NAACL 2025 Papers

This paper provides a proof of concept that audio of tabletop role-playing games (TTRPG) could serve as a challenge for diarization systems. TTRPGs are carried out mostly by conversation. Participants often alter their voices to indicate that they are talking as a fictional character. Audio processing systems are susceptible to voice conversion with or without technological assistance. TTRPG present a conversational phenomenon in which voice conversion is an inherent characteristic for an immersive gaming experience. This could make it more challenging for diarizers to pick the real speaker and determine that impersonating is just that. We present the creation of a small TTRPG audio dataset and compare it against the AMI and the ICSI corpus. The performance of two diarizers, pyannote.audio and wespeaker, were evaluated. We observed that TTRPGs’ properties result in a higher confusion rate for both diarizers.Additionally, wespeaker strongly underestimates the number of speakers in the TTRPG audio files.We propose TTRPG audio as a promising challenge for diarization systems.

Recent works have explored using language models for planning problems. One approach examines translating natural language descriptions of planning tasks into structured planning languages, such as the planning domain definition language (PDDL). Existing evaluation methods struggle to ensure semantic correctness and rely on simple or unrealistic datasets. To bridge this gap, we introduce Planetarium, a benchmark designed to evaluate language models’ ability to generate PDDL code from natural language descriptions of planning tasks. Planetarium features a novel PDDL equivalence algorithm that flexibly evaluates the correctness of generated PDDL against ground truth, along with a dataset of 145,918 text-to-PDDL pairs across 73 unique state combinations with varying levels of difficulty. Finally, we evaluate several API-access and open-weight language models that reveal this task’s complexity. For example, 96.1% of the PDDL problem descriptions generated by GPT-4o are syntactically parseable, 94.4% are solvable, but only 24.8% are semantically correct, highlighting the need for a more rigorous benchmark for this problem.

Recent studies have raised concerns about the potential threats large language models (LLMs) pose to academic integrity and copyright protection. Yet, their investigation is predominantly focused on literal copies of original texts. Also, how LLMs can facilitate the detection of LLM-generated plagiarism remains largely unexplored. To address these gaps, we introduce PlagBench, a dataset of 46.5K synthetic text pairs that represent three major types of plagiarism: verbatim copying, paraphrasing, and summarization. These samples are generated by three advanced LLMs. We rigorously validate the quality of PlagBench through a combination of fine-grained automatic evaluation and human annotation. We then utilize this dataset for two purposes: (1) to examine LLMs’ ability to transform original content into accurate paraphrases and summaries, and (2) to evaluate the plagiarism detection performance of five modern LLMs alongside three specialized plagiarism checkers. Our results show that GPT-3.5 Turbo can produce high-quality paraphrases and summaries without significantly increasing text complexity compared to GPT-4 Turbo. However, in terms of detection, GPT-4 outperforms other LLMs and commercial detection tools by 20%, highlights the evolving capabilities of LLMs not only in content generation but also in plagiarism detection. Data and source code are available at https://github.com/Brit7777/plagbench.

This work presents a speech-to-text system “Pisets” for scientists and journalists which is based on a three-component architecture aimed at improving speech recognition accuracy while minimizing errors and hallucinations associated with the Whisper model. The architecture comprises primary recognition using Wav2Vec2, false positive filtering via the Audio Spectrogram Transformer (AST), and final speech recognition through Whisper. The implementation of curriculum learning methods and the utilization of diverse Russian-language speech corpora significantly enhanced the system’s effectiveness. Additionally, advanced uncertainty modeling techniques were introduced, contributing to further improvements in transcription quality. The proposed approaches ensure robust transcribing of long audio data across various acoustic conditions compared to WhisperX and the usual Whisper model. The source code of “Pisets” system is publicly available at GitHub: https://github.com/bond005/pisets.

