AACL 2025 Papers

The rapid growth of social networks has led to a surge in social bots, which often disseminate low-quality content and may manipulate public opinion, posing threats to online security. Although recent GNN-based bot detection methods perform strongly, they still face two major challenges. First, deep GNNs are prone to over-smoothing: neighbor aggregation blends bot and human node representations, obscuring bot-specific features. Second, social graphs are dominated by human–human and human–bot connections, while direct bot–bot links are scarce, making it difficult for effective bot representations to propagate within GNNs. To address these issues, we propose a Topology-Aware Gated Graph Neural Network () to detect social bots. employs topology-aware data augmentation to synthesize realistic bot nodes that preserve the original graph structure, mitigating class imbalance; it also introduces a hierarchical gating mechanism that restructures node embeddings into a tree format, selectively filtering noise and enhancing discriminative features. Experiments on three standard benchmark datasets show that consistently surpasses leading baselines in highly imbalanced settings, delivering superior accuracy and robustness.

Visual reasoning is crucial for multimodal large language models (MLLMs) to address complex chart queries, yet high-quality rationale data remains scarce. Existing methods leveraged (M)LLMs for data generation, but direct prompting often yields limited precision and diversity. In this paper, we propose Chain of Functions (CoF), a novel programmatic reasoning data generation pipeline that utilizes freely-explored reasoning paths as supervision to ensure data precision and diversity. Specifically, it starts with human-free exploration among the atomic functions (e.g., maximum data and arithmetic operations) to generate diverse function chains, which are then translated into linguistic rationales and questions with only a moderate open-sourced LLM. CoF provides multiple benefits: 1) Precision: function-governed generation reduces hallucinations compared to freeform generation; 2) Diversity: enumerating function chains enables varied question taxonomies; 3) Explainability: function chains serve as built-in rationales, allowing fine-grained evaluation beyond overall accuracy; 4) Practicality: it eliminates reliance on extremely large models. Employing CoF, we construct the ChartCoF dataset, with 1.4k complex reasoning Q&A for fine-grained analysis and 50k Q&A for reasoning enhancement.Experiments show that ChartCoF improves performance for MLLMs on widely used benchmarks, and the fine-grained evaluation on ChartCoF reveals varying performance across question taxonomies and step numbers for each MLLM. Furthermore, the novel paradigm of function-governed rationale generation in CoF could inspire broader applications beyond charts. The code and data have been publicly available at https://github.com/microsoft/Chain-of-Functions.

Session-based recommendation systems have attracted growing interest for their ability to provide personalized recommendations based on users’ in-session behaviors. While ID-based methods have shown strong performance, they often struggle with long-tail items and overlook valuable textual information. To incorporate text information, various approaches have been proposed, generally employing a naive fusion framework. Interestingly, this approach often fails to outperform the best single-modality baseline. Further exploration indicates a potential imbalance issue in the naive fusion method, where the ID tends to dominate the training and the text is undertrained. This issue indicates that the naive fusion method might not be as effective in combining ID and text as once believed. To address this, we propose AlterRec, an alternative training framework that separates the optimization of ID and text to avoid the imbalance issue. AlterRec also designs an effective strategy to enhance the interaction between the two modalities, facilitating mutual interaction and more effective text integration. Extensive experiments demonstrate the effectiveness of AlterRec in session-based recommendation.

Human behaviors are often guided or constrained by social norms, which are defined as shared, commonsense rules. For example, underlying an action report a witnessed crime are social norms that inform our conduct, such as It is expected to be brave to report crimes. Current AI systems that assess valence (i.e., support or oppose) of human actions by leveraging large-scale data training not grounded on explicit norms may be difficult to explain, and thus untrustworthy. Emulating human assessors by considering social norms can help AI models better understand and predict valence. While multiple norms come into play, conflicting norms can create tension and directly influence human behavior. For example, when deciding whether to report a witnessed crime, one may balance bravery against self-protection. In this paper, we introduce ClarityEthic, a novel ethical assessment approach, to enhance valence prediction and explanation by generating conflicting social norms behind human actions, which strengthens the moral reasoning capabilities of language models by using a contrastive learning strategy. Extensive experiments demonstrate that our method outperforms strong baseline approaches, and human evaluations confirm that the generated social norms provide plausible explanations for the assessment of human behaviors.

