using a novel speech-based preference data of 20K samples collected by annotators who listen to response pairs. Both human and automatic evaluation show that both prompting and preference learning increase the speech-suitability of popular instruction tuned LLMs. More interestingly, we show that these methods are additive; combining them achieves the best win rates in head-to-head comparison, resulting in responses that are preferred or tied to the base model in 76.2% of comparisons on average. Lastly, we share lexical, syntactical, and qualitative analyses that elicit how our studied methods differ with baselines in generating more speech-suitable responses.

Recent advances in automatic quality estimation for machine translation have exclusively focused on written language, leaving the speech modality underexplored. In this work, we formulate the task of quality estimation for speech translation (SpeechQE), construct a benchmark, and evaluate a family of systems based on cascaded and end-to-end architectures. In this process, we introduce a novel end-to-end system leveraging pre-trained text LLM. Results suggest that end-to-end approaches are better suited to estimating the quality of direct speech translation than using quality estimation systems designed for text in cascaded systems. More broadly, we argue that quality estimation of speech translation needs to be studied as a separate problem from that of text, and release our [data and models](https://github.com/h-j-han/SpeechQE) to guide further research in this space.

Specialized lexicons are collections of words with associated constraints such as special definitions, specific roles, and intended target audiences. These constraints are necessary for content generation and documentation tasks (e.g., writing technical manuals or children’s reading materials), where the goal is to reduce the ambiguity of text content and increase its overall readability for a specific group of audience. Understanding how large language models can capture these constraints can help researchers build better, more impactful tools for wider use beyond the NLP community. Towards this end, we introduce SpeciaLex, a benchmark for evaluating a language model’s ability to follow specialized lexicon-based constraints across 18 diverse subtasks with 1,785 test instances covering core tasks of Checking, Identification, Rewriting, and Open Generation. We present an empirical evaluation of 15 open and closed-source LLMs and discuss insights on how factors such as model scale, openness, setup, and recency affect performance upon evaluating with the benchmark.

Large Language Models (LLMs) have become essential in advancing natural language processing (NLP) tasks, but their sequential token generation limits inference speed. Multi-Draft Speculative Decoding (MDSD) offers a promising solution by using a smaller draft model to generate multiple token sequences, which the target LLM verifies in parallel.However, current heuristic approaches, such as Recursive Rejection Sampling (RRS), suffer from low acceptance rates in subsequent drafts, limiting the advantages of using multiple drafts. Meanwhile, Optimal Transport with Membership Cost (OTM) can theoretically improve acceptance rates, but its computational cost is too high for real-time use.We present SpecHub, a novel, efficient sampling-verification method for MDSD that improves acceptance rates with only linear computational overhead. By simplifying the OTM problem into a compact Linear Programming model, SpecHub significantly reduces computational complexity. It further accelerates sampling by leveraging a sparse joint distribution, focusing computation on high-probability token sequences.%It integrates seamlessly into existing MDSD frameworks.In extensive experiments, Spechub consistently generates 0.05-0.27 and 0.02-0.16 more tokens per step than RRS and RRS without replacement. We attach our code at https://github.com/MasterGodzilla/Speculative_decoding_OT.

Recently, Denoising Diffusion Probabilistic Models (DDPMs) have attained leading performances across a diverse range of generative tasks. However, in the field of speech synthesis, although DDPMs exhibit impressive performance, their prolonged training duration and substantial inference costs hinder practical deployment. Existing approaches primarily focus on enhancing inference speed, while approaches to accelerate training—a key factor in the costs associated with adding or customizing voices—often necessitate complex modifications to the model, compromising their universal applicability. To address the aforementioned challenges, we propose an inquiry: is it possible to enhance the training/inference speed and performance of DDPMs by modifying the speech signal itself? In this paper, we double the training and inference speed of Speech DDPMs by simply redirecting the generative target to the wavelet domain. This method not only achieves comparable or superior performance to the original model in speech synthesis tasks but also demonstrates its versatility. By investigating and utilizing different wavelet bases, our approach proves effective not just in speech synthesis, but also in speech enhancement.

