Text_Generation
-
Context Tokens are Anchors: Understanding the Repetition Curse in dMLLMs from an Information Flow Perspective
2026-01-28 11:54:42Qiyan Zhao, Xiaofeng Zhang, Shuochen Chang, Qianyu Chen, Xiaosong Yuan, Xuhang Chen, Luoqi Liu, Jiajun Zhang, Xu-Yao Zhang, Da-Han WangAbstract
Recent diffusion-based Multimodal Large Language Models (dMLLMs) suffer from high inference latency and therefore rely on caching techniques to accelerate decoding. However, the application of cache mechanisms often introduces undesirable repetitive text generation, a phenomenon we term the \textbf{Repeat Curse}. To better investigate underlying mechanism behind this issue, we analyze repetition generation through the lens of information flow. Our work reveals three key findings: (1) context tokens aggregate semantic information as anchors and guide the final predictions; (2) as information propagates across layers, the entropy of context tokens converges in deeper layers, reflecting the model's growing prediction certainty; (3) Repetition is typically linked to disruptions in the information flow of context tokens and to the inability of their entropy to converge in deeper layers. Based on these insights, we present \textbf{CoTA}, a plug-and-play method for mitigating repetition. CoTA enhances the attention of context tokens to preserve intrinsic information flow patterns, while introducing a penalty term to the confidence score during decoding to avoid outputs driven by uncertain context tokens. With extensive experiments, CoTA demonstrates significant effectiveness in alleviating repetition and achieves consistent performance improvements on general tasks. Code is available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2601.20520
PDF
-
ProToken: Token-Level Attribution for Federated Large Language Models
2026-01-27 14:53:12Waris Gill, Ahmad Humayun, Ali Anwar, Muhammad Ali GulzarAbstract
Federated Learning (FL) enables collaborative training of Large Language Models (LLMs) across distributed data sources while preserving privacy. However, when federated LLMs are deployed in critical applications, it remains unclear which client(s) contributed to specific generated responses, hindering debugging, malicious client identification, fair reward allocation, and trust verification. We present ProToken, a novel Provenance methodology for Token-level attribution in federated LLMs that addresses client attribution during autoregressive text generation while maintaining FL privacy constraints. ProToken leverages two key insights to enable provenance at each token: (1) transformer architectures concentrate task-specific signals in later blocks, enabling strategic layer selection for computational tractability, and (2) gradient-based relevance weighting filters out irrelevant neural activations, focusing attribution on neurons that directly influence token generation. We evaluate ProToken across 16 configurations spanning four LLM architectures (Gemma, Llama, Qwen, SmolLM) and four domains (medical, financial, mathematical, coding). ProToken achieves 98% average attribution accuracy in correctly localizing responsible client(s), and maintains high accuracy when the number of clients are scaled, validating its practical viability for real-world deployment settings.
Abstract (translated)
URL
https://arxiv.org/abs/2601.19672
PDF
-
One Token Is Enough: Improving Diffusion Language Models with a Sink Token
2026-01-27 14:32:36Zihou Zhang, Zheyong Xie, Li Zhong, Haifeng Liu, Shaosheng CaoAbstract
Diffusion Language Models (DLMs) have emerged as a compelling alternative to autoregressive approaches, enabling parallel text generation with competitive performance. Despite these advantages, there is a critical instability in DLMs: the moving sink phenomenon. Our analysis indicates that sink tokens exhibit low-norm representations in the Transformer's value space, and that the moving sink phenomenon serves as a protective mechanism in DLMs to prevent excessive information mixing. However, their unpredictable positions across diffusion steps undermine inference robustness. To resolve this, we propose a simple but effective extra sink token implemented via a modified attention mask. Specifically, we introduce a special token constrained to attend solely to itself, while remaining globally visible to all other tokens. Experimental results demonstrate that introducing a single extra token stabilizes attention sinks, substantially improving model performance. Crucially, further analysis confirms that the effectiveness of this token is independent of its position and characterized by negligible semantic content, validating its role as a robust and dedicated structural sink.
