论文检索

Qiong Wu, Tan Yue, Jianxin Liang, Zhen Li, Kai He, Shuai Zhao, Dongyan Zhao

The rapid progress of large language models (LLMs) has increased the demand for efficient and reliable evaluation of question answering (QA) systems. Existing evaluation methods either rely on rule-based matching with shallow semantic understanding or adopt LLM-as-a-Judge approaches that incur high cost and latency while offering limited error interpretability. Accordingly, we propose HiEval, a curriculum learning based hierarchical framework for QA task evaluation that supports both quick scoring and fine-grained error analysis. HiEval contains a quick scoring model (HiEval-QS) that predicts three-level correctness labels, and an error analysis model (HiEval-EA) that identifies incorrect responses into five error types. HiEval incorporates a class-balanced focal loss to handle label imbalance, experience replay to prevent forgetting, and contrastive unlikelihood optimization to improve error discrimination. We also construct two large-scale human-annotated evaluation datasets collected from 50 QA-related datasets, covering 8 task types and release two challenging benchmarks. Extensive experiments show that HiEval achieves state-of-the-art performance on both quick scoring and error analysis tasks, outperforming all baseline methods, including GPT-5, while being approximately 25\times faster.

Cehao Yang, Xueyuan Lin, Xiaojun Wu, Chengjin Xu, Xuhui Jiang, Honghao Liu, Hui Xiong, Jian Guo

A practical approach to activate long chain-of-thoughts reasoning ability in large language models is to perform supervised fine-tuning on instruction datasets synthesized by strong large reasoning models, offering a cost-effective alternative to reinforcement learning. However, large-scale instruction sets incur significant training overhead, while effective strategies for automatic data selection still remain unexplored. We propose Select2Reason, a novel and efficient instruction-tuning data selection framework for long-CoT reasoning. From the perspective of emergence of rethinking behaviors like self-correction and backtracking, we investigate metrics that may determine the quality of long-CoT instructions. Select2Reason leverages a difficulty-aware reward model to estimate the learning value of questions and jointly incorporates a reasoning trace length-based heuristic through a weighted scheme for ranking to prioritize high-utility examples. Empirical results on OpenR1-Math-220k demonstrate that fine-tuning LLM on only 10% of the data selected by our method achieves performance competitive with or superior to full-data tuning and open-source baseline across nine competition-level mathematical benchmarks and four broader reasoning tasks. Further experiments highlight the scalability in varying data size, efficiency during inference, and adaptability to other instruction pools of Select2Reason with minimal cost.

Hellina Hailu Nigatu, Bethelhem Yemane Mamo, Bontu Fufa Balcha, Debora Taye Tesfaye, Elbethel Daniel Zewdie, Ikram Behiru Nesiru, Jitu Ewnetu Hailu, Senait Mengesha Yayo

As low-resourced languages are increasingly incorporated into NLP research, there is an emphasis on collecting large-scale datasets. But in prioritizing quantity over quality, we risk 1) building language technologies that perform poorly for these languages and 2) producing harmful content that perpetuates societal biases. In this paper, we investigate the quality of Machine Translation (MT) datasets for three low-resourced languages–Afan Oromo, Amharic, and Tigrinya, with a focus on the gender representation in the datasets. Our findings demonstrate that while training data has a large representation of political and religious domain text, benchmark datasets are focused on news, health, and sports. We also found a large skew towards the male gender–in names of persons, the grammatical gender of verbs, and in stereotypical depictions in the datasets. Further, we found harmful and toxic depictions against women, which were more prominent for the language with the largest amount of data, underscoring that quantity does not guarantee quality. We hope that our work inspires further inquiry into the datasets collected for low-resourced languages and prompts early mitigation of harmful content. WARNING: This paper contains discussion of NSFW content that some may find disturbing.

