EMNLP 2025 Papers

While large language models (LLMs) excel at machine translation (MT), the impact of how LLMs utilize different forms of contextual information on discourse-level phenomena remains underexplored. We systematically investigate how different forms of context such as prior source sentences, models’ generated hypotheses, and reference translations influence standard MT metrics and specific discourse phenomena (formality, pronoun selection, and lexical cohesion). Evaluating multiple LLMs across multiple domains and language pairs, our findings consistently show that context boosts both translation and discourse-specific performance. Notably, the context strategy of combining source text with the model’s own prior hypotheses effectively improves discourse consistency without gold references, demonstrating effective use of model’s own imperfect generations as diverse contextual cues.

Frontier language models are increasingly based on the Mixture of Experts (MoE) architecture, boosting the efficiency of training and inference by sparsely activating parameters. Nevertheless, training from scratch on trillions of tokens remains so expensive that most users can only finetune these models. In this work, we combine parameter reuse of dense models for the MoE layers ("upcycling”) with a novel, adaptive Nexus router that can integrate new experts into an existing trained model without hurting the performance on previous domains. Our router leverages the knowledge of each expert’s training data distribution via domain embeddings to initialize the router, improving specialization and allowing it to adapt faster to new domains than a standard MoE router. Nexus overturns the strict sequential separation between training and finetuning in classical approaches, allowing more powerful improvements to existing models at a later stage through long token-horizon trainings on new pretraining data. Our experiments show that Nexus achieves a relative gain of up to 2.1% over the baseline for initial upcycling, and an 18.8% relative gain for extending the MoE to a new domain with a new expert by using limited finetuning data. This flexibility of Nexus can power an open-source ecosystem where every user continuously assembles their own MoE-mix from a multitude of dense models.

Fact-checking real-world claims, particularly numerical claims, is inherently complex that require multistep reasoning and numerical reasoning for verifying diverse aspects of the claim. Although large language models (LLMs) including reasoning models have made tremendous advances, they still fall short on fact-checking real-world claims that require a combination of compositional and numerical reasoning. They are unable to understand nuance of numerical aspects, and are also susceptible to the reasoning drift issue, where the model is unable to contextualize diverse information resulting in misinterpretation and backtracking of reasoning process. In this work, we systematically explore scaling test-time compute (TTS) for LLMs on the task of fact-checking complex numerical claims, which entails eliciting multiple reasoning paths from an LLM. We train a verifier model (VERIFIERFC) to navigate this space of possible reasoning paths and select one that could lead to the correct verdict. We observe that TTS helps mitigate the reasoning drift issue, leading to significant performance gains for fact-checking numerical claims. To improve compute efficiency in TTS, we introduce an adaptive mechanism that performs TTS selectively based on the perceived complexity of the claim. This approach achieves 1.8x higher efficiency than standard TTS, while delivering a notable 18.8% performance improvement over single-shot claim verification methods. Our code and data can be found at https://github.com/VenkteshV/VerifierFC

Language is not only a tool for communication but also a medium for human cognition and reasoning. If, as linguistic relativity suggests, the structure of language shapes cognitive patterns, then large language models (LLMs) trained on human language may also internalize the habitual logical structures embedded in different languages. To examine this hypothesis, we introduce BICAUSE, a structured bilingual dataset for causal reasoning, which includes semantically aligned Chinese and English samples in both forward and reversed causal forms. Our study reveals three key findings: (1) LLMs exhibit typologically aligned attention patterns, focusing more on causes and sentence-initial connectives in Chinese, while showing a more balanced distribution in English. (2) Models internalize language-specific preferences for causal components order and often rigidly apply them to atypical inputs, leading to degraded performance, especially in Chinese. (3) When causal reasoning succeeds, model representations converge toward semantically aligned abstractions across languages, indicating a shared understanding beyond surface form. Overall, these results suggest that LLMs not only mimic surface linguistic forms but also internalize the reasoning biases shaped by language. Rooted in cognitive linguistic theory, this phenomenon is for the first time empirically verified through structural analysis of model internals.

Recent text-to-SQL models have achieved strong performance, but their effectiveness remains largely confined to SQLite due to dataset limitations. However, real-world applications require SQL generation across multiple dialects with varying syntax and specialized features, which remains a challenge for current models. The main obstacle in building a dialect-aware model lies in acquiring high-quality dialect-specific data. Data generated purely through static prompting—without validating SQLs via execution—tends to be noisy and unreliable. Moreover, the lack of real execution environments in the training loop prevents models from grounding their predictions in executable semantics, limiting generalization despite surface-level improvements from data filtering. This work introduces ExeSQL, a text-to-SQL framework with execution-driven, agentic bootstrapping. The method consists of iterative query generation, execution-based filtering (e.g., rejection sampling), and preference-based training, enabling the model to adapt to new SQL dialects through verifiable, feedback-guided learning. Experiments show that ExeSQL bridges the dialect gap in text-to-SQL, achieving average improvements of 15.2%, 10.38%, and 4.49% over GPT-4o on PostgreSQL, MySQL, and Oracle, respectively, across multiple datasets of varying difficulty.

