EMNLP 2024 Papers

Prior work has explored the possibility of using the semantic information obtained from embedding representations to quantify social stereotypes, leveraging techniques such as word embeddings combined with a list of traits (Garg et al., 2018; Charlesworth et al., 2022) or semantic axes (An et al., 2018; Lucy et al., 2022). However, these approaches have struggled to fully capture the variability in stereotypes across different conceptual domains for the same social group (e.g., black in science, health, and art), in part because the identity of a word and the associations formed during pre-training can dominate its contextual representation (Field and Tsvetkov, 2019). This study explores the ability to recover stereotypes from the contexts surrounding targeted entities by utilizing state-of-the-art text embedding models and adaptive semantic axes enhanced by large language models (LLMs). Our results indicate that the proposed pipeline not only surpasses token-based methods in capturing in-domain framing but also effectively tracks stereotypes over time and along domain-specific semantic axes for in-domain texts. Our research highlights the potential of employing text embedding models to achieve a deeper understanding of nuanced social stereotypes.

Large language models (LLMs) have shown excellent capability for solving reasoning problems. Existing approaches do not differentiate the question difficulty when designing prompting methods for them. Clearly, a simple method cannot elicit sufficient knowledge from LLMs to answer a hard question. Meanwhile, a sophisticated one will force the LLM to generate redundant or even inaccurate intermediate steps toward a simple question. Consequently, the performance of existing methods fluctuates among various questions.In this work, we propose Adaption-of-Thought (AdoT), an adaptive method to improve LLMs for the reasoning problem, which first measures the question difficulty and then tailors demonstration set construction and difficulty-adapted retrieval strategies for the adaptive demonstration construction. Experimental results on three reasoning tasks prove the superiority of our proposed method, showing an absolute improvement of up to 5.5% on arithmetic reasoning, 7.4% on symbolic reasoning, and 2.3% on commonsense reasoning. Our codes and implementation details are available at: https://github.com/NLPGM/AdoT

API integration is crucial for enterprise systems, as it enables seamless interaction between applications within workflows. However, the diversity and complexity of the API landscape present significant challenges in combining API calls based on user intent.Existing methods rely on named entity recognition (NER) and knowledge graphs, but struggle to generate more complex control flow structures, such as conditionals and loops.We propose a novel framework that leverages the success of large language models (LLMs) in code generation to integrate APIs based on natural language input. Our approach involves fine-tuning an LLM using automatically generated API flows derived from OpenAPI specifications.We further evaluate the effectiveness of enforcing the syntax and schema adherence through constrained decoding.To enable systematic comparison, we introduce targeted test suites to assess the generalization capabilities of these approaches and their ability to retain structured knowledge.Our findings show that LLMs fine-tuned on OpenAPI specifications can (a) learn structural API constraints implicitly during training, and (b) achieve significant improvements in both in-distribution and out-of-distribution performance over NER and retrieval-augmented generation (RAG)-based approaches.

Existing works show that augmenting the training data of pre-trained language models (PLMs) for classification tasks fine-tuned via parameter-efficient fine-tuning methods (PEFT) using both clean and adversarial examples can enhance their robustness under adversarial attacks. However, this adversarial training paradigm often leads to performance degradation on clean inputs and requires frequent re-training on the entire data to account for new, unknown attacks. To overcome these challenges while still harnessing the benefits of adversarial training and the efficiency of PEFT, this work proposes a novel approach, called AdpMixup, that combines two paradigms: (1) fine-tuning through adapters and (2) adversarial augmentation via mixup to dynamically leverage existing knowledge from a set of pre-known attacks for robust inference. Intuitively, AdpMixup fine-tunes PLMs with multiple adapters with both clean and pre-known adversarial examples and intelligently mixes them up in different ratios during prediction. Our experiments show AdpMixup achieves the best trade-off between training efficiency and robustness under both pre-known and unknown attacks, compared to existing baselines on five downstream tasks across six varied black-box attacks and 2 PLMs. The code is available at https://github.com/nguyentuc/adapters_mixup.

In the last decade, the generalization and adaptation abilities of deep learning models were typically evaluated on fixed training and test distributions. Contrary to traditional deep learning, large language models (LLMs) are (i) even more overparameterized, (ii) trained on unlabeled text corpora curated from the Internet with minimal human intervention, and (iii) trained in an online fashion. These stark contrasts prevent researchers from transferring lessons learned on model generalization and adaptation in deep learning contexts to LLMs.To this end, our short paper introduces empirical observations that aim to shed light on further training of already pretrained language models. Specifically, we demonstrate that training a model on a text domain could degrade its perplexity on the test portion of the same domain. We observe with our subsequent analysis that the performance degradation is positively correlated with the similarity between the additional and the original pretraining dataset of the LLM. Our further token-level perplexity analysis reveals that the perplexity degradation is due to a handful of tokens that are not informative about the domain. We hope these findings will guide us in determining when to adapt a model vs when to rely on its foundational capabilities.