Abstract (translated)
URL
https://arxiv.org/abs/2601.19657
PDF
-
Cross-Examination Framework: A Task-Agnostic Diagnostic for Information Fidelity in Text-to-Text Generation
2026-01-27 08:30:13Tathagata Raha, Clement Christophe, Nada Saadi, Hamza A Javed, Marco AF Pimentel, Ronnie Rajan, Praveenkumar KanithiAbstract
Traditional metrics like BLEU and BERTScore fail to capture semantic fidelity in generative text-to-text tasks. We adapt the Cross-Examination Framework (CEF) for a reference-free, multi-dimensional evaluation by treating the source and candidate as independent knowledge bases. CEF generates verifiable questions from each text and performs a cross-examination to derive three interpretable scores: Coverage, Conformity, and Consistency. Validated across translation, summarization and clinical note-generation, our framework identifies critical errors, such as content omissions and factual contradictions, missed by standard metrics. A key contribution is a systematic robustness analysis to select a stable judge model. Crucially, the strong correlation between our reference-free and with-reference modes validates CEF's reliability without gold references. Furthermore, human expert validation demonstrates that CEF mismatching questions align with meaning-altering semantic errors higher than with non-semantic errors, particularly excelling at identifying entity-based and relational distortions.
Abstract (translated)
URL
https://arxiv.org/abs/2601.19350
PDF
-
dLLM-ASR: A Faster Diffusion LLM-based Framework for Speech Recognition
2026-01-25 16:39:52Wenjie Tian, Bingshen Mu, Guobin Ma, Xuelong Geng, Zhixian Zhao, Lei XieAbstract
Automatic speech recognition (ASR) systems based on large language models (LLMs) achieve superior performance by leveraging pretrained LLMs as decoders, but their token-by-token generation mechanism leads to inference latency that grows linearly with sequence length. Meanwhile, discrete diffusion large language models (dLLMs) offer a promising alternative, enabling high-quality parallel sequence generation with pretrained decoders. However, directly applying native text-oriented dLLMs to ASR leads to a fundamental mismatch between open-ended text generation and the acoustically conditioned transcription paradigm required by ASR. As a result, it introduces unnecessary difficulty and computational redundancy, such as denoising from pure noise, inflexible generation lengths, and fixed denoising steps. We propose dLLM-ASR, an efficient dLLM-based ASR framework that formulates dLLM's decoding as a prior-guided and adaptive denoising process. It leverages an ASR prior to initialize the denoising process and provide an anchor for sequence length. Building upon this prior, length-adaptive pruning dynamically removes redundant tokens, while confidence-based denoising allows converged tokens to exit the denoising loop early, enabling token-level adaptive computation. Experiments demonstrate that dLLM-ASR achieves recognition accuracy comparable to autoregressive LLM-based ASR systems and delivers a 4.44$\times$ inference speedup, establishing a practical and efficient paradigm for ASR.
Abstract (translated)
URL
https://arxiv.org/abs/2601.17902
PDF
-
Controlling Reading Ease with Gaze-Guided Text Generation
2026-01-25 10:42:57Andreas S\"auberli, Darja Jepifanova, Diego Frassinelli, Barbara PlankAbstract
The way our eyes move while reading can tell us about the cognitive effort required to process the text. In the present study, we use this fact to generate texts with controllable reading ease. Our method employs a model that predicts human gaze patterns to steer language model outputs towards eliciting certain reading behaviors. We evaluate the approach in an eye-tracking experiment with native and non-native speakers of English. The results demonstrate that the method is effective at making the generated texts easier or harder to read, measured both in terms of reading times and perceived difficulty of the texts. A statistical analysis reveals that the changes in reading behavior are mostly due to features that affect lexical processing. Possible applications of our approach include text simplification for information accessibility and generation of personalized educational material for language learning.