Standard language models employ unique, monolithic embeddings for each token, potentially limiting their ability to capture the multifaceted nature of word meanings. We investigate whether tokens can be more effectively represented through a compositional structure that accumulates diverse semantic facets. To explore this, we propose Aggregate Semantic Grouping (ASG), a novel approach leveraging Product Quantization (PQ). We apply ASG to standard transformer architectures (mBERT, XLM-R, mT5) and evaluate this representational scheme across diverse tasks (NLI, NER, QA), as well as a biomedical domain-specific benchmark (BC5CDR) using BioBERT. Our findings demonstrate that representing tokens compositionally via ASG achieves extreme compression in embedding parameters (0.4–0.5%) while maintaining >95% task performance relative to the base model, even in generative tasks and extends to both cross lingual transfer and domain-specific settings. These results validate the principle that tokens can be effectively modeled as combinations of shared semantic building blocks. ASG offers a simple yet concrete method for achieving this, showcasing how compositional representations can capture linguistic richness while enabling compact yet semantically rich models.

Text-rich images are ubiquitous in real-world applications, serving as a critical medium for conveying complex information and facilitating accessibility.Despite recent advances driven by Multimodal Large Language Models (MLLMs), existing benchmarks suffer from limited scale, fragmented scenarios, and evaluation protocols that fail to fully capture holistic image understanding.To address these gaps, we present TIU-Bench, a large-scale, multilingual benchmark comprising over 100,000 full-image annotations and 22,000 rigorously validated question-answer (QA) pairs that span 18 subtasks across diverse real-world scenarios.TIU-Bench introduces a novel full-image structured output format that jointly models geometric, textual, and relational information, enabling fine-grained evaluation of perception and reasoning capabilities. Furthermore, we propose a two-stage understanding framework named T2TIU, which first generates a structured representation of the entire image and subsequently conducts reasoning on this representation to address complex visual-textual queries.Extensive experiments on 10 state-of-the-art generative models highlight the challenges and opportunities in advancing text-rich image understanding.Our benchmark and framework provide a comprehensive platform for developing and evaluating next-generation multimodal AI systems.

Machine translation quality has steadily improved over the years, achieving near-perfect translations in recent benchmarks.These high-quality outputs make it difficult to distinguish between state-of-the-art models and to identify areas for future improvement.In this context, automatically identifying texts where machine translation systems struggle holds promise for developing more discriminative evaluations and guiding future research.In this work, we address this gap by formalizing the task of translation difficulty estimation, defining a text’s difficulty based on the expected quality of its translations.We introduce a new metric to evaluate difficulty estimators and use it to assess both baselines and novel approaches.Finally, we demonstrate the practical utility of difficulty estimators by using them to construct more challenging benchmarks for machine translation. Our results show that dedicated models outperform both heuristic-based methods and LLM-as-a-judge approaches, with sentinel-src achieving the best performance.Thus, we release two improved models for difficulty estimation, sentinel-src-24 and sentinel-src-25, which can be used to scan large collections of texts and select those most likely to challenge contemporary machine translation systems.

Infodemics and health misinformation have significant negative impact on individuals and society, exacerbating confusion and increasing hesitancy in adopting recommended health measures. Recent advancements in generative AI, capable of producing realistic, human-like text and images, have significantly accelerated the spread and expanded the reach of health misinformation, resulting in an alarming surge in its dissemination. To combat the infodemics, most existing work has focused on developing misinformation datasets from social media and fact-checking platforms, but has faced limitations in topical coverage, inclusion of AI-generation, and accessibility of raw content. To address these gaps, we present MM-Health, a large scale multimodal misinformation dataset in the health domain consisting of 34,746 news article encompassing both textual and visual information. MM-Health includes human-generated multimodal information (5,776 articles) and AI-generated multimodal information (28,880 articles) from various SOTA generative AI models. Additionally, We benchmarked our dataset against three tasks—reliability checks, originality checks, and fine-grained AI detection—demonstrating that existing SOTA models struggle to accurately distinguish the reliability and origin of information. Our dataset aims to support the development of misinformation detection across various health scenarios, facilitating the detection of human and machine-generated content at multimodal levels.