Yong Wu, Weihang Pan, Ke Li, Chen Binhui, Ping Li, Binbin Lin

Large language models (LLMs) have shown remarkable reasoning capabilities, yet aligning such abilities to small language models (SLMs) remains a challenge due to distributional mismatches and limited model capacity. Existing reasoning datasets, typically designed for powerful LLMs, often lead to degraded performance when directly applied to weaker models. In this work, we introduce Dynamic Adaptation of Reasoning Trajectories (DART), a novel data adaptation framework that bridges the capability gap between expert reasoning trajectories and diverse SLMs. Instead of uniformly imitating expert steps, DART employs a selective imitation strategy guided by step-wise adaptability estimation via solution simulation. When expert steps surpass the student’s capacity—signaled by an Imitation Gap—the student autonomously explores alternative reasoning paths, constrained by outcome consistency. We validate DART across multiple reasoning benchmarks and model scales, demonstrating that it significantly improves generalization and data efficiency over static fine-tuning. Our method enhances supervision quality by aligning training signals with the student’s reasoning capabilities, offering a scalable solution for reasoning alignment in resource-constrained models.

Huacan Chai, Zijie Cao, Maolin Ran, Yingxuan Yang, Jianghao Lin, Xin Peng, Hairui Wang, Renjie Ding, Ziyu Wan, Muning Wen 等

Large language models (LLMs) have achieved impressive success in single-turn function calling, yet real-world applications such as travel planning or multi-stage data analysis typically unfold across multi-turn conversations. In these settings, LLMs must not only issue accurate function calls at each step but also maintain progress awareness, the ability to summarize past interactions and plan future actions to ensure coherent, long-horizon task execution. Existing approaches, however, either reduce multi-turn training to isolated single-turn samples, which neglects task-level planning, or employ end-to-end reinforcement learning (RL) that struggles with redundancy and lacks explicit integration of progress awareness. To overcome these limitations, we introduce Progra, a framework that explicitly incorporates progress awareness into LLM training for multi-turn function calling. Progra combines (i) a Progress Awareness Generation (PAG) pipeline, which automatically constructs datasets coupling conversation summaries with future task planning, and (ii) a Progress Awareness-Guided Reinforcement Learning (PAG-RL) algorithm, which integrates progress awareness into RL training to reduce contextual redundancy and improve alignment between local actions and global task completion. Empirical results on two public benchmarks demonstrate that Progra significantly outperforms existing methods, highlighting the effectiveness of progress awareness in enabling robust and efficient multi-turn function calling. Our code is available at https://github.com/FatCatCHC/Progra .

Yuanchen Shi, Longyin Zhang, Guodong Zhou, Fang Kong

Dangerous speech detection is a well-studied task, but existing approaches typically treat utterances in isolation, relying on binary labels that ignore who is speaking and in what mental state. We formulate a context-dependent variant of this task by grounding it in Theory-of-Mind (ToM). In cognitive science, ToM studies how humans attribute latent mental states-such as emotions, intentions, and actions-to others. We argue that such states are key signals for assessing the risk of an utterance. Building on this view, we construct ToM-DS, a 79K-instance dataset where each utterance is paired with structured speaker profiles, ToM states (emotion, intent, action), and topic hierarchies. During data construction, we first identify context-dependent sentences and generate diverse safe and dangerous scenarios surrounding them. High-quality annotations are obtained with state-of-the-art LLMs and a multi-stage cross-agent validation pipeline, yielding a comprehensive and reliable resource for context-dependent dangerous speech detection and fine-grained risk level classification. We further propose ToMGuard, a lightweight model with a dynamic ToM attention mechanism that adaptively weighs different mental-state cues. ToMGuard outperforms strong proprietary and open-source LLMs with significantly fewer parameters. Experimental results show that ToMGuard sets a new benchmark for context-dependent dangerous speech detection and risk level classification on ToM-DS.

