AACL 2025 Papers

The rapid proliferation of frontier model agents promises significant societal advances but also raises concerns about systemic risks arising from unsafe interactions.Collusion to the disadvantage of others has been identified as a central form of undesirable agent cooperation.The use of information hiding (steganography) in agent communications could render such collusion practically undetectable.This underscores the need for investigations into the possibility of such behaviours emerging and the robustness corresponding countermeasures.To investigate this problem we design two approaches – a gradient-based reinforcement learning (GBRL) method and an in-context reinforcement learning (ICRL) method – for reliably eliciting sophisticated LLM-generated linguistic text steganography.We demonstrate, for the first time, that unintended steganographic collusion in LLMs can arise due to mispecified reward incentives during training.Additionally, we find that standard mitigations — both passive oversight of model outputs and active mitigation through communication paraphrasing — are not fully effective at preventing this steganographic communication.Our findings imply that (i) emergence of steganographic collusion is a plausible concern that should be monitored and researched, and (ii) preventing emergence may require innovation in mitigation techniques.

This paper investigates defenses in LLM-based evaluation, where prompt injection attacks can manipulate scores by deceiving the evaluation system. We formalize blind attacks as a class in which candidate answers are crafted independently of the true answer. To counter such attacks, we propose an evaluation framework that combines standard and counterfactual evaluation. Experiments show it significantly improves attack detection with minimal performance trade-offs for recent models.

Pruning techniques have been studied to construct small models for efficiency, yet the effect of cross-lingual, which shows language performance transferability, is understudied in this field. In this work, we investigate cross-lingual effects in multilingual large language model compression using iterative pruning and recovery. We employ structured layer pruning with LoRA-based recovery and knowledge distillation, testing whether calibration languages different from target evaluation languages can preserve multilingual performance. Experiments on Qwen2.5-7B and Llama3.1-8B demonstrate that any recovery language consistently outperforms no-recovery baselines, with even low-resource languages like Swahili providing ~5% improvements. In contrast to expectations, dominant pretraining languages do not always yield the best results, where Indonesian achieves the highest performance in Llama3.1-8B, while Japanese performs the best in Qwen2.5-7B. Our findings reveal that cross-lingual calibration effectively maintains multilingual capabilities in the iterative pruning.

Large Language Models (LLMs) demonstrate strong reasoning and self-correction abilities in high-resource languages like English, but their performance remains limited in low-resource languages such as Korean. In this study, we investigate whether reinforcement learning (RL) can enhance Korean reasoning abilities to a degree comparable to English. Our findings reveal that RL alone yields limited improvements when applied to models lacking inherent Korean reasoning capabilities. To address this, we explore several fine-tuning strategies and show that aligning the model’s internal reasoning processes with Korean inputs—particularly by tuning Korean-specific neurons in early layers—is key to unlocking RL’s effectiveness. We introduce a self-correction code-switching dataset to facilitate this alignment and observe significant performance gains in both mathematical reasoning and self-correction tasks. Ultimately, we conclude that the crucial factor in multilingual reasoning enhancement is not injecting new linguistic knowledge, but effectively eliciting and aligning existing reasoning capabilities. Our study provides a new perspective on how internal translation and neuron-level tuning contribute to multilingual reasoning alignment in LLMs.

Recently, Large Language Models (LLMs) have dominated much of the artificial intelligence scene with their ability to process and generate natural languages. However, the majority of LLM research and development remains English-centric, leaving low-resource languages such as those in the Southeast Asian (SEA) region under-represented. To address this representation gap, we introduce Llama-SEA-LION-v3-8B-IT and Gemma-SEA-LION-v3-9B-IT, two cutting-edge multilingual LLMs designed for SEA languages. The SEA-LION family of LLMs supports 11 SEA languages, namely English, Chinese, Indonesian, Vietnamese, Malay, Thai, Burmese, Lao, Filipino, Tamil, and Khmer. Our work leverages large-scale multilingual continued pre-training with a comprehensive post-training regime involving multiple stages of instruction fine-tuning, alignment, and model merging. Evaluation results on multilingual benchmarks indicate that our models achieve state-of-the-art performance across LLMs supporting SEA languages. We open-source the models to benefit the wider SEA community.