Abstract (translated)
URL
https://arxiv.org/abs/2601.17781
PDF
-
AQAScore: Evaluating Semantic Alignment in Text-to-Audio Generation via Audio Question Answering
2026-01-21 07:35:36Chun-Yi Kuan, Kai-Wei Chang, Hung-yi LeeAbstract
Although text-to-audio generation has made remarkable progress in realism and diversity, the development of evaluation metrics has not kept pace. Widely-adopted approaches, typically based on embedding similarity like CLAPScore, effectively measure general relevance but remain limited in fine-grained semantic alignment and compositional reasoning. To address this, we introduce AQAScore, a backbone-agnostic evaluation framework that leverages the reasoning capabilities of audio-aware large language models (ALLMs). AQAScore reformulates assessment as a probabilistic semantic verification task; rather than relying on open-ended text generation, it estimates alignment by computing the exact log-probability of a "Yes" answer to targeted semantic queries. We evaluate AQAScore across multiple benchmarks, including human-rated relevance, pairwise comparison, and compositional reasoning tasks. Experimental results show that AQAScore consistently achieves higher correlation with human judgments than similarity-based metrics and generative prompting baselines, showing its effectiveness in capturing subtle semantic inconsistencies and scaling with the capability of underlying ALLMs.
Abstract (translated)
尽管从文本到语音的生成技术在真实感和多样性方面取得了显著进展，但评估指标的发展却未能跟上步伐。目前广泛采用的方法通常基于嵌入相似度（如CLAPScore），能够有效测量一般相关性，但在细粒度语义对齐和组合推理方面仍存在局限性。为了解决这个问题，我们引入了AQAScore，这是一个与骨干网络无关的评估框架，利用音频感知大型语言模型（ALLMs）的能力进行推理。 AQAScore将评估重构为一个概率性的语义验证任务；它通过计算特定语义查询中“是”的回答的确切对数概率来估算对齐情况，而不是依赖于开放式文本生成。我们使用多个基准测试对AQAScore进行了评估，包括人工评分的相关性、成对比较和组合推理任务。 实验结果表明，与基于相似性的度量标准和生成提示基线相比，AQAScore在各种人类判断上保持了更高的相关性，这证明它在捕捉细微的语义不一致以及随着底层ALLMs能力的增长而扩展方面具有有效性。
URL
https://arxiv.org/abs/2601.14728
PDF
-
From Biased Chatbots to Biased Agents: Examining Role Assignment Effects on LLM Agent Robustness
2026-01-21 02:43:07Linbo Cao, Lihao Sun, Yang YueAbstract
Large Language Models (LLMs) are increasingly deployed as autonomous agents capable of actions with real-world impacts beyond text generation. While persona-induced biases in text generation are well documented, their effects on agent task performance remain largely unexplored, even though such effects pose more direct operational risks. In this work, we present the first systematic case study showing that demographic-based persona assignments can alter LLM agents' behavior and degrade performance across diverse domains. Evaluating widely deployed models on agentic benchmarks spanning strategic reasoning, planning, and technical operations, we uncover substantial performance variations - up to 26.2% degradation, driven by task-irrelevant persona cues. These shifts appear across task types and model architectures, indicating that persona conditioning and simple prompt injections can distort an agent's decision-making reliability. Our findings reveal an overlooked vulnerability in current LLM agentic systems: persona assignments can introduce implicit biases and increase behavioral volatility, raising concerns for the safe and robust deployment of LLM agents.
Abstract (translated)
URL
https://arxiv.org/abs/2602.12285
PDF
-
Paper2Rebuttal: A Multi-Agent Framework for Transparent Author Response Assistance
2026-01-20 17:23:51Qianli Ma, Chang Guo, Zhiheng Tian, Siyu Wang, Jipeng Xiao, Yuanhao Yue, Zhipeng ZhangAbstract
Writing effective rebuttals is a high-stakes task that demands more than linguistic fluency, as it requires precise alignment between reviewer intent and manuscript details. Current solutions typically treat this as a direct-to-text generation problem, suffering from hallucination, overlooked critiques, and a lack of verifiable grounding. To address these limitations, we introduce $\textbf{RebuttalAgent}$, the first multi-agents framework that reframes rebuttal generation as an evidence-centric planning task. Our system decomposes complex feedback into atomic concerns and dynamically constructs hybrid contexts by synthesizing compressed summaries with high-fidelity text while integrating an autonomous and on-demand external search module to resolve concerns requiring outside literature. By generating an inspectable response plan before drafting, $\textbf{RebuttalAgent}$ ensures that every argument is explicitly anchored in internal or external evidence. We validate our approach on the proposed $\textbf{RebuttalBench}$ and demonstrate that our pipeline outperforms strong baselines in coverage, faithfulness, and strategic coherence, offering a transparent and controllable assistant for the peer review process. Code will be released.