The performance of Transformer models has been enhanced by increasing the number of parameters and the length of the processed text. Consequently, fine-tuning the entire model becomes a memory-intensive process. High-performance methods for parameter-efficient fine-tuning (PEFT) typically work with Attention blocks and often overlook MLP blocks, which contain about half of the model parameters. We propose a new selective PEFT method, namely SparseGrad, that performs well on MLP blocks. We transfer layer gradients to a space where only about 1% of the layer’s elements remain significant. By converting gradients into a sparse structure, we reduce the number of updated parameters. We apply SparseGrad to fine-tune BERT and RoBERTa for the NLU task and LLaMa-2 for the Question-Answering task. In these experiments, with identical memory requirements, our method outperforms LoRA and MeProp, robust popular state-of-the-art PEFT approaches.

Large language models (LLMs) have shown remarkable achievements across various language tasks. To enhance the performance of LLMs in scientific literature services, we developed the scientific literature LLM (SciLit-LLM) through pre-training and supervised fine-tuning on scientific literature, building upon the iFLYTEK Spark LLM. Furthermore, we present a knowledge service system Spark Research Assistant (SparkRA) based on our SciLit-LLM. SparkRA is accessible online and provides three primary functions: literature investigation, paper reading, and academic writing. As of July 30, 2024, SparkRA has garnered over 50,000 registered users, with a total usage count exceeding 1.3 million.

Large language models (LLMs) can now generate and recognize poetry. But what do LLMs really know about poetry? We develop a task to evaluate how well LLMs recognize one aspect of English-language poetry—poetic form—which captures many different poetic features, including rhyme scheme, meter, and word or line repetition. By using a benchmark dataset of over 4.1k human expert-annotated poems, we show that state-of-the-art LLMs can successfully identify both common and uncommon fixed poetic forms—such as sonnets, sestinas, and pantoums—with surprisingly high accuracy. However, performance varies significantly by poetic form; the models struggle to identify unfixed poetic forms, especially those based on topic or visual features. We additionally measure how many poems from our benchmark dataset are present in popular pretraining datasets or memorized by GPT-4, finding that pretraining presence and memorization may improve performance on this task, but results are inconclusive. We release a benchmark evaluation dataset with 1.4k public domain poems and form annotations, results of memorization experiments and data audits, and code.

Large Language Models have demonstrates remarkable performance in solving math problems, a hallmark of human intelligence.Despite high success rates on current benchmarks, however, these often feature simple problems with only one or two unknowns, which do not sufficiently challenge their reasoning capacities. This paper introduces a novel benchmark, BeyondX, designed to address these limitations by incorporating problems with multiple unknowns. Recognizing the challenges in proposing multi-unknown problems from scratch, we developed BeyondX using an innovative automated pipeline that progressively increases complexity by expanding the number of unknowns in simpler problems. Empirical study on BeyondX reveals that the performance of existing LLMs, even those fine-tuned specifically on math tasks, significantly decreases as the number of unknowns increases - with a performance drop of up to 70% observed in GPT-4. To tackle these challenges, we propose the Formulate-and-Solve strategy, a generalized prompting approach that effectively handles problems with an arbitrary number of unknowns. Our findings reveal that this strategy not only enhances LLM performance on the BeyondX benchmark but also provides deeper insights into the computational limits of LLMs when faced with more complex mathematical challenges.

Although human evaluation remains the gold standard for open-domain dialogue evaluation, the growing popularity of automated evaluation using Large Language Models (LLMs) has also extended to dialogue. However, most frameworks leverage benchmarks that assess older chatbots on aspects such as fluency and relevance, which are not reflective of the challenges associated with contemporary models. In fact, a qualitative analysis on Soda. (Kim et al., 2023), a GPT-3.5 generated dialogue dataset, suggests that current chatbots may exhibit several recurring issues related to coherence and commonsense knowledge, but generally produce highly fluent and relevant responses.Noting the aforementioned limitations, this paper introduces Soda-Eval, an annotated dataset based on Soda that covers over 120K turn-level assessments across 10K dialogues, where the annotations were generated by GPT-4. Using Soda-Eval as a benchmark, we then study the performance of several open-access instruction-tuned LLMs, finding that dialogue evaluation remains challenging. Fine-tuning these models improves performance over few-shot inferences, both in terms of correlation and explanation.

