NAACL 2025 Papers

We present TRACE, a novel system for live *common ground* tracking in situated collaborative tasks. With a focus on fast, real-time performance, TRACE tracks the speech, actions, gestures, and visual attention of participants, uses these multimodal inputs to determine the set of task-relevant propositions that have been raised as the dialogue progresses, and tracks the group’s epistemic position and beliefs toward them as the task unfolds. Amid increased interest in AI systems that can mediate collaborations, TRACE represents an important step forward for agents that can engage with multiparty, multimodal discourse.

Large language models (LLMs) have shown impressive capabilities across diverse settings, but still struggle as the length and complexity of the context increases. To address this challenge, we propose Thinking Recursively and Dynamically (ThReaD). THREAD frames model generation as a thread of execution that, based on the context, can run to completion or dynamically spawn new threads. By spawning, threads can offload work (e.g., thinking, retrieving information) to child threads, which only return tokens needed for the parent thread to do its work. We apply THREAD in the settings of LLM task solving and question answering, where the dynamic threading allows the model to recursively decompose the given task or question into progressively simpler sub-problems that can be solved by separate child threads. We test THREAD, implemented using a few-shot learning approach, on diverse benchmarks for agent tasks and data-grounded question answering. THREAD achieves state-of-the-art performance with GPT-4 and GPT-3.5 on these benchmarks, including ALFWorld, TextCraft, and WebShop, along with two new benchmarks, DataCommons QA and MIMIC-III ICU QA. In addition, THREAD outperforms existing frameworks by 10% to 50% absolute points with smaller models, including Llama-3-8b and CodeLlama-7b.

Large Language Models (LLMs) have achieved impressive results in Machine Translation (MT). However, human evaluations reveal that LLM-generated translations still contain various errors. Notably, feeding the error information back into the LLMs can facilitate self-refinement, leading to enhanced translation quality. Motivated by these findings, we introduce TEaR (Translate, Estimate, and Refine), a systematic LLM-based self-refinement framework aimed at bootstrapping translation performance. Our key results show that: 1) TEaR framework enables LLMs to improve their translation quality relying solely on self-feedback, measured by both automatic metrics and Multidimensional Quality Metrics (MQM) scores; 2) TEaR autonomously selects improvements, ensuring a robust translation quality baseline while outperforming both internal refinement and external feedback methods. Error analysis and iterative refinement experiments show its ability to continuously reduce translation errors and enhance overall translation quality. Our code and data are publicly available at https://github.com/fzp0424/self_correct_mt.

For program languages, testing plays a crucial role in the software development cycle, enabling the detection of bugs, vulnerabilities, and other undesirable behaviors. To perform software testing, testers need to write code snippets that execute the program under test. Recently, researchers have recognized the potential of large language models (LLMs) in software testing. However, there remains a lack of fair comparisons between different LLMs in terms of test case generation capabilities.In this paper, we propose TestEval, a novel benchmark for test case generation with LLMs. We collect 210 Python programs from an online programming platform, LeetCode, and design three different tasks: overall coverage, targeted line/branch coverage, and targeted path coverage. We further evaluate 17 popular LLMs, including both commercial and open-source ones, on TestEval. We find that generating test cases to cover specific program lines/branches/paths is still challenging for current LLMs, indicating a lack of ability to comprehend program logic and execution paths.

Current Large Language Models (LLMs) exhibit limited ability to understand table structures and to apply precise numerical reasoning, which is crucial for tasks such as table question answering and table-based fact verification. To address these challenges, we introduce our Tool-Augmented Reasoning framework for Tables (TART), which integrates LLMs with specialized tools. TART contains three key components: a table formatter to ensure accurate data representation, a tool maker to develop specific computational tools, and an explanation generator to maintain explainability. We also present the TOOLTAB dataset, a new benchmark designed specifically for training LLMs in table–tool integration. Our experiments indicate that TART achieves substantial improvements over existing methods (e.g., Chain-of-Thought) by improving both the precision of data processing and the clarity of the reasoning process. Notably, TART paired with CodeLlama achieves 90.0% of the accuracy of the closed-sourced LLM GPT-3.5-turbo, highlighting its robustness in diverse real-world scenarios. Both code and data are openly available at https://github.com/XinyuanLu00/TART.