Hoan My Tran, Damien Lolive, Aghilas Sini, Arnaud Delhay, Pierre-Francois Marteau, David Guennec

Detecting speech deepfakes is critical for protecting society against fraud, identity theft, and the misuse of modern speech synthesis technologies. Despite recent progress, existing countermeasures often exhibit limited generalization to unseen spoofing attacks, particularly in out-of-domain evaluation settings, even when achieving strong in-domain performance. Transformer architectures have become ubiquitous in anti-spoofing, serving both as feature extractors (e.g., wav2vec 2.0) and as classifiers. However, deep transformer stacks exhibit substantial representational redundancy across adjacent layers, with similarity increasing toward deeper layers. As a result, task-specific specialization is largely concentrated in the final layers, while shallow layers remain underutilized during fine-tuning. In this work, we analyze the layer-wise behavior of transformer-based classifiers for speech deepfake detection and propose a training strategy that explicitly aligns shallow and intermediate representations with those of the final transformer layer. By encouraging all layers to mimic the task-specialized representation learned at depth, the model more effectively exploits early-layer features while preserving discriminative capacity in deeper layers. This design improves robustness to unseen spoofing attacks and enhances out-of-domain generalization. Extensive experiments across multiple benchmark datasets demonstrate consistent performance gains over strong baselines.

Zimu Jia, Mingjie Xu, Andrew Estornell, Jiaheng Wei

The efficacy of Large Vision-Language Models (LVLMs) is critically dependent on the quality of their training data, requiring a precise balance between visual fidelity and instruction-following capability. Existing datasets, however, are plagued by inconsistent quality, and current data filtering methods rely on coarse-grained scores that lack the granularity to identify nuanced semantic flaws like logical fallacies or factual errors. This creates a fundamental bottleneck in developing more reliable models. To address this, we make three core contributions. First, we construct a large-scale, 300K-sample benchmark by systematically injecting diverse, subtle defects to provide a challenging testbed for data auditing. Second, we introduce a novel "Decomposition-then-Evaluation" paradigm that breaks model responses into constituent cognitive components: visual description, subjective inference, and factual claim, enabling targeted analysis. Third, we instantiate this paradigm via EVIAN (Explainable Visual Instruction-tuning Data AuditiNg), an automated framework that evaluates these components along the orthogonal axes of Image-Text Consistency, Logical Coherence, and Factual Accuracy. Our empirical findings challenge the prevailing scale-centric paradigm: a model fine-tuned on a compact, high-quality subset curated by EVIAN consistently surpassed models trained on orders-of-magnitude larger datasets. We also reveal that dividing complex auditing into verifiable subtasks enables robust curation, and that Logical Coherence is the most critical factor in data quality evaluation.

Kuo Tian, Pengfei Sun, Zhen Wu, Junran Ding, Xinyu Dai

The autonomous synthesis of deep research reports represents a critical frontier for Large Language Models (LLMs), demanding sophisticated information orchestration and non-linear narrative logic. Current approaches rely on rigid predefined linear workflows, which cause error accumulation, preclude global restructuring from subsequent insights, and ultimately limit in-depth multimodal fusion and report quality. We propose CogGen, a Cognitively inspired recursive framework for deep research report Generation. Leveraging a Hierarchical Recursive Architecture to simulate cognitive writing, CogGen enables flexible planning and global restructuring. To extend this recursivity to multimodal content, we introduce Abstract Visual Representation (AVR): a concise intent-driven language that iteratively refines visual-text layouts without pixel-level regeneration overhead. We further present CLEF, a Cognitive Load Evaluation Framework, and curate a new benchmark from Our World in Data (OWID). Extensive experiments show CogGen achieves state-of-the-art results among open-source systems, generating reports comparable to professional analysts’ outputs and surpassing Gemini Deep Research. Our code and dataset will be publicly available upon publication.