The exponential growth of scientific publications has overwhelmed reviewers and researchers, with top conferences receiving thousands of submissions annually. Reviewers must assess feasibility, novelty, and impact under tight deadlines, often lacking tools to identify relevant prior work. Early-career researchers face similar challenges, with limited support to navigate fast-evolving fields. Existing LLM-based systems struggle with static retrieval, surface-level features, and lack multi-hop reasoning, leading to shallow or hallucinated assessments. Scientific evaluation requires a deep, relational understanding, which current retrieval-augmented generation (RAG) methods fail to achieve. We introduce SciCompanion, a graph-grounded reasoning framework for structured scientific evaluation. Given a paper or abstract-like input, SciCompanion builds a dynamic knowledge graph from recent publications, domain-specific databases, and curated metadata. It employs multi-hop reasoning to iteratively construct contextual graphs and generate structured critiques, enabling deeper exploration of scientific literature. Unlike sentiment-biased LLM evaluations, SciCompanion directly optimizes retrieval and graph refinement using Group Relative Policy Optimization (GRPO), producing reviews aligned with expert judgments. Experiments on ICLR and ACL datasets show that SciCompanion reduces evaluation error by over 30% compared to prompting-only baselines and allows smaller models to outperform larger ones. Evaluations across three datasets, using metrics for retrieval accuracy, semantic overlap, and multi-hop sensitivity, along with a case study, demonstrate SciCompanion’s robustness and versatility.

Compound AI (CAI) systems, also referred to as LLM Agents, combine LLMs with retrievers and tools to enable information-seeking applications in the real-world. Thus, ensuring these systems perform reliably is critical. However, traditional evaluation using benchmark datasets and aggregate metrics often fails to capture their true operational performance. This is because understanding the operational efficacy of these information-seeking systems requires the ability to probe their behavior across a spectrum of simulated scenarios to identify potential failure modes. Thus, we present a behavior-driven evaluation framework that generates test specifications - explicit descriptions of expected system behaviors in specific scenarios - aligned with real usage contexts. These test specifications serve as formal declarations of system requirements that are then automatically transformed into concrete test cases. Specifically, our framework operates in two phases: (1) generating diverse test specifications via submodular optimization over semantic diversity and document coverage of the tests, and (2) implementing these specifications through graph-based pipelines supporting both tabular and textual sources. Evaluations on QuAC & HybriDialogue datasets, across SoTA LLMs, reveal that our framework identifies failure modes missed by traditional metrics, demonstrating failure rates twice as high as human-curated datasets.

Human empathy builds on the shared pragmatic common ground among different languages. However, existing human empathy data is limited to English. Inspired by multilingual coactivation as the neurocognitive underpinning of human bilingual proficiency, which predicts empathy, we integrate language-independent diffusion processes to facilitate the cross-lingual transfer of empathy. Taking Chinese language varieties as the target domain, automatic and human evaluations demonstrate successful transfers of source empathy into target contexts without compromising linguistic naturalness. The results of this work offer empirical clues on the importance of pragmatic transferability of empathy and its cross-lingual effects in conversation.

Natural Language to SQL (NL2SQL) provides a new model-centric paradigm that simplifies database access for non-technical users by converting natural language queries into SQL commands. Recent advancements, particularly those integrating Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) reasoning, have made significant strides in enhancing NL2SQL performance. However, challenges such as inaccurate task decomposition and keyword extraction by LLMs remain major bottlenecks, often leading to errors in SQL generation. While existing datasets aim to mitigate these issues by fine-tuning models, they struggle with over-fragmentation of tasks and lack of domain-specific keyword annotations, limiting their effectiveness.To address these limitations, we present DeKeyNLU, a novel dataset which contains 1,500 meticulously annotated QA pairs aimed at refining task decomposition and enhancing keyword extraction precision for the RAG pipeline. Fine-tuned with DeKeyNLU, we propose DeKeySQL, a RAG-based NL2SQL pipeline that employs three distinct modules for user question understanding, entity retrieval, and generation to improve SQL generation accuracy. We benchmarked multiple model configurations within DeKeySQL RAG pipeline. Experimental results demonstrate that fine-tuning with DeKeyNLU significantly improves SQL generation accuracy on both BIRD (62.31% to 69.10%) and Spider (84.2% to 88.7%) dev datasets.

Web banner advertisements, which are placed on websites to guide users to a targeted landing page (LP), are still often selected manually because human preferences are important in selecting which ads to deliver. To automate this process, we propose a new benchmark, BannerBench, to evaluate the human preference-driven banner selection process using vision-language models (VLMs). This benchmark assesses the degree of alignment with human preferences in two tasks: a ranking task and a best-choice task, both using sets of five images derived from a single LP. Our experiments show that VLMs are moderately correlated with human preferences on the ranking task. In the best-choice task, most VLMs perform close to chance level across various prompting strategies. These findings suggest that although VLMs have a basic understanding of human preferences, most of them struggle to pinpoint a single suitable option from many candidates.