Instruction tuning has achieved unprecedented success in NLP, turning large language models into versatile chatbots. However, the increasing variety and volume of instruction datasets demand significant computational resources. To address this, it is essential to extract a small and highly informative subset (i.e., Coreset) that achieves comparable performance to the full dataset. Achieving this goal poses non-trivial challenges: 1) data selection requires accurate data representations that reflect the training samples’ quality, 2) considering the diverse nature of instruction datasets, and 3) ensuring the efficiency of the coreset selection algorithm for large models. To address these challenges, we propose Task-Agnostic Gradient Clustered COreset Selection (TAGCOS). Specifically, we leverage sample gradients as the data representations, perform clustering to group similar data, and apply an efficient greedy algorithm for coreset selection. Experimental results show that our algorithm, selecting only 5% of the data, surpasses other unsupervised methods and achieves performance close to that of the full dataset.

Retrieval-Augmented Generation (RAG) has emerged as a prominent method for incorporating domain knowledge into Large Language Models (LLMs). While RAG enhances response relevance by incorporating retrieved domain knowledge in the context, retrieval errors can still lead to hallucinations and incorrect answers. To recover from retriever failures, domain knowledge is injected by fine-tuning the model to generate the correct response, even in the case of retrieval errors. However, we observe that without systematic knowledge augmentation, fine-tuned LLMs may memorize new information but still fail to extract relevant domain knowledge, leading to poor performance. In this work, we present a novel framework that significantly enhances the fine-tuning process by augmenting the training data in two ways – context augmentation and knowledge paraphrasing. In context augmentation, we create multiple training samples for a given QA pair by varying the relevance of the retrieved information, teaching the model when to ignore and when to rely on retrieved content. In knowledge paraphrasing, we finetune with multiple answers to the same question, enabling LLMs to better internalize specialized knowledge. To mitigate catastrophic forgetting due to fine-tuning, we add a domain-specific identifier to a question and also utilize a replay buffer containing general QA pairs. Experimental results demonstrate the efficacy of our method over existing techniques, achieving up to 10% relative gain in token-level recall while preserving the LLM’s generalization capabilities.

LLMs have boosted progress in many AI applications. Recently, there were attempts to distill the vast knowledge of LLMs into information retrieval systems. Those distillation methods mostly use output probabilities of LLMs which are unavailable in the latest black-box LLMs. We propose Syntriever, a training framework for retrievers using synthetic data from black-box LLMs. Syntriever consists of two stages. Firstly in the distillation stage, we synthesize relevant and plausibly irrelevant passages and augmented queries using chain-of-thoughts for the given queries. LLM is asked to self-verify the synthetic data for possible hallucinations, after which retrievers are trained with a loss designed to cluster the embeddings of relevant passages. Secondly in the alignment stage, we align the retriever with the preferences of LLMs. We propose a preference modeling called partial Plackett-Luce ranking to learn LLM preferences with regularization which prevents the model from deviating excessively from that trained in the distillation stage. Experiments show that Syntriever achieves state-of-the-art performances on benchmark datasets from various domains in nDCG@K. the source code is available in https://github.com/kmswin1/Syntriever

The NLP community has broadly focused on text-only approaches of cognitive state tasks, but audio can provide vital missing cues through prosody. We posit that text-to-speech models learn to track aspects of cognitive state in order to produce naturalistic audio, and that the signal audio models implicitly identify is orthogonal to the information that language models exploit. We present Synthetic Audio Data fine-tuning (SAD), a framework where we show that 7 tasks related to cognitive state modeling benefit from multimodal training on both text and zero-shot synthetic audio data from an off-the-shelf TTS system. We show an improvement over the text-only modality when adding synthetic audio data to text-only corpora. Furthermore, on tasks and corpora that do contain gold audio, we show our SAD framework achieves competitive performance with text and synthetic audio compared to text and gold audio.

Existing approaches to multilingual text detoxification are hampered by the scarcity of parallel multilingual datasets. In this work, we introduce a pipeline for the generation of multilingual parallel detoxification data. We also introduce SynthDetoxM, a manually collected and synthetically generated multilingual parallel text detoxification dataset comprising 16,000 high-quality detoxification sentence pairs across German, French, Spanish and Russian. The data was sourced from different toxicity evaluation datasets and then rewritten with nine modern open-source LLMs in few-shot setting. Our experiments demonstrate that models trained on the produced synthetic datasets have superior performance to those trained on the human-annotated MultiParaDetox dataset even in data limited setting. Models trained on SynthDetoxM outperform all evaluated LLMs in few-shot setting. We release our dataset and code to help further research in multilingual text detoxification.

