EMNLP 2025 Papers

Legal mathematical reasoning is essential for applying large language models (LLMs) in high-stakes legal contexts, where outputs must be both mathematically accurate and procedurally compliant. However, existing legal LLMs lack structured numerical reasoning, and open-domain models, though capable of calculations, often overlook mandatory legal steps. To address this, we present LexNum, the first Chinese legal mathematical reasoning benchmark, covering three representative scenarios where each instance reflects legally grounded procedural flows. We further propose LexPam, a two-stage reinforcement learning framework for efficient legal reasoning training. Leveraging curriculum learning, we use a stronger teacher model to partition data into basic and challenging subsets. A lightweight 1.5B student model is then fine-tuned with Group Relative Policy Optimization, which avoids costly value networks and enables stable training from sparse, end-of-sequence rewards. The first stage improves accuracy and format; the second introduces a novel reward to guide procedural alignment via task-specific legal elements. Experiments show that existing models perform poorly on LexNum, while LexPam enhances both mathematical accuracy and legal coherence, and generalizes effectively across tasks and domains.

Prompt privacy is crucial, especially when using online large language models (LLMs), due to the sensitive information often contained within prompts. While LLMs can enhance prompt privacy through text rewriting, existing methods primarily focus on document-level rewriting, neglecting the rich, multi-granular representations of text. This limitation restricts LLM utilization to specific tasks, overlooking their generalization and in-context learning capabilities, thus hindering practical application. To address this gap, we introduce DP-GTR, a novel three-stage framework that leverages local differential privacy (DP) and the composition theorem via group text rewriting. DP-GTR is the first framework to integrate both document-level and word-level information while exploiting in-context learning to simultaneously improve privacy and utility, effectively bridging local and global DP mechanisms at the individual data point level. Experiments on CommonSense QA and DocVQA demonstrate that DP-GTR outperforms existing approaches, achieving a superior privacy-utility trade-off. Furthermore, our framework is compatible with existing rewriting techniques, serving as a plug-in to enhance privacy protection. Our code is publicly available at anonymous.4open.science for reproducibility.

Visual Question Answering (VQA) requires a vision-language model to reason over both visual and textual inputs to answer questions about images. In this work, we investigate whether incorporating explicit semantic information, in the form of Abstract Meaning Representation (AMR) graphs, can enhance model performance—particularly in low-resource settings where training data is limited. We augment two vision-language models, LXMERT and BLIP-2, with sentence- and document-level AMRs and evaluate their performance under both full and reduced training data conditions. Our findings show that in well-resourced settings, models (in particular the smaller LXMERT) are negatively impacted by incorporating AMR without specialized training. However, in low-resource settings, AMR proves beneficial: LXMERT achieves up to a 13.1% relative gain using sentence-level AMRs. These results suggest that while addition of AMR can lower the performance in some settings, in a low-resource setting AMR can serve as a useful semantic prior, especially for lower-capacity models trained on limited data.

We introduce SQLSpace, a human-interpretable, generalizable, compact representation for text-to-SQL examples derived with minimal human intervention. We demonstrate the utility of these representations in evaluation with three use cases: (i) closely comparing and contrasting the composition of popular NL2SQL benchmarks to identify unique dimensions of examples they evaluate, (ii) understanding model performance at a granular level beyond overall accuracy scores, and (iii) improving model performance through targeted query rewriting based on learned correctness estimation. We show that SQLSpace enables analysis that would be difficult with raw examples alone: it reveals compositional differences between benchmarks, exposes performance patterns obscured by accuracy alone, and supports modeling of query success.

The collaborative paradigm of large and small language models (LMs) effectively balances performance and cost, yet its pivotal challenge lies in precisely pinpointing the moment of invocation when hallucinations arise in small LMs. Previous optimization efforts primarily focused on post-processing techniques, which were separate from the reasoning process of LMs, resulting in high computational costs and limited effectiveness. In this paper, we propose a practical invocation evaluation metric called AttenHScore, which calculates the accumulation and propagation of hallucinations during the generation process of small LMs, continuously amplifying potential reasoning errors. By dynamically adjusting the detection threshold, we achieve more accurate real-time invocation of large LMs. Additionally, considering the limited reasoning capacity of small LMs, we leverage uncertainty-aware knowledge reorganization to assist them better capture critical information from different text chunks. Extensive experiments reveal that our AttenHScore outperforms most baselines in enhancing real-time hallucination detection capabilities across multiple QA datasets, especially when addressing complex queries. Moreover, our strategies eliminate the need for additional model training and display flexibility in adapting to various transformer-based LMs. Our code is available at https://github.com/Robot2050/AttenHScore.