Large language models (LLMs) are typically optimized for resource-rich languages like English, exacerbating the gap between high-resource and underrepresented languages. This work presents a detailed analysis of strategies for developing a model capable of following instructions in a low-resource language, specifically Basque, by focusing on three key stages: pre-training, instruction tuning, and alignment with human preferences. Our findings demonstrate that continual pre-training with a high-quality Basque corpus of around 600 million words improves natural language understanding (NLU) of the foundational model by over 12 points. Moreover, instruction tuning and human preference alignment using automatically translated datasets proved highly effective, resulting in a 24-point improvement in instruction-following performance. The resulting models, Llama-eus-8B and Llama-eus-8B-instruct, establish a new state-of-the-art for Basque in the sub-10B parameter category.

Task-Oriented Dialogue (TOD) systems are designed to fulfill user requests through natural language interactions, yet existing systems often produce generic, monotonic responses that lack individuality and fail to adapt to users’ personal attributes. To address this, we introduce PicPersona-TOD, a novel dataset that incorporates user images as part of the persona, enabling personalized responses tailored to user-specific factors such as age or emotional context. This is facilitated by first impressions, dialogue policy-guided prompting, and the use of external knowledge to reduce hallucinations. Human evaluations confirm that our dataset enhances user experience, with personalized responses contributing to a more engaging interaction. Additionally, we introduce a new NLG model, Pictor, which not only personalizes responses, but also demonstrates robust performance across unseen domains.

AIs can beat humans in game environments; however, how helpful those agents are to human remains understudied. We augment Cicero, a natural language agent that demonstrates superhuman performance in Diplomacy, to generate both move and message advice based on player intentions. A dozen Diplomacy games with novice and experienced players, with varying advice settings, show that some of the generated advice is beneficial. It helps novices compete with experienced players and in some instances even surpass them. The mere presence of advice can be advantageous, even if players do not follow it.

Personalizing large language models (LLMs) is essential for delivering tailored interactions that improve user experience. Many existing personalization methods require fine-tuning LLMs for each user, rendering them prohibitively expensive for widespread adoption. Although retrieval-based approaches offer a more compute-efficient alternative, they still depend on large, high-quality datasets that are not consistently available for all users. To address this challenge, we propose Chameleon, a scalable and efficient personalization approach that uses (1) self-generated personal preference data and (2) representation editing to enable quick and cost-effective personalization. Our experiments on various tasks, including those from the LaMP personalization benchmark, show that Chameleon efficiently adapts models to personal preferences, improving instruction-tuned models and outperforms two personalization baselines by an average of 40% across two model architectures.

Suggested questions (SQs) provide an effective initial interface for users to engage with their documents in AI-powered reading applications. In practical reading sessions, users have diverse backgrounds and reading goals, yet current SQ features typically ignore such user information, resulting in homogeneous or ineffective questions. We introduce a pipeline that generates personalized SQs by incorporating reader profiles (professions and reading goals) and demonstrate its utility in two ways: 1) as an improved SQ generation pipeline that produces higher quality and more diverse questions compared to current baselines, and 2) as a data generator to fine-tune extremely small models that perform competitively with much larger models on SQ generation. Our approach can not only serve as a drop-in replacement in current SQ systems to immediately improve their performance but also help develop on-device SQ models that can run locally to deliver fast and private SQ experience.

Large language models (LLMs) demonstrate exceptional capabilities in various scenarios. However, they suffer from much redundant information and are sensitive to the position of key information in long context scenarios. To address these challenges, we present Perception Compressor, a training-free prompt compression framework. It includes a perception retriever that leverages guiding questions and instruction to retrieve the most relevant demonstrations, a dual-slope ratio allocator to dynamically allocate compression ratios and open-book ratios, and a semi-guided iterative compression that retains key information at the token level while removing tokens that distract the LLM. We conduct extensive experiments on long context benchmarks, i.e., NaturalQuestions, LongBench, and MuSiQue. Experiment results show that Perception Compressor outperforms existing methods by a large margin, achieving state-of-the-art performance.