Model compression has emerged as a mainstream solution to reduce memory usage and computational overhead. This paper proposes GQSA, a novel model compression framework specifically designed for LLMs. Traditional methods typically focus exclusively on either quantization or sparsification, but relying on a single strategy often results in significant performance loss at high compression rates. In contrast, GQSA integrates quantization and sparsification in a tightly coupled manner, leveraging GPU-friendly structured group sparsity and quantization for efficient acceleration. Building upon system-algorithm co-design principles, we propose a two-stage sparse optimization strategy that ensures the performance superiority of the compressed model. On the engine side, we introduce a “task-centric” parallel strategy, which, to the best of our knowledge, is the first application in the domain of sparse computing. Compared to the traditional 2:4 sparse method, the GQSA offers a more flexible and adjustable sparsity rate, as well as a higher weight compression rate, and is efficiently compatible with weight-only quantization methods. Experimental results demonstrate that, under the GQSA W4S50% compression setting, the model’s accuracy surpasses that of both 2:4 pruning and W2 quantization. Furthermore, at the inference level, GQSA outperforms W2 by 1.26 \times and 2:4 pruning by 2.35 \times in terms of speed.

The pervasiveness of large language models (LLMs) in enterprise settings has also brought forth a significant amount of risks associated with their usage. Guardrails technologies aim to mitigate this risk by filtering LLMs’ input/output text through various detectors. However, developing and maintaining robust detectors has many challenges, one of which is the difficulty in acquiring production-quality labeled data on real LLM outputs before deployment. In this work, we propose STAR, a simple yet intuitive solution to generate production-like labeled data for LLMs’ guardrails development. STAR is based on two key ideas: (i) using self-automated back-querying to synthetically generate data, paired with (ii) a sparse human-in-the-loop clustering technique to label the data. The aim of self-automated back-querying is to construct a parallel corpus roughly representative of the original dataset and resembling real LLM output. We then infuse existing datasets with our synthetically generated examples to produce robust training data for our detectors. We test our technique on one of the most difficult and nuanced detectors: the identification of health advice in LLM output, and demonstrate improvement versus other solutions. Our detector is able to outperform GPT-4o by up to 3.48%, despite having 400x less parameters.

Model collapse—a phenomenon where models degrade in performance due to indiscriminate use of synthetic data—is well studied. However, its role in bias amplification—the progressive reinforcement of pre-existing social biases in Large Language Models (LLMs)—remains underexplored. In this paper, we formally define the conditions for bias amplification and demonstrate through statistical simulations that bias can intensify even in the absence of sampling errors, the primary driver of model collapse. Empirically, we investigate political bias amplification in GPT-2 using a custom-built benchmark for sentence continuation tasks. Our findings reveal a progressively increasing right-leaning bias. Furthermore, we evaluate three mitigation strategies—Overfitting, Preservation, and Accumulation—and show that bias amplification persists even when model collapse is mitigated. Finally, a mechanistic interpretation identifies distinct sets of neurons responsible for model collapse and bias amplification, suggesting they arise from different underlying mechanisms.

In recent years, dense retrieval has been the focus of information retrieval (IR) research. While effective, dense retrieval produces uninterpretable dense vectors, and suffers from the drawback of large index size. Learned sparse retrieval (LSR) has emerged as promising alternative, achieving competitive retrieval performance while also being able to leverage the classical inverted index data structure for efficient retrieval. However, limited works have explored scaling LSR beyond BERT scale. In this work, we identify two challenges in training large language models (LLM) for LSR: (1) training instability during the early stage of contrastive training; (2) suboptimal performance due to pre-trained LLM’s unidirectional attention. To address these challenges, we propose two corresponding techniques: (1) a lightweight adaptation training phase to eliminate training instability; (2) two model variants to enable bidirectional information. With these techniques, we are able to train LSR models with 8B scale LLM, and achieve competitive retrieval performance with reduced index size. Furthermore, we are among the first to analyze the performance-efficiency tradeoff of LLM-based LSR model through the lens of model quantization. Our findings provide insights into adapting LLMs for efficient retrieval modeling.

Multimodal sentiment analysis identifies human emotional tendencies by analyzing text, visual, and auditory modalities. In most studies, the textual modality is usually considered to contain the most emotional information and is regarded as the dominant modality. Existing methods mostly map auxiliary modalities into a semantic space close to the dominant modality, which overly relies on the dominant modality. In this work, we propose a Feature Decomposition-Augmentation (FeaDA) framework, which aims to elevate the role of auxiliary modalities in multimodal data fusion. We first design a projector to decompose auxiliary modalities into partial features, which contain features for emotion judgment, and then utilize these decomposed features to guide the fusion process with KL loss, thereby enhancing the status of auxiliary modality fusion. To verify the effectiveness of our method, we conducted experiments on the CMU-MOSI, CMU-MOSEI, and CH-SIMS datasets. The experimental results show that our FeaDA framework outperforms mutilmodal sentiment analysis methods of the same type in main metrics. Our code is available at https://github.com/PowerLittleYin/FeaDA-main.