As large language models (LLMs) are increasingly deployed in enterprise settings, controlling model behavior based on user roles becomes an essential requirement. Existing safety methods typically assume uniform access and focus on preventing harmful or toxic outputs, without addressing role-specific access constraints. In this work, we investigate whether LLMs can be fine-tuned to generate responses that reflect the access privileges associated with different organizational roles. We explore three modeling strategies: a BERT-based classifier, an LLM-based classifier, and role-conditioned generation. To evaluate these approaches, we construct two complementary datasets. The first is adapted from existing instruction-tuning corpora through clustering and role labeling, while the second is synthetically generated to reflect realistic, role-sensitive enterprise scenarios. We assess model performance across varying organizational structures and analyze robustness to prompt injection, role mismatch, and jailbreak attempts.

Hassles and uplifts are psychological constructs of individuals’ positive or negative responses to daily minor incidents, with cumulative impacts on mental health. These concepts are largely overlooked in NLP, where existing tasks and models focus on identifying general sentiment expressed in text. These, however, cannot satisfy targeted information needs in psychological inquiry. To address this, we introduce Hassles and Uplifts Detection (HUD), a novel NLP application to identify these constructs in social media language.We evaluate various language models and task adaptation approaches on a probing dataset collected from a private, real-time emotional venting platform. Some of our models achieve F scores close to 80%. We also identify open opportunities to improve affective language understanding in support of studies in psychology.

Large language models (LLMs) have demonstrated strong capabilities in simulating social roles and generating human-like behaviors. However, their effectiveness in predicting real-world user behavior under continuous memory accumulation remains largely unexplored. Most existing studies focus on short-term interactions or static personas, neglecting the dynamic nature of users’ historical experiences in social media environments. To address this gap, we introduce FineRob, a novel dataset for fine-grained behavior prediction of social media users, which includes long-term memory traces from 1,866 users across three platforms. Each behavior is decomposed into three elements: object, type, and content, resulting in 78.6k QA records.We identify that as memory accumulates, prediction accuracy drops significantly due to the model’s difficulty in accessing detailed historical information. We further propose the OM-CoT fine-tuning framework to enhance the model’s ability to process and utilize long-term memory. Experimental results show that our method effectively reduces the performance degradation caused by memory growth, improving fine-grained behavior prediction. .

Large language models (LLMs) have significantly advanced generative applications in natural language processing (NLP). Recent trends in model architectures revolve around efficient variants of transformers or state-space/gated-recurrent models (SSMs, GRMs). However, prevailing SSM/GRM-based methods often emulate only a single attention head, potentially limiting their expressiveness. In this work, we propose MossNet, a novel mixture-of-state-space-experts architecture that emulates a linear multi-head attention (MHA). MossNet leverages a mixture-of-experts (MoE) implementation not only in channel-mixing multi-layered perceptron (MLP) blocks but also in the time-mixing SSM kernels to realize multiple “attention heads.” Extensive experiments on language modeling and downstream evaluations show that MossNet outperforms both transformer- and SSM-based architectures of similar model size and data budgets. Larger variants of MossNet, trained on trillions of tokens, further confirm its scalability and superior performance. In addition, real-device profiling on a Samsung Galaxy S24 Ultra and an Nvidia A100 GPU demonstrate favorable runtime speed and resource usage compared to similarly sized baselines. Our results suggest that MossNet is a compelling new direction for efficient, high-performing recurrent LLM architectures.

With the rapid development of spatial audio technologies today, applications in AR, VR and other scenarios have garnered extensive attention. Unlike traditional mono sound, spatial audio offers a more realistic and immersive auditory experience. Despite notable progress in the field, there remains a lack of comprehensive surveys that systematically organize and analyze these methods and their underlying technologies. In this paper, we provide a comprehensive overview of spatial audio and systematically review recent literature in the area. To address this, we chronologically outline existing work related to spatial audio and categorize these studies based on input-output representations, as well as generation and understanding tasks, thereby summarizing various research aspects of spatial audio. In addition, we review related datasets, evaluation metrics, and benchmarks, offering insights from both training and evaluation perspectives. Related materials are available at https://github.com/dieKarotte/ASAudio.