This work provides an explanatory view of how LLMs can apply moral reasoning to both criticize and defend sexist language. We assessed eight large language models, all of which demonstrated the capability to provide explanations grounded in varying moral perspectives for both critiquing and endorsing views that reflect sexist assumptions. With both human and automatic evaluation, we show that all eight models produce comprehensible and contextually relevant text, which is helpful in understanding diverse views on how sexism is perceived. Also, through analysis of moral foundations cited by LLMs in their arguments, we uncover the diverse ideological perspectives in models’ outputs, with some models aligning more with progressive or conservative views on gender roles and sexism.Based on our observations, we caution against the potential misuse of LLMs to justify sexist language. We also highlight that LLMs can serve as tools for understanding the roots of sexist beliefs and designing well-informed interventions. Given this dual capacity, it is crucial to monitor LLMs and design safety mechanisms for their use in applications that involve sensitive societal topics, such as sexism.

Fine-tuning large language models (LLMs) has achieved remarkable performance across various natural language processing tasks, yet it demands more and more memory as model sizes keep growing. To address this issue, the recently proposed Memory-efficient Zeroth-order (MeZO) methods attempt to fine-tune LLMs using only forward passes, thereby avoiding the need for a backpropagation graph. However, significant performance drops and a high risk of divergence have limited their widespread adoption. In this paper, we propose the Adaptive Zeroth-order Tensor-Train Adaption (AdaZeta) framework, specifically designed to improve the performance and convergence of the ZO methods. To enhance dimension-dependent ZO estimation accuracy, we introduce a fast-forward, low-parameter tensorized adapter. To tackle the frequently observed divergence issue in large-scale ZO fine-tuning tasks, we propose an adaptive query number schedule that guarantees convergence. Detailed theoretical analysis and extensive experimental results on Roberta-Large and Llama-2-7B models substantiate the efficacy of our AdaZeta framework in terms of accuracy, memory efficiency, and convergence speed.

Recent advancements in large language models (LLMs) have been remarkable. Users face a choice between using cloud-based LLMs for generation quality and deploying local-based LLMs for lower computational cost. The former option is typically costly and inefficient, while the latter usually fails to deliver satisfactory performance for reasoning steps requiring deliberate thought processes. In this work, we propose a novel LLM utilization paradigm that facilitates the collaborative operation of large cloud-based LLMs and smaller local-deployed LLMs. Our framework comprises two primary modules: the local agent instantiated with a relatively smaller LLM, handling less complex reasoning steps, and the cloud agent equipped with a larger LLM, managing more intricate reasoning steps. This collaborative processing is enabled through an adaptive mechanism where the local agent introspectively identifies errors and proactively seeks assistance from the cloud agent, thereby effectively integrating the strengths of both locally-deployed and cloud-based LLMs, resulting in significant enhancements in task completion performance and efficiency. We evaluate AdaSwitch across 7 benchmarks, ranging from mathematical reasoning and complex question answering, using various types of LLMs to instantiate the local and cloud agents. The empirical results show that AdaSwitch effectively improves the performance of the local agent, and sometimes achieves competitive results compared to the cloud agent while utilizing much less computational overhead.

Mixture of experts (MoE) has become the standard for constructing production-level large language models (LLMs) due to its promise to boost model capacity without causing significant overheads. Nevertheless, existing MoE methods usually enforce a constant top-k routing for all tokens, which is arguably restrictive because various tokens (e.g., "<EOS>” vs. “apple”) may require various numbers of experts for feature abstraction. Lifting such a constraint can help make the most of limited resources and unleash the potential of the model for downstream tasks. In this sense, we introduce **AdaMoE** to realize token-adaptive routing for MoE, where different tokens are permitted to select a various number of experts. AdaMoE makes minimal modifications to the vanilla MoE with top-k routing—it simply introduces a fixed number of *null experts*, which do not consume any FLOPs, to the expert set and increases the value of k. AdaMoE does not force each token to occupy a fixed number of null experts but ensures the average usage of the null experts with a load-balancing loss, leading to an adaptive number of null/true experts used by each token. AdaMoE exhibits a strong resemblance to MoEs with expert choice routing while allowing for trivial auto-regressive modeling. AdaMoE is easy to implement and can be effectively applied to pre-trained (MoE-)LLMs. Extensive studies show that AdaMoE can reduce average expert load (FLOPs) while achieving superior performance. For example, on the ARC-C dataset, applying our method to fine-tuning Mixtral-8x7B can reduce FLOPs by 14.5% while increasing accuracy by 1.69%.Code is available at [this link](https://github.com/CengZihao/AdaMoE).