Tang Guowei, Tianwen Qian, Huanran Zheng, Wang Yifei, Xiaoling Wang

Real-time, continuous understanding of visual signals is essential for real-world interactive AI applications, and poses a fundamental system-level challenge. Existing research on streaming video understanding, however, typically focuses on isolated aspects such as question-answering accuracy under limited visual context or improvements in encoding efficiency, while largely overlooking practical deployability under realistic resource constraints. To bridge this gap, we introduce StreamingEval, a unified evaluation framework for assessing the streaming video understanding capabilities of Video-LLMs under realistic constraints. StreamingEval benchmarks both mainstream offline models and recent online video models under a standardized protocol, explicitly characterizing the trade-off between efficiency, storage and accuracy. Specifically, we adopt a fixed-capacity memory bank to normalize accessible historical visual context, and jointly evaluate visual encoding efficiency, text decoding latency, and task performance to quantify overall system deployability. Extensive experiments across multiple datasets reveal substantial gaps between current Video-LLMs and the requirements of realistic streaming applications, providing a systematic basis for future research in this direction. Codes will be released upon acceptance.

Maosen Zhang, Jianshuo Dong, Lu Boting, Li Wenyue, Xiaoping Zhang, Tianwei Zhang, Han Qiu

Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to leverage external knowledge, but also exposes valuable RAG databases to leakage attacks. As RAG systems grow more complex and LLMs exhibit stronger instruction-following capabilities, existing studies fall short of systematically assessing RAG leakage risks. We present LeakDojo, a configurable framework for controlled evaluation of RAG leakage. Using LeakDojo, we benchmark six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems. Our study reveals that (1) query generation and adversarial instructions contribute independently to leakage, with overall leakage well approximated by their product; (2) stronger instruction-following capability correlates with higher leakage risk; and (3) improvements in RAG faithfulness can introduce increased leakage risk. These findings provide actionable insights for understanding and mitigating RAG leakage in practice. Our codebase is available at https://github.com/yeasen-z/LeakDojo.

Ivanhoé Botcazou, Tassadit Amghar, Sylvain Lamprier, Frédéric Saubion

Modern neural language models achieve high accuracy in text generation, yet precise control over generation length remains underdeveloped. In this paper, we first investigate a recent length control method based on Reverse Positional Embeddings (RPE) and show its limits when control is requested beyond the training distribution. In particular, using a discrete countdown signal tied to the absolute remaining token count leads to instability. To provide robust length control, we introduce Progress Ratio Embeddings (PRE), as continuous embeddings tied to a trigonometric impatience signal. PRE integrates seamlessly into standard Transformer architectures, providing stable length fidelity without degrading text accuracy under standard evaluation metrics. We further show that PRE generalizes well to unseen target lengths. Experiments on two widely used news-summarization benchmarks and a popular question generation dataset validate these findings.

Jun Xue, Zhuolin Yi, Yihuan Huang, Yanzhen Ren, Yujie Chen, Cunhang Fan, Zicheng Su, Yongcheng Zhang, Bo Cai

With the rapid advancement of speech generation technologies, the threat posed by speech deepfakes in real-time communication (RTC) scenarios has intensified. However, existing detection studies mainly focus on offline simulations and struggle to cope with the complex distortions introduced during RTC transmission, including unknown speech enhancement processes (e.g., noise suppression) and codec compression. To address this challenge, we present the first large-scale speech deepfake dataset tailored for RTC scenarios, termed RTCFake, totaling approximately 600 hours. The dataset is constructed by transmitting speech through multiple mainstream social media and conferencing platforms (e.g., Zoom), enabling precise pairing between offline and online speech. In addition, we propose a phoneme-guided consistency learning (PCL) strategy that enforces models to learn platform-invariant semantic structural representations. In this paper, the RTCFake dataset is divided into training, development, and evaluation sets. The evaluation set further includes both unseen RTC platforms and unseen complex noise conditions, thereby providing a more realistic and challenging evaluation benchmark for speech deepfake detection. Furthermore, the proposed PCL strategy achieves significant improvements in both cross-platform generalization and noise robustness, offering an effective and generalizable modeling paradigm.