To alleviate the high cost of manually annotating Question Answering (QA) datasets, Question Generation (QG) requires the model to generate a question related to the given answer and passage. This work primarily focuses on Multi-Span Question Generation (MSQG), where the generated question corresponds to multiple candidate answers. Existing QG methods may not suit MSQG as they typically overlook the correlation between the candidate answers and generate trivial questions, which limits the quality of the synthetic datasets. Based on the observation that relevant entities typically share the same relationship with the same entity, we propose REGULAR, a framework of RElation-GUided MuLti-SpAn Question GeneRation. REGULAR first converts passages into relation graphs and extracts candidate answers from the relation graphs. Then, REGULAR utilizes a QG model to generate a set of candidate questions and a QA model to obtain the best question. We construct over 100,000 questions using Wikipedia corpora, named REGULAR-WIKI, and conduct experiments to compare our synthetic datasets with other synthetic QA datasets. The experiment results show that models trained with REGULAR-WIKI achieve the best performance. We also conduct ablation studies and statistical analysis to verify the quality of our synthetic dataset. Our code and data are available at https://github.com/PluseLin/REGULAR.

Emotions are a fundamental facet of human experience, varying across individuals, cultural contexts, and nationalities. Given the recent success of Large Language Models (LLMs) as role-playing agents, we examine whether LLMs exhibit emotional stereotypes when assigned nationality-specific personas. Specifically, we investigate how different countries are represented in pre-trained LLMs through emotion attributions and whether these attributions align with cultural norms. To provide a deeper interpretive lens, we incorporate four key cultural dimensions, namely Power Distance, Uncertainty Avoidance, Long-Term Orientation, and Individualism, derived from Hofstede’s cross-cultural framework. Our analysis reveals significant nationality-based differences, with emotions such as shame, fear, and joy being disproportionately assigned across regions. Furthermore, we observe notable misalignment between LLM-generated and human emotional responses, particularly for negative emotions, highlighting the presence of reductive and potentially biased stereotypes in LLM outputs.

Recent work in Differential Privacy with Natural Language Processing (DP NLP) has proposed numerous promising techniques in the form of text rewriting mechanisms. In the evaluation of these mechanisms, an often-ignored aspect is that of dataset size, or rather, the effect of dataset size on a mechanism’s efficacy for utility and privacy preservation. In this work, we are the first to introduce this factor in the evaluation of DP text privatization, where we design utility and privacy tests on large-scale datasets with dynamic split sizes. We run these tests on datasets of varying size with up to one million texts, and we focus on quantifying the effect of increasing dataset size on the privacy-utility trade-off. Our findings reveal that dataset size plays an integral part in evaluating DP text rewriting mechanisms; additionally, these findings call for more rigorous evaluation procedures in DP NLP, as well as shed light on the future of DP NLP in practice and at scale.

Visually rich documents (VRDs) challenge retrieval-augmented generation (RAG) with layout-dependent semantics, brittle OCR, and evidence spread across complex figures and structured tables. This survey examines how Multimodal Large Language Models (MLLMs) are being used to make VRD retrieval practical for RAG. We organize the literature into three roles: Modality-Unifying Captioners, Multimodal Embedders, and End-to-End Representers. We compare these roles along retrieval granularity, information fidelity, latency and index size, and compatibility with reranking and grounding. We also outline key trade-offs and offer some practical guidance on when to favor each role.Finally, we identify promising directions for future research, including adaptive retrieval units, model size reduction, and the development of evaluation methods.

A stereotype is a generalised claim about a social group. Such claims change with culture and context and are often phrased in everyday language, which makes them hard to detect: the State of the Art Large Language Models (LLMs) reach only 68% macro-F1 on the yes/no task “does this sentence contain a stereotype?”. We present HEARTS, a Holistic framework for Explainable, sustAinable and Robust Text Stereotype detection that brings together NLP and social-science. The framework is built on the Expanded Multi-Grain Stereotype Dataset (EMGSD), 57201 English sentences that cover gender, profession, nationality, race, religion and LGBTQ+ topics, adding 10% more data for under-represented groups while keeping high annotator agreement (\kappa = 0.82). Fine-tuning the lightweight ALBERT-v2 model on EMGSD raises binary detection scores to 81.5% macro-F1, matching full BERT while producing 200\times less CO_2. For Explainability, we blend SHAP and LIME token level scores and introduce a confidence measure that increases when the model is correct (\rho = 0.18). We then use HEARTS to assess 16 SOTA LLMs on 1050 neutral prompts each for stereotype propagation: stereotype rates fall by 23% between model generations, yet clear differences remain across model families (LLaMA > Gemini > GPT > Claude). HEARTS thus supplies a practical, low-carbon and interpretable toolkit for measuring stereotype bias in language.

