NAACL 2025 论文列表

Classification tasks in NLP are typically addressed by selecting a pre-trained language model (PLM) from a model hub, and fine-tuning it for the task at hand. However, given the very large number of PLMs that are currently available, a practical challenge is to determine which of them will perform best for a specific downstream task. With this paper, we introduce TransformerRanker, a lightweight library that efficiently ranks PLMs for classification tasks without the need for computationally costly fine-tuning. Our library implements current approaches for transferability estimation (LogME, H-Score, kNN), in combination with layer aggregation options, which we empirically showed to yield state-of-the-art rankings of PLMs (Garbas et al., 2024). We designed the interface to be lightweight and easy to use, allowing users to directly connect to the HuggingFace Transformers and Dataset libraries. Users need only select a downstream classification task and a list of PLMs to create a ranking of likely best-suited PLMs for their task. We make TransformerRanker available as a pip-installable open-source library.

Even though large language models (LLMs) have demonstrated remarkable capability in solving various natural language tasks, the capability of an LLM to follow human instructions is still an area of active development. Recent works (Ouyang et al., 2022; Rafailov et al., 2023; Zhang et al., 2023) have shown great improvements in instruction-following capability through additional training for instruction-following tasks. However, the mechanisms responsible for effective instruction-following capabilities remain inadequately understood. Here, we introduce a simplified instruction-following task and use synthetic datasets to analyze a Transformer-based causal language model. Our findings suggest that the model learns task-specific information by clustering data within its hidden space, with this clustering process evolving dynamically during learning. We also demonstrate how this phenomenon assists the model in handling unseen instances, and validate our results in a more realistic setting. We further present applications in pre-training and alignment, inspired by clustering.

In this paper, we introduce Transform Retrieval, a novel approach aimed at improving Textual Entailment Retrieval within the framework of Retrieval-Augmented Generation (RAG). While RAG has shown promise in enhancing Large Language Models by retrieving relevant documents to extract specific knowledge or mitigate hallucination, current retrieval methods often prioritize relevance without ensuring the retrieved documents semantically support answering the queries. Transform Retrieval addresses this gap by transforming query embeddings to better align with semantic entailment without re-encoding the document corpus. We achieve this by using a transform model and employing a contrastive learning strategy to optimize the alignment between transformed query embeddings and document embeddings for better entailment.We evaluated the framework using BERT as frozen pre-trained encoder and compared it with a fully fine-tuned skyline model. Experimental results show that Transform Retrieval with simple MLP consistently approaches the skyline across multiple datasets, demonstrating the method’s effectiveness. The high performance on HotpotQA highlights its strength in many-to-many retrieval scenarios.

As post-training processes utilize increasingly large datasets and base models continue to grow in size, the computational demands and implementation challenges of existing algorithms are escalating significantly. In this paper, we propose modeling the changes at the logits level during post-training using a separate neural network (i.e., the value network). After training this network on a small base model using demonstrations, this network can be seamlessly integrated with another pre-trained models during inference, enabling them to achieve similar capability enhancements. We systematically investigate the best practices for this paradigm in terms of pre-training weights and connection schemes. We demonstrate that the resulting value network has broad transferability across pre-trained models of different parameter sizes within the same family, models undergoing continuous pre-training within the same family, and models with different vocabularies across families. In certain cases, it can achieve performance comparable to full-parameter fine-tuning. Furthermore, we explore training methods to enhance transferability, which effectively improve the transfer performance of the value model across models of various parameter scales and prevent overfitting to the base model used during training.

Generative language models have shown significant potential in single-turn Text-to-SQL. However, their performance does not extend equivalently to multi-turn Text-to-SQL. This is primarily due to generative language models’ inadequacy in handling the complexities of context information and dynamic schema linking in multi-turn interactions. In this paper, we propose a framework named Track-SQL, which enhances generative language models with dual-extractive modules designed to track schema and contextual changes in multi-turn Text-to-SQL. Specifically, Track-SQL incorporates a Semantic-enhanced Schema Extractor and a Schema-aware Context Extractor. Experimental results demonstrate that Track-SQL achieves state-of-the-art performance on the SparC and CoSQL datasets. Furthermore, detailed ablation studies reveal that Track-SQL significantly improves execution accuracy in multi-turn interactions by 7.1% and 9.55% on these datasets, respectively. Our implementation will be open-sourced at https://github.com/DMIRLAB-Group/Track-SQL.