Large language models predominantly reflect Western cultures, largely due to the dominance of English-centric training data. This imbalance presents a significant challenge, as LLMs are increasingly used across diverse contexts without adequate evaluation of their cultural competence in non-English languages, including Persian. To address this gap, we introduce PerCul, a carefully constructed dataset designed to assess the sensitivity of LLMs toward Persian culture. PerCul features story-based, multiple-choice questions that capture culturally nuanced scenarios.Unlike existing benchmarks, PerCul is curated with input from native Persian annotators to ensure authenticity and to prevent the use of translation as a shortcut. We evaluate several state-of-the-art multilingual and Persian-specific LLMs, establishing a foundation for future research in cross-cultural NLP evaluation. Our experiments demonstrate a 11.3% gap between best closed source model and layperson baseline while the gap increases to 21.3% by using the best open-weight model. You can access the dataset from here:https://huggingface.co/datasets/teias-ai/percul

Large Language Models (LLMs) have been recently adopted in interactive systems requiring communication. As the false belief in a model can harm the usability of such systems, LLMs should not have cognitive biases that humans have. Psychologists especially focus on the False Consensus Effect (FCE), a cognitive bias where individuals overestimate the extent to which others share their beliefs or behaviors, because FCE can distract smooth communication by posing false beliefs. However, previous studies have less examined FCE in LLMs thoroughly, which needs more consideration of confounding biases, general situations, and prompt changes. Therefore, in this paper, we conduct two studies to examine the FCE phenomenon in LLMs. In Study 1, we investigate whether LLMs have FCE. In Study 2, we explore how various prompting styles affect the demonstration of FCE. As a result of these studies, we identified that popular LLMs have FCE. Also, the result specifies the conditions when FCE becomes more or less prevalent compared to normal usage.

We present PeerQA, a real-world, scientific, document-level Question Answering (QA) dataset. PeerQA questions have been sourced from peer reviews, which contain questions that reviewers raised while thoroughly examining the scientific article. Answers have been annotated by the original authors of each paper. The dataset contains 579 QA pairs from 208 academic articles, with a majority from ML and NLP, as well as a subset of other scientific communities like Geoscience and Public Health.PeerQA supports three critical tasks for developing practical QA systems: Evidence retrieval, unanswerable question classification, and answer generation. We provide a detailed analysis of the collected dataset and conduct experiments establishing baseline systems for all three tasks. Our experiments and analyses reveal the need for decontextualization in document-level retrieval, where we find that even simple decontextualization approaches consistently improve retrieval performance across architectures. On answer generation, PeerQA serves as a challenging benchmark for long-context modeling, as the papers have an average size of 12k tokens.

Existing multimodal summarization approaches struggle with scenarios involving numerous images as input, leading to a heavy load for readers. Summarizing both the input text and numerous images helps readers quickly grasp the key points of multimodal input. This paper introduces a novel task, Numerous Images-Oriented Multimodal Summarization (NIMMS). To benchmark this task, we first construct the dataset based on a public multimodal summarization dataset. Considering that most existing metrics evaluate summaries from a unimodal perspective, we propose a new Multimodal Information evaluation (M-info) method, measuring the differences between the generated summary and the multimodal input. Finally, we compare various summarization methods on NIMMS and analyze associated challenges. Experimental results have shown that M-info correlates more closely with human judgments than five widely used metrics. Meanwhile, existing models struggle with summarizing numerous images. We hope that this research will shed light on the development of multimodal summarization. Furthermore, our code and dataset will be released to the public.

This paper presents Patent-CR, the first dataset created for the patent claim revision task in English. It includes both initial patent applications rejected by patent examiners and the final granted versions. Unlike normal text revision tasks that predominantly focus on enhancing sentence quality, such as grammar correction and coherence improvement, patent claim revision aims at ensuring the claims meet stringent legal criteria. These criteria are beyond novelty and inventiveness, including clarity of scope, technical accuracy, language precision, and legal robustness. We assess various large language models (LLMs) through professional human evaluation, including general LLMs with different sizes and architectures, text revision models, and domain-specific models. Our results indicate that LLMs often bring ineffective edits that deviate from the target revisions. In addition, domain-specific models and the method of fine-tuning show promising results. Notably, GPT-4 outperforms other tested LLMs, but further revisions are still necessary to reach the examination standard. Furthermore, we demonstrate the inconsistency between automated and human evaluation results, suggesting that GPT-4-based automated evaluation has the highest correlation with human judgment. This dataset, along with our preliminary empirical research, offers invaluable insights for further exploration in patent claim revision.