Scaling test-time computation, generating and analyzing multiple or sequential outputs for a single input, has become a promising strategy for improving the reliability and quality of large language models (LLMs), as evidenced by advances in uncertainty quantification and multi-step reasoning. A key shared component is semantic clustering, which groups outputs that differ in form but convey the same meaning. Semantic clustering enables estimation of the distribution over the semantics of outputs and helps avoid redundant exploration of reasoning paths. However, existing approaches typically rely on external models, which introduce substantial computational overhead and often fail to capture context-aware semantics. We propose Latent Semantic Clustering (LSC), a lightweight and context-sensitive method that leverages the generator LLM’s internal hidden states for clustering, eliminating the need for external models. Our extensive experiment across various LLMs and datasets shows that LSC significantly improves the computational efficiency of test-time scaling while maintaining or exceeding the performance of existing methods.

Mathematical reasoning poses significant challenges for Large Language Models (LLMs) due to its demand for multi-step reasoning and abstract conceptual integration. While recent test-time scaling techniques rely heavily on high-quality, challenging problems, the scarcity of Olympiad-level math problems remains a bottleneck. We introduce CogAtom, a novel cognitive atom-based framework for synthesizing mathematically rigorous and cognitively diverse problems. Unlike prior approaches, CogAtom models problem construction as a process of selecting and recombining fundamental reasoning units, cognitive atoms, extracted from human-authored solutions. A diversity-promoting random walk algorithm enables exploration of the cognitive atom space, while a constraint-based recombination mechanism ensures logical soundness and structural validity. The combinatorial nature of the graph structure provides a near-infinite space of reasoning paths, and the walk algorithm systematically explores this space to achieve large-scale synthesis of high-quality problems; meanwhile, by controlling the number of cognitive atoms, we can precisely adjust problem difficulty, ensuring diversity, scalability, and controllability of the generated problems. Experimental results demonstrate that CogAtom outperforms existing methods in accuracy, reasoning depth, and diversity, generating problems that closely match the difficulty of AIME while exceeding it in structural variation. Our work offers a cognitively grounded pathway toward scalable, high-quality math problem generation.Our code is publicly available at https://github.com/Icarus-1111/CogAtom.

In this paper, we reported our experiments with various strategies to improve code-mixed humour and sarcasm detection. Particularly, we tried three approaches: (i) native sample mixing, (ii) multi-task learning (MTL), and (iii) prompting and instruction finetuning very large multilingual language models (VMLMs). In native sample mixing, we added monolingual task samples to code-mixed training sets. In MTL learning, we relied on native and code-mixed samples of a semantically related task (hate detection in our case). Finally, in our third approach, we evaluated the efficacy of VMLMs via few-shot context prompting and instruction finetuning. Some interesting findings we got are (i) adding native samples improved humor (raising the F1-score up to 6.76%) and sarcasm (raising the F1-score up to 8.64%) detection, (ii) training MLMs in an MTL framework boosted performance for both humour (raising the F1-score up to 10.67%) and sarcasm (increment up to 12.35% in F1-score) detection, and (iii) prompting and instruction finetuning VMLMs couldn’t outperform the other approaches. Finally, our ablation studies and error analysis discovered the cases where our model is yet to improve. We provided our code for reproducibility.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) with external knowledge but incurs significant inference costs due to lengthy retrieved contexts. While context compression mitigates this issue, existing methods apply fixed compression rates—over-compressing simple queries or under-compressing complex ones. We propose Adaptive Context Compression for RAG (ACC-RAG), a framework that dynamically adjusts compression rates based on input complexity, optimizing inference efficiency without loss of accuracy. ACC-RAG combines a hierarchical compressor (for multi-granular embeddings) with a context selector to retain minimal sufficient information, akin to human skimming. Evaluated on Wikipedia and five QA datasets, ACC-RAG outperforms fixed-rate methods and unlocks >4× faster inference versus standard RAG while maintaining or improving accuracy.

In this paper, we propose a method to improve the reasoning capabilities of Visual Question Answering (VQA) systems by integrating Dense Passage Retrievers (DPRs) with Vision Language Models (VLMs). While recent works focus on the application of knowledge graphs and chain-of-thought reasoning, we recognize that the complexity of graph neural networks and end-to-end training remain significant challenges. To address these issues, we introduce Relevance Guided VQA (RG-VQA), a retriever-generator pipeline that uses DPRs to efficiently extract relevant information from structured knowledge bases. Our approach ensures scalability to large graphs without significant computational overhead. Experiments on the ScienceQA dataset show that RG-VQA achieves state-of-the-art performance, surpassing human accuracy and outperforming GPT-4 by more than . This demonstrates the effectiveness of RG-VQA in boosting the reasoning capabilities of VQA systems and its potential for practical applications.