COLING 2025 Papers

“Jailbreak” is a major safety concern of Large Language Models (LLMs), which occurs when malicious prompts lead LLMs to produce harmful outputs, raising issues about the reliability and safety of LLMs. Therefore, an effective evaluation of jailbreaks is very crucial to develop its mitigation strategies. However, our research reveals that many jailbreaks identified by current evaluations may actually be hallucinations—erroneous outputs that are mistaken for genuine safety breaches. This finding suggests that some perceived vulnerabilities might not represent actual threats, indicating a need for more precise red teaming benchmarks. To address this problem, we propose the Benchmark for reliABilitY and jailBreak haLlUcination Evaluation (BabyBLUE). BabyBLUE introduces a specialized validation framework including various evaluators to enhance existing jailbreak benchmarks, ensuring outputs are useful malicious instructions. Additionally, BabyBLUE presents a new dataset as an augmentation to the existing red teaming benchmarks, specifically addressing hallucinations in jailbreaks, aiming to evaluate the true potential of jailbroken LLM outputs to cause harm to human society.

In the Multimodal Sentiment Analysis task, most existing approaches focus on extracting modality-consistent information from raw unimodal data and integrating it into multimodal representations for sentiment classification. However, these methods often assume that all modalities contribute equally to model performance, prioritizing the extraction and enhancement of consistent information, while overlooking the adverse effects of noise caused by modality inconsistency. In contrast to these approaches, this paper introduces a novel approach namely text-guided Hierarchical Noise Eliminator (t-HNE). This model consists of a two-stage denoising phase and a feature recovery phase. Firstly, textual information is injected into both visual and acoustic modalities using an attention mechanism, aiming to reduce intra-modality noise in the visual and acoustic representations. Secondly, it further mitigates inter-modality noise by maximizing the mutual information between textual representations and the respective visual and acoustic representations. Finally, to address the potential loss of modality-invariant information during denoising, the fused multimodal representation is refined through contrastive learning with each unimodal representation except the textual. Extensive experiments conducted on the CMU-MOSI and CMU-MOSEI datasets demonstrate the efficacy of our approach.

As large language models (LLMs) continue to advance, aligning them with human preferences has become a critical objective. In this paper, we introduce stepwise DPO (sDPO), an innovative extension of the recently popularized Direct Preference Optimization (DPO) technique for alignment tuning. sDPO systematically partitions the available preference datasets and applies them incrementally, rather than utilizing the entire dataset simultaneously. This stepwise manner enables the integration of progressively more aligned reference models within the DPO training framework. Our empirical results demonstrate that sDPO not only enhances the alignment precision of reference models but also significantly improves the overall performance of the final model, surpassing other prominent LLMs with larger parameter counts.

Electronic Medical Records (EMRs), while integral to modern healthcare, present challenges for clinical reasoning and diagnosis due to their complexity and information redundancy. To address this, we proposed medIKAL (Integrating Knowledge Graphs as Assistants of LLMs), a framework that combines Large Language Models (LLMs) with knowledge graphs (KGs) to enhance diagnostic capabilities. medIKAL assigns weighted importance to entities in medical records based on their type, enabling precise localization of candidate diseases within KGs. It innovatively employs a residual network-like approach, allowing initial diagnosis by the LLM to be merged into KG search results. Through a path-based reranking algorithm and a fill-in-the-blank style prompt template, it further refined the diagnostic process. We validated medIKAL’s effectiveness through extensive experiments on a newly introduced open-sourced Chinese EMR dataset, demonstrating its potential to improve clinical diagnosis in real-world settings.

Language typology databases enhance multi-lingual Natural Language Processing (NLP) by improving model adaptability to diverse linguistic structures. The widely-used lang2vec toolkit integrates several such databases, but its coverage remains limited at 28.9%. Previous work on automatically increasing coverage predicts missing values based on features from other languages or focuses on single features, we propose to use textual data for better-informed feature prediction. To this end, we introduce a multi-lingual Part-of-Speech (POS) tagger, achieving over 70% accuracy across 1,749 languages, and experiment with external statistical features and a variety of machine learning algorithms. We also introduce a more realistic evaluation setup, focusing on likely to be missing typology features, and show that our approach outperforms previous work in both setups.