Recent advances in singing voice synthesis (SVS) have attracted substantial attention from both academia and industry. With the advent of large language models and novel generative paradigms, producing controllable, high‐fidelity singing voices has become an attainable goal. Yet the field still lacks a comprehensive survey that systematically analyzes deep‐learning‐based singing voice systems and their enabling technologies.To address the aforementioned issue, this survey first categorizes existing systems by task type and then organizes current architectures into two major paradigms: cascaded and end-to-end approaches. Moreover, we provide an in-depth analysis of core technologies, covering singing modeling and control techniques. Finally, we review relevant datasets, annotation tools, and evaluation benchmarks that support training and assessment. In appendix, we introduce training strategies and further discussion of SVS. This survey provides an up-to-date review of the literature on SVS models, which would be a useful reference for both researchers and engineers. Related materials are available at https://github.com/David-Pigeon/SyntheticSingers.

Recent advances in large language models (LLMs) have significantly improved mathematical problem-solving. Among various techniques, paraphrasing problem statements has emerged as a promising strategy to enhance model understanding and accuracy.We define twelve paraphrasing types grounded in mathematics education theory and analyze their impact on LLM performance across different configurations. To automate selection, we propose a Paraphrase Type Selector that predicts effective paraphrases for each problem.Experiments on MATH-500, SVAMP, and AIME shows consistent performance gain from paraphrased problems. On MATH-500 with LLaMA 3.1-8B, combining the original with the best five paraphrased problems improves accuracy by +8.4%, with the selector achieving an additional +1.33% gain.

Nigeria is the most populous country in Africa with a population of more than 200 million people. More than 500 languages are spoken in Nigeria and it is one of the most linguistically diverse countries in the world. Despite this, natural language processing (NLP) research has mostly focused on the following four languages: Hausa, Igbo, Nigerian-Pidgin, and Yoruba (i.e <1% of the languages spoken in Nigeria). This is in part due to the unavailability of textual data in these languages to train and apply NLP algorithms. In this work, we introduce Ibom—a dataset for machine translation and topic classification in four Coastal Nigerian languages from the Akwa Ibom State region: Anaang, Efik, Ibibio, and Oro. These languages are not represented in Google Translate or in major benchmarks such as Flores-200 or SIB-200. We focus on extending Flores-200 benchmark to these languages, and further align the translated texts with topic labels based on SIB-200 classification dataset. Our evaluation shows that current LLMs perform poorly on machine translation for these languages in both zero-and-few shot settings. However, we find the few-shot samples to steadily improve topic classification with more shots.

We present an automated pipeline for estimating Verb Frame Frequencies (VFFs), the frequency with which a verb appears in particular syntactic frames. VFFs provide a powerful window into syntax in both human and machine language systems, but existing tools for calculating them are limited in scale, accuracy, or accessibility. We use large language models (LLMs) to generate a corpus of sentences containing 476 English verbs. Next, by instructing an LLM to behave like an expert linguist, we had it analyze the syntactic structure of the sentences in this corpus. This pipeline outperforms two widely used syntactic parsers across multiple evaluation datasets. Furthermore, it requires far fewer resources than manual parsing (the gold-standard), thereby enabling rapid, scalable VFF estimation. Using the LLM parser, we produce a new VFF database with broader verb coverage, finer-grained syntactic distinctions, and explicit estimates of the relative frequencies of structural alternates commonly studied in psycholinguistics. The pipeline is easily customizable and extensible to new verbs, syntactic frames, and even other languages. We present this work as a proof of concept for automated frame frequency estimation, and release all code and data to support future research.

Multilingual vision–language models (VLMs) promise universal image–text retrieval, yet their social biases remain under‐explored.We perform the first systematic audit of four public multilingual CLIP variants—M‐CLIP, NLLB‐CLIP, CAPIVARA‐CLIP, and the debiased SigLIP‐2—covering ten languages that differ in resource availability and morphological gender marking.Using balanced subsets of FairFace and the PATA stereotype suite in a zero‐shot setting, we quantify race and gender bias and measure stereotype amplification.Contrary to the intuition that multilinguality mitigates bias, every model exhibits stronger gender skew than its English‐only baseline.CAPIVARA‐CLIP shows its largest biases precisely in the low‐resource languages it targets, while the shared encoder of NLLB‐CLIP and SigLIP‐2 transfers English gender stereotypes into gender‐neutral languages; loosely coupled encoders largely avoid this leakage.Although SigLIP‐2 reduces agency and communion skews, it inherits—and in caption‐sparse contexts (e.g., Xhosa) amplifies—the English anchor’s crime associations.Highly gendered languages consistently magnify all bias types, yet gender‐neutral languages remain vulnerable whenever cross‐lingual weight sharing imports foreign stereotypes.Aggregated metrics thus mask language‐specific “hot spots,” underscoring the need for fine‐grained, language‐aware bias evaluation in future multilingual VLM research.