The assessment of argument quality depends on well-established logical, rhetorical, and dialectical properties that are unavoidably subjective: multiple valid assessments may exist, there is no unequivocal ground truth. This aligns with recent paths in machine learning, which embrace the co-existence of different perspectives. However, this potential remains largely unexplored in NLP research on argument quality. One crucial reason seems to be the yet unexplored availability of suitable datasets. We fill this gap by conducting a systematic review of argument quality datasets. We assign them to a multi-layered categorization targeting two aspects: (a) What has been annotated: we collect the quality dimensions covered in datasets and consolidate them in an overarching taxonomy, increasing dataset comparability and interoperability. (b) Who annotated: we survey what information is given about annotators, enabling perspectivist research and grounding our recommendations for future actions. To this end, we discuss datasets suitable for developing perspectivist models (i.e., those containing individual, non-aggregated annotations), and we showcase the importance of a controlled selection of annotators in a pilot study.

Current multimodal machine translation (MMT) systems rely on fully supervised data (i.e sentences with their translations and accompanying images), which is costly to collect and prevents the extension of MMT to language pairs with no such data. We propose a method to bypass the need for fully supervised data to train MMT systems, using multimodal English data only. Our method ( ZeroMMT) consists in adapting a strong text-only machine translation (MT) model by training it jointly on two objectives: visually conditioned masked language modelling and the Kullback-Leibler divergence between the original MT and new MMT outputs. We evaluate on standard MMT benchmarks and on CoMMuTE, a contrastive test set designed to evaluate how well models use images to disambiguate translations. ZeroMMT obtains disambiguation results close to state-of-the-art MMT models trained on fully supervised examples. To prove that ZeroMMT generalizes to languages with no fully supervised training data, we extend CoMMuTE to three new languages: Arabic, Russian and Chinese. We also show that we can control the trade-off between disambiguation capabilities and translation fidelity at inference time using classifier-free guidance and without any additional data. Our code, data and trained models are publicly accessible.

The annotation and exploration of large text corpora, both automatic and manual, presents significant challenges across multiple disciplines, including linguistics, digital humanities, biology, and legal science. These challenges are exacerbated by the heterogeneity of processing methods, which complicates corpus visualization, interaction, and integration. To address these issues, we introduce the Unified Corpus Explorer (UCE), a standardized, dockerized, open-source and dynamic Natural Language Processing (NLP) application designed for flexible and scalable corpus navigation. Herein, UCE utilizes the UIMA format for NLP annotations as a standardized input, constructing interfaces and features around those annotations while dynamically adapting to the corpora and their extracted annotations. We evaluate UCE based on a user study and demonstrate its versatility as a corpus explorer based on generative AI.

Recent advancements in Large Language Models (LLMs) have sparked widespread concerns about their safety. Recent work demonstrates that safety alignment of LLMs can be easily removed by fine-tuning with a few adversarially chosen instruction-following examples, i.e., fine-tuning attacks. We take a further step to understand fine-tuning attacks in multilingual LLMs. We first discover cross-lingual generalization of fine-tuning attacks: using a few adversarially chosen instruction-following examples in one language, multilingual LLMs can also be easily compromised (e.g., multilingual LLMs fail to refuse harmful prompts in other languages). Motivated by this finding, we hypothesize that safety-related information is language-agnostic and propose a new method termed Safety Information Localization (SIL) to identify the safety-related information in the model parameter space. Through SIL, we validate this hypothesis and find that only changing 20% of weight parameters in fine-tuning attacks can break safety alignment across all languages. Furthermore, we provide evidence to the alternative pathways hypothesis for why freezing safety-related parameters does not prevent fine-tuning attacks, and we demonstrate that our attack vector can still jailbreak LLMs adapted to new languages.