Reward models are essential for aligning large language models (LLMs) with human preferences. However, most open-source multilingual reward models are primarily trained on preference datasets in high-resource languages, resulting in unreliable reward signals for low-resource Indic languages. Collecting large-scale, high-quality preference data for these languages is prohibitively expensive, making preference-based training approaches impractical. To address this challenge, we propose RELIC, a novel in-context learning framework for reward modeling in low-resource Indic languages. RELIC trains a retriever with a pairwise ranking objective to select in-context examples from auxiliary high-resource languages that most effectively highlight the distinction between preferred and less-preferred responses. Extensive experiments on three preference datasets—PKU-SafeRLHF, WebGPT, and HH-RLHF—using state-of-the-art open-source reward models demonstrate that RELIC significantly improves reward model accuracy for low-resource Indic languages, consistently outperforming existing example selection methods. For example, on Bodo—a low-resource Indic language—using a LLaMA-3.2-3B reward model, RELIC achieves a 12.81% and 10.13% improvement in accuracy over zero-shot prompting and state-of-the-art example selection method, respectively

Critique ability, a meta-cognitive capability of humans, presents significant challenges for LLMs to improve. While utilizing human annotation can enhance critique ability effectively, most recent works primarily rely on supervised fine-tuning (SFT) using critiques generated by a single LLM like GPT-4, which is more scalable and cost-effective.However, such model-generated critiques often suffer from inherent flaws due to the complexity of critique. Consequently, fine-tuning LLMs on these flawed critiques not only limits performance but also propagates errors into the learned model.To address this issue, we propose MultiCritique, a unified framework that leverages multi-agent feedback to improve critique ability in both the supervised fine-tuning (SFT) and reinforcement learning (RL) stages.In the SFT stage, MultiCritique aggregates high-quality multi-agent critiques through a fine-grained meta-critique mechanism. In the RL stage, preference critiques are constructed and refined by validating their contributions to revisions, thereby enhancing robustness of RL in improving critique ability.Based on MultiCritique, we construct SFT and RL datasets. Extensive experimental results on two benchmarks highlight the key benefits of our dataset, including superior quality, enhanced data efficiency, strong generalization on unseen tasks, and improvements in the general capability of LLMs.Notably, our fine-tuned 7B model significantly surpasses advanced 7B-13B models, approaching advanced 70B LLMs and GPT-4.Resources have been publicly available.

Automatic language identification is frequently framed as a multi-class classification problem. However, when creating digital corpora for less commonly written languages, it may be more appropriate to consider it a data mining problem. For these varieties, one knows ahead of time that the vast majority of documents are of little interest. By minimizing resources spent on classifying such documents, we can create corpora covering previously overlooked languages faster than existing pipelines. To demonstrate the effectiveness of the targeted mining perspective, we introduce a new pipeline that can filter a single snapshot in two hours. We also provide web corpora for several French-based Creoles.

Large Language Models (LLMs) often exhibit social biases inherited from their training data. While existing benchmarks evaluate bias by term-based mode through direct term associations between demographic terms and bias terms, LLMs have become increasingly adept at avoiding biased responses, leading to seemingly low levels of bias. However, biases persist in subtler, contextually hidden forms that traditional benchmarks fail to capture. We introduce the Description-based Bias Benchmark (DBB), a novel dataset designed to assess bias at the semantic level that bias concepts are hidden within naturalistic, subtly framed contexts in real-world scenarios rather than superficial terms. We analyze six state-of-the-art LLMs, revealing that while models reduce bias in response at the term level, they continue to reinforce biases in nuanced settings. Data, code, and results are available at https://github.com/JP-25/Description-based-Bias-Benchmark.

This paper establishes a formal information-theoretic framework for image captioning, conceptualizing captions as compressed linguistic representations that selectively encode semantic units in images. Our framework posits that good image captions should balance three key aspects: informationally sufficient, minimally redundant, and readily comprehensible by humans. By formulating these aspects as quantitative measures with adjustable weights, our framework provides a flexible foundation for analyzing and optimizing image captioning systems across diverse task requirements. To demonstrate its applicability, we introduce the Pyramid of Captions (PoCa) method, which generates enriched captions by integrating local and global visual information. We present both theoretical proof that PoCa improves caption quality under certain assumptions, and empirical validation of its effectiveness across various image captioning models and datasets.