Instructions augmentation is a crucial step for unleashing the full potential of large language models (LLMs) in downstream tasks. Existing Self-Instruct methods primarily simulate new instructions from a few initial instructions with in-context learning. However, our study identifies a critical flaw in this approach: even with GPT4o, it cannot generate complex instructions of length ≥ 100, which is necessary in complex tasks such as code completion.To address this issue, our key insight is that fine-tuning open source LLMs with only ten examples can produce complex instructions that maintain distributional consistency for complex reasoning tasks. We introduce Ada-Instruct, an adaptive instruction generator developed through fine-tuning. We empirically validated Ada-Instruct’s efficacy across different applications. The results highlight Ada-Instruct’s capacity to generate long, intricate, and distributionally consistent instructions.

Large language models (LLMs) capable of casual conversation have recently become widely available. We hypothesize that users of conversational systems want a more personalized experience, and existing work shows that users are highly receptive to personalized questions (PQs). Question Generation tasks, however, focus on factual questions from textual excerpts. To create a PQ generator, we first identify over 400 real user interests by anonymously aggregating ~39K user models. We then populate prompt templates with these 400 interests and use an LLM to generate PQs customized to user interests. The result is PerQs, a novel corpus of ~19K question/answer pairs. We evaluate PerQs at scale in the unique context of the Alexa Prize. Our results show significant positive effects on perceived conversation quality. We then fine-tune, deploy, and evaluate PerQy, a neural model that generates PQs in real-time. When evaluated against several competitive LLM baselines, PerQy produced the most natural and engaging responses.

For abstractive text summarization, laborious data annotation and time-consuming model training become two high walls, hindering its further progress. Active Learning, selecting a few informative instances for annotation and model training, sheds light on solving these issues. However, only few active learning-based studies focus on abstractive text summarization and suffer from low stability, effectiveness, and efficiency. To solve the problems, we propose a novel LLM-determined curriculum active learning framework. Firstly, we design a prompt to ask large language models to rate the difficulty of instances, which guides the model to train on from easier to harder instances. Secondly, we design a novel active learning strategy, i.e., Certainty Gain Maximization, enabling to select instances whose distribution aligns well with the overall distribution. Experiments show our method can improve stability, effectiveness, and efficiency of abstractive text summarization backbones.

Given the prompt “Rome is in”, can we steer a language model to flip its prediction of an incorrect token “France” to a correct token “Italy” by only multiplying a few relevant activation vectors with scalars? We argue that successfully intervening on a model is a prerequisite for interpreting its internal workings. Concretely, we establish a three-term objective: a successful intervention should flip the correct with the wrong token and vice versa (effectiveness), and leave other tokens unaffected (faithfulness), all while being sparse (minimality). Using gradient-based optimization, this objective lets us learn (and later evaluate) a specific kind of efficient and interpretable intervention: activation scaling only modifies the signed magnitude of activation vectors to strengthen, weaken, or reverse the steering directions already encoded in the model. On synthetic tasks, this intervention performs comparably with steering vectors in terms of effectiveness and faithfulness, but is much more minimal allowing us to pinpoint interpretable model components. We evaluate activation scaling from different angles, compare performance on different datasets, and make activation scalars a learnable function of the activation vectors themselves to generalize to varying-length prompts.

Large language models(LLMs) have been adopted to process textual task description and accomplish procedural planning in embodied AI tasks because of their powerful reasoning ability. However, there is still lack of study on how vision language models(VLMs) behave when multi-modal task inputs are considered. Counterfactual planning that evaluates the model’s reasoning ability over alternative task situations are also under exploited. In order to evaluate the planning ability of both multi-modal and counterfactual aspects, we propose ActPlan-1K. ActPlan-1K is a multi-modal planning benchmark constructed based on ChatGPT and household activity simulator iGibson2. The benchmark consists of 153 activities and 1,187 instances. Each instance describing one activity has a natural language task description and multiple environment images from the simulator. The gold plan of each instance is action sequences over the objects in provided scenes. Both the correctness and commonsense satisfaction are evaluated on typical VLMs. It turns out that current VLMs are still struggling at generating human-level procedural plans for both normal activities and counterfactual activities. We further provide automatic evaluation metrics by finetuning over BLEURT model to facilitate future research on our benchmark.