Zihao Yi, Qingxuan Jiang, Ruotian Ma, Xingyu Chen, Qu Yang, Mengru Wang, Fanghua Ye, Ying Shen, Zhaopeng Tu, Xiaolong Li 等

Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ”Deceitful” and ”Manipulative”, often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

Jiajun Wu, Jian Yang, Wei Zhang, Linzheng Chai, Yuchi Ma, Ensheng Shi, Yuqing Ma, Zhoujun Li, Xianglong Liu

Large language models (LLMs) have demonstrated remarkable capabilities in code generation tasks. However, their effectiveness heavily relies on supervised training with extensive labeled (e.g., question-answering pairs) or unlabeled datasets (e.g., code snippets), which are often expensive and difficult to obtain at scale. To address this limitation, this paper introduces a method IPC, an unsupervised framework that leverages Internal Probing of LLMs for Code generation without any external corpus, even unlabeled code snippets. We introduce the problem space probing, test understanding probing, solution space probing, and knowledge consolidation and reinforcement to probe the internal knowledge and confidence patterns existing in LLMs. Further, IPC identifies reliable code candidates through self-consistency mechanisms and representation-based quality estimation to train UCoder (coder with unsupervised learning). We validate the proposed approach across multiple code benchmarks, demonstrating that unsupervised methods can achieve competitive performance compared to supervised approaches while significantly reducing the dependency on labeled data and computational resources. Analytic experiments reveal that internal model states contain rich signals about code quality and correctness, and that properly harnessing these signals enables effective unsupervised learning for code generation tasks, opening new directions for training code LLMs in resource-constrained scenarios.

Zhili Pu, Lantian Zhang, Hao Duan, Zhixing Zhang, Keyun Zhu, Yongqi Fan, Ruihui Hou, Tong Ruan, Yun Tang

Large language models (LLMs) offer a new paradigm for molecular property prediction (MPP), yet a semantic gap between natural language and molecular representations limits LLMs’ ability to capture structure–activity relationships (SAR). Recent approaches have explored injecting structure-level information into LLMs, primarily modeling associations based on statistical regularities. However, these methods are prone to misinterpreting coincidental associations as general principles, imposing a bottleneck on predictive performance. To tackle the challenges above, we propose MCLE-Mol, an ML–LLM–Rule collaborative framework for MPP. It bridges the semantic gap by injecting ML-derived substructure attribution values into LLMs, utilizing Context-Calibrated Substructure Attribution Rules (CCSAR) to calibrate these attributions under specific chemical contexts to mitigate such misinterpretation. In addition, MCLE-Mol introduces a low-cost continual evolution strategy that updates CCSAR with frozen model parameters to adapt to dynamic chemical spaces. Experiments on multiple benchmark datasets demonstrate that MCLE-Mol outperforms all baselines, successfully resolving the trade-off between predictive performance and interpretability.

Jiaming Zhou, Yujie Guo, Shiwan Zhao, Yao Lu, Jianye Wang, Haoqin Sun, Hui Wang, Yong Qin

Automatic speech recognition (ASR) for children remains challenging due to developmental variability and the scarcity of high-quality corpora, especially for Mandarin and its dialects. In this paper, we present ChildTalk, a large-scale Chinese child speech corpus designed to address this gap. It contains 112.5 hours of speech from 498 children (aged 2–8) and 500 caregivers, recorded as natural child–caregiver conversations. Unlike prior Mandarin child ASR corpora that mainly release isolated utterances, ChildTalk provides full-length dialogues with complete transcriptions, preserving turn-taking and discourse context. To our knowledge, it is the first publicly available Mandarin child speech corpus with full-length dialogues and systematic coverage of standard Mandarin, eight Mandarin dialect subgroups, and two additional dialects (Southern Min and Jin). We benchmark end-to-end models trained from scratch, large pre-trained ASR models fine-tuned on ChildTalk, omni-modal LLMs in a zero-shot setting, and commercial speech transcription APIs. Fine-tuning on ChildTalk consistently improves both in-domain and cross-domain performance. These results indicate that ChildTalk provides a challenging, broad-coverage testbed for Chinese child ASR, dialect robustness, and dialogue-level modeling. The dataset will be made freely available for all academic purposes.