Large Language Models (LLMs) are increasingly aligned with human preferences through Reinforcement Learning from Human Feedback (RLHF). Among RLHF methods, Group Relative Policy Optimization (GRPO) has gained attention for its simplicity and strong performance, notably eliminating the need for a learned value function. However, GRPO implicitly assumes a balanced domain distribution and uniform semantic alignment across groups—assumptions that rarely hold in real-world datasets. When applied to multi-domain, imbalanced data, GRPO disproportionately optimizes for dominant domains, neglecting underrepresented ones and resulting in poor generalization and fairness. We propose Domain-Informed Self-Consistency Policy Optimization (DISCO), a principled extension to GRPO that addresses inter-group imbalance with two key innovations. Domain-aware reward scaling counteracts frequency bias by reweighting optimization based on domain prevalence. Difficulty-aware reward scaling leverages prompt-level self-consistency to identify and prioritize uncertain prompts that offer greater learning value. Together, these strategies promote more equitable and effective policy learning across domains. Extensive experiments across multiple LLMs and skewed training distributions show that DISCO improves generalization, outperforms existing GRPO variants by 5% on Qwen3 models, and sets new state-of-the-art results on multi-domain alignment benchmarks.

Rapid developments of large language models have revolutionized many NLP tasks for English data. Unfortunately, the models and their evaluations for low-resource languages are being overlooked, especially for languages in South Asia. Although there are more than 650 languages in South Asia, many of them either have very limited computational resources or are missing from existing language models.Thus, a concrete question to be answered is: Can we assess the current stage and challenges to inform our NLP community and facilitate model developments for South Asian languages? In this survey, we have comprehensively examined current efforts and challenges of NLP models for South Asian languages by retrieving studies since 2020, with a focus on transformer-based models, such as BERT, T5, & GPT. We present advances and gaps across 3 essential aspects: data, models, & tasks, such as available data sources, fine-tuning strategies, & domain applications. Our findings highlight substantial issues, including missing data in critical domains (e.g., health), code-mixing, and lack of standardized evaluation benchmarks. Our survey aims to raise awareness within the NLP community for more targeted data curation, unify benchmarks tailored to cultural and linguistic nuances of South Asia, and encourage an equitable representation of South Asian languages. The complete list of resources is available at: https://github.com/trust-nlp/LM4SouthAsia-Survey.

Parameter-efficient fine-tuning (PEFT) is essential for adapting large language models (LLMs), with low rank adaptation (LoRA) being the most popular approach. However, LoRA suffers from slow convergence, and some recent LoRA variants, such as PiSSA, primarily rely on Singular Value Decomposition (SVD) for initialization, leading to expensive computation. To mitigate these problems, we resort to Nyström method, which follows a three-matrix manipulation. Therefore, we first introduce StructuredLoRA (SLoRA), investigating to introduce a small intermediate matrix between the low-rank matrices (A) and (B). Secondly, we propose NyströmLoRA (NLoRA), which leverages Nyström-based initialization for SLoRA to improve its effectiveness and efficiency. Finally, we propose IntermediateTune (IntTune) to explore fine-tuning exclusively the intermediate matrix of NLoRA to furthermore boost LLMs’ efficiency. We evaluate our methods on 5 natural language generation (NLG) tasks and 8 natural language understanding (NLU) tasks. On GSM8K, SLoRA and NLoRA achieve accuracies of 56.48% and 57.70%, surpassing LoRA by 33.52% and 36.41% with only 3.67M additional trainable parameters. IntTune boosts average NLG performance over LoRA by 7.45% while using only 1.25% of its parameters. These results demonstrate the efficiency and effectiveness of our approach in enhancing model performance with minimal parameter overhead.

Parameter generation has emerged as a novel paradigm for neural network development, offering an alternative to traditional neural network training by synthesizing high-quality model weights directly. In the context of Low-Rank Adaptation (LoRA) for evolving (i.e, constantly updated) large language models (LLMs), this approach promises efficient adaptation without costly retraining. However, existing methods face critical limitations in simultaneously achieving scalability and controllability. In this paper, we introduce \texttt{ORAL}, a novel \textit{conditional recurrent diffusion} framework that addresses these challenges. \texttt{ORAL} incorporates a novel conditioning mechanism that integrates model architecture and textual task specifications, enabling the generation of task-specific LoRA parameters that can seamlessly transfer across evolving foundation models. Our approach successfully scales to billions-of-parameter LLMs and maintains controllability. Through extensive experiments across seven language tasks, four vision tasks, and three multimodal tasks using five pre-trained LLMs, we demonstrate that \texttt{ORAL} generates high-quality LoRA parameters that achieve comparable or superior performance to vanilla trained counterparts.