Large Language Models (LLMs) have recently made significant advancements in tackling complex tasks, such as retrieving hard-to-find information and solving intricate problems. Consequently, various approaches have been proposed to integrate LLMs into recommender systems, primarily by embedding them within existing architectures or training them on the recommendation data. However, most existing methods fail to effectively incorporate user-item interaction signals into pretrained LLMs due to the modality gap between interaction data and the LLM’s internal knowledge. To address this challenge, we propose the Item-Language Model (ILM) to enhance LLMs for recommendation. ILM consists of two main components: An item-language representation learning module, where an ILM encoder is pretrained to generate text-aligned item representations. And an item-language co-training module, where the ILM encoder is integrated into a pretrained LLM for the recommendation tasks. Extensive experiments demonstrate the superior performance of our approach over several state-of-the-art methods, validating the importance of text-aligned item representations in bridging this modality gap. Our ablation studies further reveal the effectiveness of our model design for integrating the interaction knowledge into LLMs for recommendation tasks. Our code is available at: https://anonymous.4open.science/r/ILM-7AD4/.

LLM-as-a-Judge has emerged as a promising alternative to human evaluators across various tasks, yet inherent biases—particularly position bias, the tendency to favor solutions based on their position within the prompt—compromise its reliability. This exploratory study evaluates position bias in LLM judges across pairwise and list-wise comparison settings, introducing three metrics: repetition stability, position consistency, and preference fairness. Our experiments, involving 15 LLM judges across MTBench and DevBench with 22 tasks and approximately 40 solution-generating models, result in over 150,000 evaluation instances. We identify Judge-Level, Candidate-Level, and Task-Level factors contributing to bias. The findings confirm that position bias is not due to random chance and varies significantly across judges and tasks. While position bias is weakly influenced by the length of prompt components, it is strongly affected by the quality gap between solutions. Our agreement and disagreement analysis among judges further provides insights into the distribution of judging difficulty across the dataset, and highlights the potential for dataset modifications.

The rapid escalation from elementary school-level to frontier problems of the difficulty for LLM benchmarks in recent years seems to bring us close enough to the “last exam” for LLMs to surpass humanity. However, is the LLMs’ remarkable reasoning ability indeed coming from true intelligence by human standards, or are they actually reciting solutions witnessed during training at an Internet level? To study this problem, we propose RoR-Bench, a novel, multi-modal benchmark for detecting LLM’s recitation behavior when asked simple reasoning problems but with conditions subtly shifted, and conduct empirical analysis on our benchmark. Surprisingly, we found existing cutting-edge LLMs unanimously exhibits extremely severe recitation behavior; by changing one phrase in the condition, top models such as OpenAI-o1 and DeepSeek-R1 can suffer 60 percent performance loss on elementary school-level arithmetic and reasoning problems. Such findings are a wake-up call to the LLM community that compels us to reevaluate the true intelligence level of cutting-edge LLMs.

Uniform Meaning Representation (UMR) is a recently developed graph-based semantic representation, which expands on Abstract Meaning Representation (AMR) in a number of ways, in particular through the inclusion of document-level information and multilingual flexibility. In order to effectively adopt and leverage UMR for downstream tasks, efforts must be placed toward developing a UMR technological ecosystem. Though only a small amount of UMR annotations have been produced to date, in this work, we investigate the first approaches to producing text from multilingual UMR graphs. Exploiting the structural similarity between UMR and AMR graphs and the wide availability of AMR technologies, we introduce (1) a baseline approach which passes UMR graphs to AMR-to-text generation models, (2) a pipeline conversion of UMR to AMR, then using AMR-to-text generation models, and (3) a fine-tuning approach for both foundation models and AMR-to-text generation models with UMR data. Our best performing models achieve multilingual BERTscores of 0.825 for English and 0.882 for Chinese, a promising indication of the effectiveness of fine-tuning approaches for UMR-to-text generation even with limited UMR data.