Numerous adversarial defense methods have been proposed to strengthen the robustness of Natural Language Processing (NLP) models against adversarial attacks. However, many of these methods rely on predetermined linguistic knowledge and assume that attackers’ synonym candidates are known, which is often unrealistic. In this work, we investigate adversarial training in the embedding space and introduce a Fast Adversarial Training (FAT) method to improve the model robustness without requiring synonym awareness. FAT leverages single-step perturbation generation and effective perturbation initialization based on two key insights: (1) adversarial perturbations generated by single-step and multi-step gradient ascent are similar, and (2) perturbations generated on the same training sample across successive epochs exhibit resemblance. By employing single-step gradient ascent and leveraging historical perturbation information, FAT not only expedites the training process but also efficiently initializes perturbations. Extensive experiments demonstrate that FAT significantly enhances the robustness of popular NLP models under scenarios where synonyms are unknown, outperforming other defense baselines under various character-level and word-level attacks.

To improve the performance and explainability of LLM-based natural language reasoning, structured reasoning can be applied to generate explicitly structured proofs. Among different methods for structured reasoning, we specifically focus on backward chaining, where the proof goal is recursively decomposed to subgoals by searching and applying rules. We argue that current LLM-based backward chaining systems (e.g. Least-to-most prompting and LAMBADA) are incomplete, as they omit crucial algorithmic components identified from the classic backward chaining algorithm in computational logic (SLD Resolution). To this end, we propose a novel backward chaining system, SymBa (Symbolic Backward Chaining), which integrates a symbolic solver and an LLM. In SymBa, the solver controls the proof process, and the LLM is only called when the solver requires new information to complete the proof. Empowered by completeness, SymBa achieves a significant improvement in seven deductive, relational, and arithmetic reasoning benchmarks compared to the baselines.

Syllogistic reasoning is crucial for Natural Language Inference (NLI). This capability is particularly significant in specialized domains such as biomedicine, where it can support automatic evidence interpretation and scientific discovery. This paper presents SylloBio-NLI, a novel framework that leverages external ontologies to systematically instantiate diverse syllogistic arguments for biomedical NLI. We employ SylloBio-NLI to evaluate Large Language Models (LLMs) on identifying valid conclusions and extracting supporting evidence across 28 syllogistic schemes instantiated with human genome pathways. Extensive experiments reveal that biomedical syllogistic reasoning is particularly challenging for zero-shot LLMs, which achieve an average accuracy between 70% on generalized modus ponens and 23% on disjunctive syllogism. At the same time, we found that few-shot prompting can boost the performance of different LLMs, including Gemma (+14%) and LLama-3 (+43%). However, a deeper analysis shows that both techniques exhibit high sensitivity to superficial lexical variations, highlighting a dependency between reliability, models’ architecture, and pre-training regime. Overall, our results indicate that, while in-context examples have the potential to elicit syllogistic reasoning in LLMs, existing models are still far from achieving the robustness and consistency required for safe biomedical NLI applications.

Audio description (AD) is a crucial accessibility service provided to blind persons and persons with visual impairment, designed to convey visual information in acoustic form. Despite recent advancements in multilingual machine translation research, the lack of well-crafted and time-synchronized AD data impedes the development of audio description translation (ADT) systems that address the needs of multilingual countries such as Switzerland. Furthermore, most ADT systems rely on text alone, and it is unclear whether incorporating visual information from video clips improves the quality of ADT outputs.In this work, we introduce SwissADT, an **emerging** ADT system for three main Swiss languages and English, designed for future use by our industry partners. By collecting well-crafted AD data augmented with video clips in German, French, Italian, and English, and leveraging the power of Large Language Models (LLMs), we aim to enhance information accessibility for diverse language populations in Switzerland by automatically translating AD scripts to the desired Swiss language. Our extensive experimental ADT results, composed of both automatic and human evaluations of ADT quality, demonstrate the promising capability of SwissADT for the ADT task. We believe that combining human expertise with the generation power of LLMs can further enhance the performance of ADT systems, ultimately benefiting a larger multilingual target population.

Enterprise customers are increasingly adopting Large Language Models (LLMs) for critical communication tasks, such as drafting emails, crafting sales pitches, and composing casual messages. Deploying such models across different regions requires them to understand diverse cultural and linguistic contexts and generate safe and respectful responses. For enterprise applications, it is crucial to mitigate reputational risks, maintain trust, and ensure compliance by effectively identifying and handling unsafe or offensive language. To address this, we introduce SweEval, a benchmark simulating real-world scenarios with variations in tone (positive or negative) and context (formal or informal). The prompts explicitly instruct the model to include specific swear words while completing the task. This benchmark evaluates whether LLMs comply with or resist such inappropriate instructions and assesses their alignment with ethical frameworks, cultural nuances, and language comprehension capabilities. In order to advance research in building ethically aligned AI systems for enterprise use and beyond, we release the dataset and code: https://github.com/amitbcp/multilingual_profanity.