Large language models (LLMs) are increasingly recognized for their exceptional generative capabilities and versatility across various tasks. However, the high inference costs associated with these models have not received adequate attention, particularly when compared to the focus on training costs in existing research. In response to this gap, our study conducts a comprehensive benchmarking of LLM inference energy across a wide range of NLP tasks, where we analyze the impact of different models, tasks, prompts, and system-related factors on inference energy. Specifically, our experiments reveal several interesting insights, including strong correlation of inference energy with output token length and response time. Also, we find that quantization and optimal batch sizes, along with targeted prompt phrases, can significantly reduce energy usage. This study is the first to thoroughly benchmark LLM inference across such a diverse range of aspects, providing insights and offering several recommendations for improving energy efficiency in model deployment.

Reasoning and linguistic skills form the cornerstone of human intelligence, facilitating problem-solving and decision-making. Recent advances in Large Language Models (LLMs) have led to impressive linguistic capabilities and emergent reasoning behaviors, fueling widespread adoption across application domains. However, LLMs still struggle with complex reasoning tasks, highlighting their systemic limitations. In this work, we focus on evaluating whether LLMs have the requisite representations to reason using two foundational relationships: “equivalence” and “inheritance”. We introduce novel tasks and benchmarks spanning six languages and observe that current SOTA LLMs often produce conflicting answers to the same questions across languages in 17.3-57.5% of cases and violate inheritance constraints in up to 37.2% cases. To enhance consistency across languages, we propose novel “Compositional Representations” where tokens are represented as composition of equivalent tokens across languages, with resulting conflict reduction (up to -4.7%) indicating benefits of shared LLM representations.

As the prevalence of phishing attacks continues to rise, there is an increasing demand for more robust detection technologies. With recent advances in AI, we discuss how to construct a reliable and practical phishing detection system using language models. For this system, we introduce the first large-scale Korean dataset for phishing detection, encompassing six types of phishing attacks. We consider multiple factors for building a real-time detection system for edge devices, such as model size, Speech-To-Text quality, split length, training technique and multi-task learning. We evaluate the model’s ability twofold: in-domain, and unseen attack detection performance which is referred to as zero-day performance. Additionally, we demonstrate the importance of accurate comparison groups and evaluation datasets, showing that voice phishing detection performs reasonably well while smishing detection remains challenging. Both the dataset and the trained model will be available upon request.

The emergence of Large Language Models has created new opportunities for building agent applications across various domains. To address the lack of targeted open benchmarks for agent frameworks, we designed a benchmark that features domain-specific, small knowledge bases, and includes a diverse set of questions categorized by type, such as simple, multi-hop, aggregation, and reasoning questions. We evaluated OpenAI’s Assistants API versus a RAG assistant built with Langchain and deployed a RAG system based on benchmark insights as a course assistant over a two-year span in a computer science course. Our findings reveal how domain-specific retrieval impacts response accuracy and highlight key challenges in real-world deployment. Notably, in smaller agentic systems with constrained knowledge bases, the primary challenge shifts from retrieval accuracy to data availability in the knowledge bases. We present insights from both benchmark evaluation and real-world usage data to guide the development of more reliable and effective agentic applications.

This work discusses how to build more rational language and multimodal agents and what criteria define rationality in intelligent systems.Rationality is the quality of being guided by reason, characterized by decision-making that aligns with evidence and logical principles. It plays a crucial role in reliable problem-solving by ensuring well-grounded and consistent solutions. Despite their progress, large language models (LLMs) often fall short of rationality due to their bounded knowledge space and inconsistent outputs. In response, recent efforts have shifted toward developing multimodal and multi-agent systems, as well as integrating modules like external tools, programming codes, symbolic reasoners, utility function, and conformal risk controls rather than relying solely on a single LLM for decision-making. This paper surveys state-of-the-art advancements in language and multimodal agents, assesses their role in enhancing rationality, and outlines open challenges and future research directions. We maintain an open repository at https://github.com/bowen-upenn/Agent_Rationality.