Large Language models (LLMs) have become a research hotspot. To accelerate the inference of LLMs, storing computed caches in memory has become the standard technique. However, as the inference length increases, growing KV caches might lead to out-of-memory issues. Many existing methods address this issue through KV cache compression, primarily by preserving key tokens throughout all layers to reduce information loss. Most of them allocate a uniform budget size for each layer to retain. However, we observe that the minimum budget sizes needed to retain essential information vary across layers and models based on the perspectives of attention and hidden state output. Building on this observation, this paper proposes a simple yet effective KV cache compression method that leverages layer uncertainty to allocate budget size for each layer. Experimental results show that the proposed method can reduce memory usage of the KV caches to only ~20% when compared to full KV inference while achieving nearly lossless performance.

Large Language Models (LLMs) have demonstrated remarkable performance through supervised fine-tuning or in-context learning using gold labels. However, this paradigm is limited by the availability of gold labels, while in certain scenarios, LLMs may need to perform tasks that are too complex for humans to provide such labels. To tackle this challenge, this study explores whether solely utilizing unlabeled data can elicit strong model capabilities. We propose a new paradigm termed zero-to-strong generalization. We iteratively prompt LLMs to annotate unlabeled data and retain high-quality labels by filtering. Surprisingly, we obverse that this iterative process gradually unlocks LLMs’ potential on downstream tasks. Our experiments on extensive classification and reasoning tasks confirm the effectiveness of our proposed framework. Our analysis indicates that this paradigm is effective for both in-context learning and fine-tuning, and for various model sizes.

The claim matching (CM) task can benefit an automated fact-checking pipeline by putting together claims that can be resolved with the same fact-check. In this work, we are the first to explore zero-shot and few-shot learning approaches to the task. We consider CM as a binary classification task and experiment with a set of instruction-following large language models (GPT-3.5-turbo, Gemini-1.5-flash, Mistral-7B-Instruct, and Llama-3-8B-Instruct), investigating prompt templates. We introduce a new CM dataset, ClaimMatch, which will be released upon acceptance. We put LLMs to the test in the CM task and find out that it can be tackled by leveraging more mature yet similar tasks such as natural language inference or paraphrase detection. We also propose a pipeline for CM, which we evaluate on texts of different lengths.

Zero-shot slot filling is a well-established subtask of Natural Language Understanding (NLU). However, most existing methods primarily focus on single-turn text data, overlooking the unique complexities of conversational dialogue. Conversational data is highly dynamic, often involving abrupt topic shifts, interruptions, and implicit references that make it difficult to directly apply zero-shot slot filling techniques, even with the remarkable capabilities of large language models (LLMs). This paper addresses these challenges by proposing strategies for automatic data annotation with slot induction and black-box knowledge distillation (KD) from a teacher LLM to a smaller model, outperforming vanilla LLMs on internal datasets by 26% absolute increase in F1 score. Additionally, we introduce an efficient system architecture for call center product settings that surpasses off-the-shelf extractive models by 34% relative F1 score, enabling near real-time inference on dialogue streams with higher accuracy, while preserving low latency.

Automatic biomedical annotation is essential for advancing medical research, diagnosis, and treatment. However, it presents significant challenges, especially when entities are not explicitly mentioned in the text, leading to difficulties in extraction of relevant information. These challenges are intensified by unclear terminology, implicit background knowledge, and the lack of labeled training data. Annotating with a specific ontology adds another layer of complexity, as it requires aligning text with a predefined set of concepts and relationships. Manual annotation is time-consuming and expensive, highlighting the need for automated systems to handle large volumes of biomedical data efficiently. In this paper, we propose an entailment-based zero-shot text classification approach to annotate biomedical text passages using the Homeostasis Imbalance Process (HOIP) ontology. Our method reformulates the annotation task as a multi-class, multi-label classification problem and uses natural language inference to classify text into related HOIP processes. Experimental results show promising performance, especially when processes are not explicitly mentioned, highlighting the effectiveness of our approach for ontological annotation of biomedical literature.