In this paper, we introduce Swan, a family of embedding models centred around the Arabic language, addressing both small-scale and large-scale use cases. Swan includes two variants: Swan-Small, based on ARBERTv2, and Swan-Large, built on ArMistral, a pretrained Arabic large language model. To evaluate these models, we propose ArabicMTEB, a comprehensive benchmark suite that assesses cross-lingual, multi-dialectal, multi-domain, and multi-cultural Arabic text embedding performance, covering eight diverse tasks and spanning 94 datasets. Swan-Large achieves state-of-the-art results, outperforming Multilingual-E5-large in most Arabic tasks, while the Swan-Small consistently surpasses Multilingual-E5-base. Our extensive evaluations demonstrate that Swan models are dialectally and culturally aware, excelling across various Arabic domains while offering significant monetary efficiency. This work significantly advances the field of Arabic language modelling and provides valuable resources for future research and applications in Arabic natural language processing. Our models and benchmarks will be made publicly accessible for research.

The rapid growth of the financial sector and the increasing focus on Environmental, Social, and Governance (ESG) considerations have created a pressing need for advanced natural language processing (NLP) tools. Despite recent advancements, there is still a notable absence of open-source Large Language Models (LLMs) that are proficient across both general finance and ESG domains, such as generating ESG reports. To address this gap, we introduce SusGen-30k, a high-quality, category-balanced dataset comprising seven financial NLP tasks. In addition, we propose TCFD-Bench, a benchmark designed to improve the evaluation of sustainability report generation. Our data-centric approach led to the development of a suite of models, SusGen-GPT, trained on the curated dataset. These models were evaluated across six adapted tasks and two off-the-shelf tasks, showing state-of-the-art performance, surpassing all other models except GPT-4. Remarkably, SusGen-GPT achieved an average score only 0.02 below GPT-4, despite using models with only 7-8B parameters compared to much larger GPT-4. This demonstrates the efficiency of our approach in delivering high performance with significantly fewer resources, addressing existing challenges and fostering further advancements in the financial and ESG research community.

Retrieval-augmented language models (RALMs) have recently shown great potential in mitigating the limitations of implicit knowledge in LLMs, such as untimely updating of the latest expertise and unreliable retention of long-tail knowledge. However, since the external knowledge base, as well as the retriever, can not guarantee reliability, potentially leading to the knowledge retrieved not being helpful or even misleading for LLM generation. In this paper, we introduce Supportiveness-based Knowledge Rewriting (SKR), a robust and pluggable knowledge rewriter inherently optimized for LLM generation. Specifically, we introduce the novel concept of “supportiveness”—which represents how effectively a knowledge piece facilitates downstream tasks. Based on supportiveness, we first design a training data curation strategy for our rewriter model, effectively identifying and filtering out poor or irrelevant rewrites to improve data efficacy. We then introduce the direct preference optimization (DPO) algorithm to align the generated rewrites to optimal supportiveness, guiding the rewriter model to summarize augmented content that better improves the final response. Comprehensive evaluations across six popular knowledge-intensive tasks and four LLMs have demonstrated the effectiveness and superiority of SKR. With only 7B parameters, SKR has shown better knowledge rewriting capability over GPT-4.

Superlatives are used to single out elements with a maximal/minimal property. Semantically, superlatives perform a set comparison: something (or some things) has the min/max property out of a set. As such, superlatives provide an ideal phenomenon for studying implicit phenomena and discourse restrictions. While this comparison set is often not explicitly defined, its (implicit) restrictions can be inferred from the discourse context the expression appears in. In this work we provide an extensive computational study on the semantics of superlatives. We propose a unified account of superlative semantics which allows us to derive a broad-coverage annotation schema. Using this unified schema we annotated a multi-domain dataset of superlatives and their semantic interpretations. We specifically focus on interpreting implicit or ambiguous superlative expressions, by analyzing how the discourse context restricts the set of interpretations. In a set of experiments we then analyze how well models perform at variations of predicting superlative semantics, with and without context. We show that the fine-grained semantics of superlatives in context can be challenging for contemporary models, including GPT-4.