Instruction tuning is widely used to unlock the abilities of Large Language Models (LLMs) in following human instructions, resulting in substantial performance improvements across various downstream tasks.Furthermore, contrastive decoding methods are employed to enhance instruction-tuned models. To further explore the potential of contrastive decoding, we introduce the Contrastive Tuning and Decoding (CTD) framework, which enhances model performance without requiring additional data or significant computational resources.When performing Contrastive Tuning, we optimize a correction model by targeting discrepancies between the original outputs and labels. During Contrastive Decoding, the correction model adjusts the logits of the SFT model using the same input to ensure better adherence to instructions.With the lightweight CTD framework, we refine the behavior of instruction-tuned models, improving their performance on the challenging SUPNATINST dataset with unfamiliar data distributions across various models and prompt formats.

When annotators disagree, predicting the labels given by individual annotators can capture nuances overlooked by traditional label aggregation. We introduce three approaches to predict individual annotator ratings on the toxicity of text by incorporating individual annotator-specific information: a neural collaborative filtering (NCF) approach, an in-context learning (ICL) approach, and an intermediate embedding-based architecture. We also study the utility of demographic information for rating prediction. NCF showed limited utility; however, integrating annotator history, demographics, and survey information permits both the embedding-based architecture and ICL to substantially improve prediction accuracy, with the embedding-based architecture outperforming the other methods. We also find that, if demographics are predicted from survey information, using these imputed demographics as features performs comparably to using true demographic data. This suggests that demographics may not provide substantial information for modeling ratings beyond what is captured in survey responses. Our findings raise considerations about the relative utility of different types of annotator information and provide new approaches for modeling annotators in subjective NLP tasks.

Commercially available models dominate academic leaderboards. While impressive, this has concentrated research on creating and adapting general-purpose models to improve NLP leaderboard standings for large language models. However, leaderboards collect many individual tasks and general-purpose models often underperform in specialized domains; domain-specific or adapted models yield superior results. This focus on large general-purpose models excludes many academics and draws attention away from areas where they can make important contributions. We advocate for a renewed focus on developing and evaluating domain- and task-specific models, and highlight the unique role of academics in this endeavor.

Abstract reasoning, the ability to reason from the abstract essence of a problem, serves as a key to generalization in human reasoning. However, eliciting language models to perform reasoning with abstraction remains unexplored. This paper seeks to bridge this gap by introducing a novel structured reasoning format called Abstraction-of-Thought (AoT). The uniqueness of AoT lies in its explicit requirement for varying levels of abstraction within the reasoning process. This approach could elicit language models to first contemplate on the abstract level before incorporating concrete details, which is overlooked by the prevailing step-by-step Chain-of-Thought (CoT) method. To align models with the AoT format, we present AoT Collection, a generic finetuning dataset consisting of 348k high-quality samples with AoT reasoning processes, collected via an automated and scalable pipeline. We finetune a wide range of language models with AoT Collection and conduct extensive evaluations on 23 unseen tasks from the challenging benchmark Big-Bench Hard. Experimental results indicate that models aligned to AoT reasoning format substantially outperform those aligned to CoT in many reasoning tasks.

Large Language Models (LLMs) are widely used in both industry and academia for various tasks, yet evaluating the consistency of generated text responses continues to be a challenge. Traditional metrics like ROUGE and BLEU show a weak correlation with human judgment. More sophisticated metrics using Natural Language Inference (NLI) have shown improved correlations but are complex to implement, require domain-specific training due to poor cross-domain generalization, and lack explainability. More recently, prompt-based metrics using LLMs as evaluators have emerged; while they are easier to implement, they still lack explainability and depend on task-specific prompts, which limits their generalizability. This work introduces Automated eXplainable Consistency Evaluation using LLMs (AXCEL), a prompt-based consistency metric which offers explanations for the consistency scores by providing detailed reasoning and pinpointing inconsistent text spans. AXCEL is also a generalizable metric which can be adopted to multiple tasks without changing the prompt. AXCEL outperforms both non-prompt and prompt-based state-of-the-art (SOTA) metrics in detecting inconsistencies across summarization by 8.7%, free text generation by 6.2%, and data-to-text conversion tasks by 29.4%. We also evaluate the influence of underlying LLMs on prompt based metric performance and recalibrate the SOTA prompt-based metrics with the latest LLMs for fair comparison. Further, we show that AXCEL demonstrates strong performance using open source LLMs.