Current rumor detectors exhibit limitations in fully exploiting responses to the source tweet as essential public opinions, and in explaining and indicating the reliability of the results obtained. Additionally, the joint utilization of both responses and the multimodal source content for detection presents challenges due to the heterogeneous nature of the data points. In this work, to address the first challenge, we initially prompt the Large Language Model (LLM) with both multimodal source content and the corresponding response set to extract contrasting evidence to enable maximal utilization of informative responses. To overcome the second challenge, we introduce an uncertainty-aware evidential evaluator to assess the evidence intensity from the multimodal source content and dual-sided reasoning, from which the final prediction is derived. As we model the second-order probability, we can effectively indicate the model’s uncertainty (i.e., the reliability) of the results. The reasoning from the correct perspective also serves as a natural language-based explanation. To this end, the third challenge is also addressed as we fully leverage the available resources. Extensive experiments validate the effectiveness, uncertainty awareness in predictions, helpful explainability for human judgment, and superior efficiency of our approach compared to contemporary works utilizing LLMs.

XTR (Lee et al., 2023) introduced an efficient multi-vector retrieval method that addresses the limitations of the ColBERT (Khattab and Zaharia, 2020model by simplifying retrieval into a single stage through a modified learning objective. While XTR eliminates the need for multistage retrieval, it doesn’t incorporate the efficiency optimizations from ColBERTv2 (Santhanam et al., 2022, which improve indexing and retrieval speed. In this work, we enhance XTR by integrating ColBERTv2’s optimizations, showing that the combined approach preserves the strengths of both models. This results in a more efficient and scalable solution for multi-vector retrieval, while maintaining XTR’s streamlined retrieval process.

In the domain of Document AI, parsing semi-structured image form is a crucial Key Information Extraction (KIE) task. The advent of pre-trained multimodal models significantly empowers Document AI frameworks to extract key information from form documents in different formats such as PDF, Word, and images. Nonetheless, form parsing is still encumbered by notable challenges like subpar capabilities in multilingual parsing and diminished recall in industrial contexts in rich text and rich visuals. In this work, we introduce a simple but effective Multimodal and Multilingual semi-structured FORM PARSER (XFormParser), which is anchored on a comprehensive Transformer-based pre-trained language model and innovatively amalgamates semantic entity recognition (SER) and relation extraction (RE) into a unified framework. Combined with Bi-LSTM, the performance of multilingual parsing is significantly improved. Furthermore, we develop InDFormSFT, a pioneering supervised fine-tuning (SFT) industrial dataset that specifically addresses the parsing needs of forms in a variety of industrial contexts. Through rigorous testing on established benchmarks, XFormParser has demonstrated its unparalleled effectiveness and robustness. Compared to existing state-of-the-art (SOTA) models, XFormParser notably achieves up to 1.79% F1 score improvement on RE tasks in language-specific settings. It also exhibits exceptional improvements in cross-task performance in both multilingual and zero-shot settings.

Large Language Models (LLMs) have demonstrated exceptional performance across a broad spectrum of cross-lingual Natural Language Processing (NLP) tasks. However, previous methods predominantly focus on leveraging parallel corpus to conduct instruction data for continuing pre-training or fine-tuning. They ignored the state of parallel data on the hidden layers of LLMs. In this paper, we demonstrate Word-level Cross-lingual Structure (WCS) of LLM which proves that the word-level embedding on the hidden layers are isomorphic between languages. We find that the hidden states of different languages’ input on the LLMs hidden layers can be aligned with an orthogonal matrix on word-level. We prove this conclusion in both mathematical and downstream task ways on two representative LLM foundations, LLaMA2 and BLOOM. Besides, we propose an Isomorphism-based Data Augmentation (IDA) method to apply the WCS on a downstream cross-lingual task, Bilingual Lexicon Induction (BLI), in both supervised and unsupervised ways. The experiment shows the significant improvement of our proposed method over all the baselines, especially on low-resource languages.

Evaluating machine translation (MT) systems for low-resource languages has long been a challenge due to the limited availability of evaluation metrics and resources. As a result, researchers in this space have relied primarily on lexical-based metrics like BLEU, TER, and ChrF, which lack semantic evaluation. In this first-of-its-kind work, we propose a novel pivot-based evaluation framework that addresses these limitations; after translating low-resource language outputs into a related high-resource language, we leverage advanced neural and embedding-based metrics for more meaningful evaluation. Through a series of experiments using five low-resource languages: Assamese, Manipuri, Kannada, Bhojpuri, and Nepali, we demonstrate how this method extends the coverage of both lexical-based and embedding-based metrics, even for languages not directly supported by advanced metrics. Our results show that the differences between direct and pivot-based evaluation scores are minimal, proving that this approach is a viable and effective solution for evaluating translations in endangered and low-resource languages. This work paves the way for more inclusive, accurate, and scalable MT evaluation for underrepresented languages, marking a significant step forward in this under-explored area of research. The code and data will be made available at https://github.com/AnanyaCoder/PivotBasedEvaluation.