We introduce Tooka-SBERT, a family of Persian sentence embedding models designed to enhance semantic understanding for Persian. The models are released in two sizes—Small (123M parameters) and Large (353M parameters)—both built upon the TookaBERT backbone. Tooka-SBERT is pretrained on the Targoman News corpus and fine-tuned using high-quality synthetic Persian sentence pair datasets to improve semantic alignment. We evaluate Tooka-SBERT on PTEB, a Persian adaptation of the MTEB benchmark, where the Large model achieves an average score of 70.54% and the Small model 69.49%, outperforming some strong multilingual baselines. Tooka-SBERT provides a compact and high-performing open-source solution for Persian sentence representation, with efficient inference suitable for both GPU and CPU environments. Our models are publicly available on Hugging Face, and the corresponding benchmark results can be viewed on the PTEB Leaderboard.

We present a novel agent-based approach for the automated claim matching task with instruction-following LLMs. We propose a two-step pipeline that first generates prompts with LLMs, to then perform claim matching as a binary classification task with LLMs. We demonstrate that LLM-generated prompts can outperform SOTA with human-generated prompts, and that smaller LLMs can do as well as larger ones in the generation process, allowing to save computational resources. We also demonstrate the effectiveness of using different LLMs for each step of the pipeline, i.e. using an LLM for prompt generation, and another for claim matching. Our investigation into the prompt generation process in turn reveals insights into the LLMs’ understanding and handling of the claim matching task.

Large language models (LLMs) have transformed machine translation, yet they have a high subword fertility issue for low-resource languages, which leads to slow inference speed and increased costs. While vocabulary expansion via continual pre-training is a common solution, it often degrades translation quality and requires large target-language corpora, which are unavailable for truly low-resource languages. To address this, we investigate tokenization efficiency through an information-theoretic lens, building on the established hypothesis that word length correlates with information content. From this perspective, we characterize tokenization inefficiency as having high fertility for low-information (highly predictable) words. Guided by this principle, we introduce a novel fine-tuning strategy that systematically identifies informationally redundant words—those with high fertility but low information content—for targeted vocabulary expansion and model fine-tuning. Experiments fine-tuning BLOOM and LLaMA-3 in English-Manipuri and other two language pairs show that our proposed method significantly reduces fertility by 50% and accelerates inference by more than 2 times, without compromising and often exceeding the translation quality of standard LLM baselines, providing a theoretically grounded solution for efficient LLM-based MT.

Soft prompts have emerged as a powerful alternative to adapters in parameter-efficient fine-tuning (PEFT), enabling large language models (LLMs) to adapt to downstream tasks without architectural changes or parameter updates. While prior work has focused on stabilizing training via parameter interaction in small neural prompt encoders, their broader potential for transfer across languages remains unexplored. In this paper, we demonstrate that a prompt encoder can play a central role in improving performance on low-performing languages—those that achieve poor accuracy even under full-model fine-tuning. We investigate a lightweight encoder paired with multi-source training on typologically diverse languages. We call this architecture-training combination the Cross-Prompt Encoder (XPE), and show that it advances the capture of abstract, transferable patterns across languages. To complement XPE, we propose a Dual Soft Prompt mechanism that combines an encoder-based prompt with a directly trained standard soft prompt. This hybrid design proves especially effective for target languages that benefit from both broadly shared structure and language-specific alignment. Text classification experiments with a transformer encoder (XLM-R) on the SIB-200 benchmark reveal a consistent trade-off: XPE is most effective for low-performing languages, while hybrid variants offer broader adaptability across multilingual settings.

Can today’s Text-to-SQL benchmarks still stretch modern LLMs? We argue no. Spider1.0 and BIRD, painstakingly hand-built, remain small, costly, and skewed toward middle complex SQL. Meanwhile, LLM-generated corpora are inexpensive but often superficial and fragile suffering from shallow nesting, semantic drift, template fatigue, and insufficient quality check.We address this gap with a Chain-of-Verifications framework that turns a handful of expert-labelled seeds into a large, reliably checked dataset at a fraction of the usual cost. The resulting corpus, AIGT2S, delivers: (1)18k Question–SQL pairs across 113 databases, 41–77% larger than current English sets; (2)55% queries in the Ultra band of our four-level difficulty taxonomy; (3)87.5% inter-annotator agreement; (4)≥80% labour and ≥98% monetary savings versus earlier efforts.Baselines including GPT-4o, Llama3, RESDSQL, and MAC-SQL, achieve at most 56% execution accuracy, indicating substantial room for improvement.