Emotional Support Conversation (ESC) systems aim to alleviate user distress. However, current Chain-of-Thought based ESC methods often employ rigid, text-only reasoning, limiting adaptability in dynamic, multimodal interactions and introducing reasoning noise that degrades support quality. To address this, we introduce “Flexible Thinking” for multimodal ESC, enabling models to adaptively select contextually relevant thinking aspects: Visual Scene, Emotion, Situation, and Response Strategy. We first construct training data by manually curating flexible thinking demonstrations on the MESC dataset, then using a Multimodal Large Language Model to synthesize these processes for the full training set. Then, we propose FIRES, a framework integrating Supervised Fine-Tuning (SFT) for initial learning with Reinforcement Learning for refinement. This two-stage approach helps FIRES transcend SFT’s generalization limits and, crucially, directly links thinking processes to response quality via tailored rewards, moving beyond imitating potentially imperfect synthetic data. Experiments on MESC and EMOTyDA datasets demonstrate FIRES’s effectiveness and generalizability in fostering higher-quality emotional support responses through adaptive reasoning.

Foundation models (FMs) are increasingly applied to bridge language and action in embodied agents, yet the operational characteristics of different integration strategies remain under-explored—especially for complex instruction following and versatile action generation in changing environments. We investigate three paradigms for robotic systems: end-to-end vision-language-action models (VLAs) that implicitly unify perception and planning, and modular pipelines using either vision-language models (VLMs) or multimodal large language models (MLLMs). Two case studies frame the comparison: instruction grounding, which probs fine-grained language understanding and cross-modal disambiguation; and object manipulation, which targets skill transfer via VLA finetuning. Our experiments reveal trade-offs in system scale, generalization and data efficiency. These findings indicate design lessons for language-driven physical agents and point to challenges and opportunities for FM-powered robotics in real-world conditions.

As large language models (LLMs) continue to evolve, efficient evaluation metrics are vital for assessing their ability to compress information and reduce redundancy. While traditional metrics like Matrix Entropy offer valuable insights, they are computationally intensive for large-scale models due to their O(n^3) time complexity with Singular Value Decomposition (SVD). To mitigate this issue, we introduce the Matrix Nuclear-Norm, which not only serves as a metric to quantify the data compression proficiency of LLM but also provides a convex approximation of matrix rank to capture both predictive discriminability and diversity. By employing the L_{1,2}\text{-norm} to further approximate the nuclear norm, we can effectively assess the model’s information compression capabilities. This approach reduces the time complexity to O(n^2) and eliminates the need for SVD computation. Consequently, the Matrix Nuclear-Norm achieves speeds 8 to 24 times faster than Matrix Entropy for the CEREBRAS-GPT model as sizes increase from 111M to 6.7B. This performance gap becomes more pronounced with larger models, as validated in tests with other models like Pythia. Additionally, evaluations on benchmarks and model responses confirm that our proposed Matrix Nuclear-Norm is a reliable, scalable, and efficient tool for assessing LLMs’ performance, striking a balance between accuracy and computational efficiency.

Large Language Models (LLMs) often exhibit hallucinations, generating factually incorrect or semantically irrelevant content in response to prompts. Chain-of-Thought (CoT) prompting can mitigate hallucinations by encouraging step-by-step reasoning, but its impact on hallucination detection remains underexplored. To bridge this gap, we conduct a systematic empirical evaluation. We begin with a pilot experiment, revealing that CoT reasoning significantly affects the LLM’s internal states and token probability distributions. Building on this, we evaluate the impact of various CoT prompting methods on mainstream hallucination detection methods across both instruction-tuned and reasoning-oriented LLMs. Specifically, we examine three key dimensions: changes in hallucination score distributions, variations in detection accuracy, and shifts in detection confidence. Our findings show that while CoT prompting helps reduce hallucination frequency, it also tends to obscure critical signals used for detection, impairing the effectiveness of various detection methods. Our study highlights an overlooked trade-off in the use of reasoning. Code is publicly available at: https://github.com/ECNU-Text-Computing/cot-hallu-detect .

Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning of large language models by decomposing weight updates into low-rank matrices, significantly reducing storage and computational overhead. While effective, standard LoRA lacks mechanisms for uncertainty quantification, leading to overconfident and poorly calibrated models. Bayesian variants of LoRA address this limitation, but at the cost of a significantly increased number of trainable parameters, partially offsetting the original efficiency gains. Additionally, these models are harder to train and may suffer from unstable convergence.In this work, we propose a novel parameter-efficient Bayesian LoRA via subspace inference, demonstrating that effective uncertainty quantification can be achieved in very low-dimensional parameter spaces. The proposed method achieves strong performance with improved calibration and generalization while maintaining computational efficiency. Our empirical findings show that, with the appropriate projection of the weight space: (1) uncertainty can be effectively modeled in a low-dimensional space, and (2) weight covariances exhibit low ranks.