Language models perform differently across languages. It has been previously suggested that morphological typology may explain some of this variability (Cotterell et al., 2018). We replicate previous analyses and find additional new evidence for a performance gap between agglutinative and fusional languages, where fusional languages, such as English, tend to have better language modeling performance than morphologically more complex languages like Turkish. We then propose and test three possible causes for this performance gap: morphological alignment of tokenizers, tokenization quality, and disparities in dataset sizes and measurement. To test the morphological alignment hypothesis, we present MorphScore, a tokenizer evaluation metric, and supporting datasets for 22 languages. We find some evidence that tokenization quality explains the performance gap, but none for the role of morphological alignment. Instead we find that the performance gap is most reduced when training datasets are of equivalent size across language types, but only when scaled according to the so-called “byte-premium”—the different encoding efficiencies of different languages and orthographies. These results suggest that languages of particular morphological types are not intrinsically advantaged or disadvantaged in language modeling. Differences in performance can be attributed to disparities in dataset size. These findings bear on ongoing efforts to improve performance for low-performing and under-resourced languages.

Scientific English is currently undergoing rapid change, with words like “delve,” “intricate,” and “underscore” appearing far more frequently than just a few years ago. It is widely assumed that scientists’ use of large language models (LLMs) is responsible for such trends. We develop a formal, transferable method to characterize these linguistic changes. Application of our method yields 21 focal words whose increased occurrence in scientific abstracts is likely the result of LLM usage. We then pose “the puzzle of lexical overrepresentation”: why are such words overused by LLMs? We fail to find evidence that lexical overrepresentation is caused by model architecture, algorithm choices, or training data. To assess whether reinforcement learning from human feedback (RLHF) contributes to the overuse of focal words, we undertake comparative model testing and conduct an exploratory online study. While the model testing is consistent with RLHF playing a role, our experimental results suggest that participants may be reacting differently to “delve” than to other focal words. With LLMs quickly becoming a driver of global language change, investigating these potential sources of lexical overrepresentation is important. We note that while insights into the workings of LLMs are within reach, a lack of transparency surrounding model development remains an obstacle to such research.

The efficacy of detectors for texts generated by large language models (LLMs) substantially depends on the availability of large-scale training data. However, white-box zero-shot detectors, which require no such data, are limited by the accessibility of the source model of the LLM-generated text. In this paper, we propose a simple yet effective black-box zero-shot detection approach based on the observation that, from the perspective of LLMs, human-written texts typically contain more grammatical errors than LLM-generated texts. This approach involves calculating the Grammar Error Correction Score (GECScore) for the given text to differentiate between human-written and LLM-generated text. Experimental results show that our method outperforms current state-of-the-art (SOTA) zero-shot and supervised methods, achieving an average AUROC of 98.62% across XSum and Writing Prompts dataset. Additionally, our approach demonstrates strong reliability in the wild, exhibiting robust generalization and resistance to paraphrasing attacks. Data and code are available at: https://github.com/NLP2CT/GECScore.

Recently, large language models (LLMs) have shown remarkable performance across various natural language processing tasks, thanks to their vast amount of knowledge. Nevertheless, they often generate unreliable responses. A common example is providing a single biased answer to an ambiguous question that could have multiple correct answers. To address this issue, in this study, we discuss methods to detect such ambiguous samples. More specifically, we propose a classifier that uses a representation from an intermediate layer of the LLM as input. This is based on observations from previous research that representations of ambiguous samples in intermediate layers are closer to those of relevant label samples in the embedding space, but not necessarily in higher layers. The experimental results demonstrate that using representations from intermediate layers detects ambiguous input prompts more effectively than using representations from the final layer. Furthermore, in this study, we propose a method to train such classifiers without ambiguity labels, as most datasets lack labels regarding the ambiguity of samples, and evaluate its effectiveness.