In automated essay scoring (AES), recent efforts have shifted toward cross-prompt settings that score essays on unseen prompts for practical applicability. However, prior methods trained with essay-score pairs of specific prompts pose challenges in obtaining prompt-generalized essay representation. In this work, we propose a grammar-aware cross-prompt trait scoring (GAPS), which internally captures prompt-independent syntactic aspects to learn generic essay representation. We acquire grammatical error-corrected information in essays via the grammar error correction technique and design the AES model to seamlessly integrate such information. By internally referring to both the corrected and the original essays, the model can focus on generic features during training. Empirical experiments validate our method’s generalizability, showing remarkable improvements in prompt-independent and grammar-related traits. Furthermore, GAPS achieves notable QWK gains in the most challenging cross-prompt scenario, highlighting its strength in evaluating unseen prompts.

The widespread practice of indiscriminate data scraping to fine-tune language models (LMs) raises significant legal and ethical concerns, particularly regarding compliance with data protection laws such as the General Data Protection Regulation (GDPR). This practice often results in the unauthorized use of personal information, prompting growing debate within the academic and regulatory communities. Recent works have introduced the concept of generating unlearnable datasets (by adding imperceptible noise to the clean data), such that the underlying model achieves lower loss during training but fails to generalize to the unseen test setting. Though somewhat effective, these approaches are predominantly designed for images and are limited by several practical constraints like requiring knowledge of the target model. To this end, we introduce **RegText**, a framework that injects imperceptible spurious correlations into natural language datasets, effectively rendering them unlearnable without affecting semantic content. We demonstrate RegText’s utility through rigorous empirical analysis of small and large LMs. Notably, RegText can restrict newer models like GPT-4o and Llama from learning on our generated data, resulting in a drop in their test accuracy compared to their zero-shot performance and paving the way for generating unlearnable text to protect public data.

Large Language Models (LLMs) have demonstrated remarkable performance across various tasks. However, they are prone to contextual hallucination, generating information that is either unsubstantiated or contradictory to the given context. Although many studies have investigated contextual hallucinations in LLMs, addressing them in long-context inputs remains an open problem. In this work, we take an initial step toward solving this problem by constructing a dataset specifically designed for long-context hallucination detection. Furthermore, we propose a novel architecture that enables pre-trained encoder models, such as BERT, to process long contexts and effectively detect contextual hallucinations through a decomposition and aggregation mechanism. Our experimental results show that the proposed architecture significantly outperforms previous models of similar size as well as LLM-based models across various metrics, while providing substantially faster inference. We publicly release our dataset and code to promote research along the same line.

To achieve lifelong human-agent interaction, dialogue agents need to constantly memorize perceived information and properly retrieve it for response generation (RG). While prior studies focus on getting rid of outdated memories to improve retrieval quality, we argue that such memories provide rich, important contextual cues for RG (e.g., changes in user behaviors) in long-term conversations. We present THEANINE, a framework for LLM-based lifelong dialogue agents. THEANINE discards memory removal and manages large-scale memories by linking them based on their temporal and cause-effect relation. Enabled by this linking structure, THEANINE augments RG with memory timelines - series of memories representing the evolution or causality of relevant past events. Along with THEANINE, we introduce TeaFarm, a counterfactual-driven evaluation scheme, addressing the limitation of G-Eval and human efforts when assessing agent performance in integrating past memories into RG. A supplementary video for THEANINE and data for TeaFarm are at https://huggingface.co/spaces/ResearcherScholar/Theanine.

Retrieval-Augmented Generation (RAG) systems have shown promise in enhancing the performance of Large Language Models (LLMs). However, these systems face challenges in effectively integrating external knowledge with the LLM’s internal knowledge, often leading to issues with misleading or unhelpful information. This work aims to provide a systematic study on knowledge checking in RAG systems. We conduct a comprehensive analysis of LLM representation behaviors and demonstrate the significance of using representations in knowledge checking. Motivated by the findings, we further develop representation-based classifiers for knowledge filtering. We show substantial improvements in RAG performance, even when dealing with noisy knowledge databases. Our study provides new insights into leveraging LLM representations for enhancing the reliability and effectiveness of RAG systems.