Basel Mousi, Fahim Dalvi, Shammur Absar Chowdhury, Firoj Alam, Nadir Durrani

Vision–language models (VLMs) can achieve high accuracy while still accepting **culturally plausible but visually incorrect** interpretations. Existing hallucination benchmarks rarely test this failure mode, particularly outside Western contexts and English. We introduce **M^2CQA**, a culturally grounded multimodal benchmark built from images spanning 17 MENA countries, paired with contrastive true and counterfactual statements in English, Arabic, and its dialects. To isolate hallucination beyond raw accuracy, we propose the **CounterFactual Hallucination Rate (CFHR)**, which measures counterfactual acceptance conditioned on correctly answering the true statement. Evaluating state-of-the-art VLMs under multiple prompting strategies, we find that CFHR rises sharply in Arabic, especially in dialects, even when true-statement accuracy remains high.Moreover, reasoning-first prompting consistently increases counterfactual hallucination, while answering before justifying improves robustness. We make the dataset publicly available for the community (https://huggingface.co/datasets/QCRI/M2CQA)).

Arnav Attri, Anuj Attri

Opinion summarization systems aggregate customer sentiments without capturing the emotional factors that drive purchasing decisions, resulting in shallow summaries that overlook the affective dimensions shaping customer experiences and fail to explain why customers feel the way they do. This gap exists because prior research has neglected the interplay between expressed opinions and their underlying emotional contexts. To bridge this gap, we introduce Emotion-Aware Opinion Summarization (EAOS), a framework leveraging Large Language Models (LLMs) to integrate emotional dimensions into opinion summaries, moving beyond conventional sentiment polarity. To support this task, we develop a large-scale (40K product–summary pairs) training dataset, an evaluation benchmark, a compact 1B-parameter model that matches 70B-scale performance via knowledge distillation, and methods for generating and evaluating emotion-aware summaries. A user study shows that 82% of readers prefer our emotion-aware summaries (p < .001), confirming that adding emotion helps in making purchase decisions.

Zhipeng Qian, Zihan Liang, Yufei Ma, Ben Chen, Huangyu Dai, Yiwei Ma, Jiayi Ji, Chenyi Lei, Han Li, Xiaoshuai Sun

Composed Image Retrieval (CIR) enables users to search for target images using both a reference image and manipulation text, offering substantial advantages over single-modality retrieval systems. However, existing CIR methods suffer from representation space fragmentation: queries and targets comprise heterogeneous modalities and are processed by distinct encoders, forcing models to bridge misaligned representation spaces only through post-hoc alignment, which fundamentally limits retrieval performance. As evidenced by t-SNE visualization in Fig.(a), this architectural asymmetry manifests as three distinct, well-separated clusters in the feature space, directly demonstrating how heterogeneous modalities create fundamentally misaligned representation spaces from initialization. In this work, we propose CSMCIR, a unified representation framework that achieves efficient query-target alignment through three synergistic components. First, we introduce a Multi-level Chain-of-Thought (MCoT) prompting strategy that guides Multimodal Large Language Models to generate discriminative, semantically compatible captions for target images, establishing modal symmetry. Building upon this, we design a symmetric dual-tower architecture where both query and target sides utilize the identical shared-parameter Q-Former for cross-modal encoding, ensuring consistent feature representations and further reducing the alignment gap. Finally, this architectural symmetry enables an entropy-based, temporally dynamic Memory Bank strategy that provides high-quality negative samples while maintaining consistency with the evolving model state. Extensive experiments on four benchmark datasets demonstrate that our CSMCIR achieves state-of-the-art performance with superior training efficiency. Comprehensive ablation studies further validate the effectiveness of each proposed component.