Large Language Models (LLMs) are increasingly used for generating code solutions, empowered by features like self-debugging and self-reflection. However, LLMs often struggle with complex programming problems without human guidance. This paper investigates the strategies employed by expert programmers to steer code-generating LLMs toward successful outcomes. Through a study involving experts using natural language to guide GPT-4, Gemini Ultra, and, Claude 3.5 Sonnet on highly difficult programming challenges, we frame our analysis using the “Socratic Feedback” paradigm for understanding effective steering strategies. By analyzing 30 conversational transcripts across all three models, we map observed feedback strategies to five stages of Socratic Questioning: Definition, Elenhus, Maieutic, Dialectic, and Counter-factual reasoning. We find evidence that by employing a combination of different Socratic feedback strategies across multiple turns, programmers successfully guided the models to solve 74% of the problems that the models initially failed to solve on their own.

While much research has explored enhancing the reasoning capabilities of large language models (LLMs) in the last few years, there is a gap in understanding the alignment of these models with social values and norms. We introduce the task of judging social acceptance. Social acceptance requires models to judge and rationalize the acceptability of people’s actions in social situations. For example, is it socially acceptable for a neighbor to ask others in the community to keep their pets indoors at night? We find that LLMs’ understanding of social acceptance is often misaligned with human consensus. To alleviate this, we introduce SocialGaze, a multi-step prompting framework, in which a language model verbalizes a social situation from multiple perspectives before forming a judgment. Our experiments demonstrate that the SocialGaze approach improves the alignment with human judgments by up to 11 F1 points with the GPT-3.5 model. We also identify biases and correlations in LLMs in assigning blame that is related to features such as the gender (males are significantly more likely to be judged unfairly) and age (LLMs are more aligned with humans for older narrators).

While the impact of social biases in language models has been recognized, prior methods for bias evaluation have been limited to binary association tests on small datasets, limiting our understanding of bias complexities. This paper proposes a novel framework for probing language models for social biases by assessing disparate treatment, which involves treating individuals differently according to their affiliation with a sensitive demographic group. We curate SoFa, a large-scale benchmark designed to address the limitations of existing fairness collections. SoFa expands the analysis beyond the binary comparison of stereotypical versus anti-stereotypical identities to include a diverse range of identities and stereotypes. Comparing our methodology with existing benchmarks, we reveal that biases within language models are more nuanced than acknowledged, indicating a broader scope of encoded biases than previously recognized. Benchmarking LMs on SoFa, we expose how identities expressing different religions lead to the most pronounced disparate treatments across all models. Finally, our findings indicate that real-life adversities faced by various groups such as women and people with disabilities are mirrored in the behavior of these models.

Vision-extended LLMs have made significant strides in Visual Question Answering (VQA). Despite these advancements, VLLMs still encounter substantial difficulties in handling queries involving long-tail entities, with a tendency to produce erroneous or hallucinated responses. In this work, we introduce a novel evaluative benchmark named SnapNTell, specifically tailored for entity-centric VQA. This task aims to test the models’ capabilities in identifying entities and providing detailed, entity-specific knowledge. We have developed the SnapNTell Dataset, distinct from traditional VQA datasets: (1) It encompasses a wide range of categorized entities, each represented by images and explicitly named in the answers; (2) It features QA pairs that require extensive knowledge for accurate responses. The dataset is organized into 22 major categories, containing 7,568 unique entities in total. For each entity, we curated 10 illustrative images and crafted 10 knowledge-intensive QA pairs. To address this novel task, we devised a scalable, efficient, and transparent retrieval-augmented multimodal LLM. Our approach markedly outperforms existing methods on the SnapNTell dataset, achieving a 66.5% improvement in the BELURT score.