Large pre-trained Vision-Language Models (VLMs) have revolutionized both computer vision and natural language processing. Despite their success, adversarial examples can still mislead VLMs into producing incorrect results. This work focuses on boosting the adversarial robustness of VLMs by searching for text prompts at the word level, rather than optimizing continuous textual embeddings. We introduce Parameter-Free Prompt Tuning (PFPT) to learn defense words that enhance resilience against adversarial attacks when appended to existing prompts, thereby offering ease of use due to the simplicity of this approach. These defense words are naturally present in the inherent vocabulary of VLMs, providing a human-readable property. PFPT employs a coarse-to-fine strategy with carefully designed optimization objectives to guide the word search. Extensive experiments demonstrate our method’s superiority over hand-engineered prompts and other state-of-the-art methods. PFPT significantly boosts accuracy and robustness, outperforming hand-engineered prompts with average gains of +4.9% and +5.8%, respectively (epsilon=1/255).

Large language models (LLMs) have become the norm in natural language processing (NLP), excelling in few-shot in-context learning (ICL) with their remarkable abilities. Nonetheless, the success of ICL largely hinges on the choice of few-shot demonstration examples, making the selection process increasingly crucial. Existing methods have delved into optimizing the quantity and semantic similarity of these examples to improve ICL performances. However, our preliminary experiments indicate that the effectiveness of ICL is limited by the length of the input context. Moreover, varying combinations of few-shot demonstration examples can significantly boost accuracy across different test samples. To address this, we propose a novel method named parallel in-context learning (ParaICL) that effectively utilizes all demonstration examples without exceeding the manageable input context length. ParaICL employs parallel batching to distribute demonstration examples into different batches according to the semantic similarities of the questions in the demonstrations to the test question. It then computes normalized batch semantic scores for each batch. A weighted average semantic objective, constrained by adaptive plausibility, is applied to select the most appropriate tokens. Through extensive experiments, we validate the effectiveness of ParaICL and conduct ablation studies to underscore its design rationale. We further demonstrate that ParaICL can seamlessly integrate with existing methods.

Recent advancements in large language models have revolutionized text generation with their remarkable capabilities. These models can produce controlled texts that closely adhere to specific requirements when prompted appropriately. However, designing an optimal prompt to control multiple attributes simultaneously can be challenging. A common approach is to linearly combine single-attribute models, but this strategy often overlooks attribute overlaps and can lead to conflicts. Therefore, we propose a novel combination strategy inspired by the Law of Total Probability and Conditional Mutual Information Minimization on generative language models. This method has been adapted for single-attribute control scenario and is termed the Palette of Language Models due to its theoretical linkage between attribute strength and generation style, akin to blending colors on an artist’s palette. Moreover, positive correlation and attribute enhancement are advanced as theoretical properties to guide a rational combination strategy design. We conduct experiments on both single control and multiple control settings, and achieve surpassing results.

Generalized zero-shot intent detection (GZID) aims to recognize the labels of utterances from both seen and unseen intents by utilizing the knowledge learned from seen intents. Enhancing the generalization ability from seen intents to unseen intents is a key challenge in the GZID setting. Existing methods attempt to tackle this challenge by distinguishing unseen intents from seen intents or focusing on enhancing the model discriminability. However, the challenge is not solved substantially as they ignore to promote the representation learning ability of the model itself and neglect to strengthen the model adaptability to new tasks, resulting in overfitting on the seen intents. In this paper, we propose a pairwise prompt-based tuning model with parameter efficient fast adaptation which involves two training steps. In the first step, we leverage hybrid contrastive learning in discriminant space and masked language modeling to make predictions at both sentence and token levels, which can enhance the model discriminability and representation learning ability respectively. In the second step, we design a pipeline for generating and filtering unseen data by only providing unseen intent labels, and utilize parameter-efficient fine-tuning to quickly adapt to unseen intents. Experiments on four intent detection datasets demonstrate that our two-step training method has better comprehension and generalization capabilities.