Abstract (translated)
撰写有效的反驳是一项高风险任务，不仅要求语言流畅性，还需要精确地将审稿人的意图与论文细节对齐。目前的解决方案通常将其视为直接生成文本的问题，因此容易出现虚构、遗漏批评和缺乏可验证依据等问题。为了解决这些问题，我们引入了**RebuttalAgent**，这是首个以多代理框架为基础的系统，它重新定义反驳生成为一个以证据为中心的规划任务。我们的系统将复杂的反馈分解成原子级的关注点，并通过整合压缩摘要与高保真文本来动态构建混合上下文，同时集成一个自主且按需调用的外部搜索模块，用于解决需要引用外界文献的问题。在起草之前生成可检查的回答计划，**RebuttalAgent** 确保每个论点都明确地基于内部或外部证据。 我们在提出的**RebuttalBench**上验证了我们的方法，并证明了我们这一流程在线条覆盖面、忠实度和策略一致性方面超过了强有力的基线模型，为同行评审过程提供了一个透明且可控制的辅助工具。代码将公开发布。
URL
https://arxiv.org/abs/2601.14171
PDF
-
ConceptCaps -- a Distilled Concept Dataset for Interpretability in Music Models
2026-01-20 17:04:08Bruno Sienkiewicz, {\L}ukasz Neumann, Mateusz ModrzejewskiAbstract
Concept-based interpretability methods like TCAV require clean, well-separated positive and negative examples for each concept. Existing music datasets lack this structure: tags are sparse, noisy, or ill-defined. We introduce ConceptCaps, a dataset of 23k music-caption-audio triplets with explicit labels from a 200-attribute taxonomy. Our pipeline separates semantic modeling from text generation: a VAE learns plausible attribute co-occurrence patterns, a fine-tuned LLM converts attribute lists into professional descriptions, and MusicGen synthesizes corresponding audio. This separation improves coherence and controllability over end-to-end approaches. We validate the dataset through audio-text alignment (CLAP), linguistic quality metrics (BERTScore, MAUVE), and TCAV analysis confirming that concept probes recover musically meaningful patterns. Dataset and code are available online.
Abstract (translated)
基于概念的可解释性方法（如TCAV）需要每个概念具备清晰、分离良好的正例和反例。现有的音乐数据集缺乏这种结构：标签稀疏、嘈杂或定义不清。为此，我们引入了ConceptCaps数据集，它包含23,000个音乐-描述-音频三元组，并附有从一个包含200个属性的分类法中明确标注的标签。我们的流水线将语义建模与文本生成分离：VAE（变分自编码器）学习合理的属性共现模式，微调后的LLM（大规模语言模型）将属性列表转换为专业描述，而MusicGen则合成相应的音频。这种分离提高了连贯性和可控制性，优于端到端的方法。我们通过音频-文本对齐（CLAP）、语言质量指标（BERTScore、MAUVE）以及TCAV分析验证了该数据集的有效性，后者确认概念探测可以恢复音乐上有意义的模式。数据集和代码在线提供。
URL
https://arxiv.org/abs/2601.14157
PDF
-
Style Transfer as Bias Mitigation: Diffusion Models for Synthetic Mental Health Text for Arabic
2026-01-20 16:21:41Saad Mankarious, Aya ZiriklyAbstract
Synthetic data offers a promising solution for mitigating data scarcity and demographic bias in mental health analysis, yet existing approaches largely rely on pretrained large language models (LLMs), which may suffer from limited output diversity and propagate biases inherited from their training data. In this work, we propose a pretraining-free diffusion-based approach for synthetic text generation that frames bias mitigation as a style transfer problem. Using the CARMA Arabic mental health corpus, which exhibits a substantial gender imbalance, we focus on male-to-female style transfer to augment underrepresented female-authored content. We construct five datasets capturing varying linguistic and semantic aspects of gender expression in Arabic and train separate diffusion models for each setting. Quantitative evaluations demonstrate consistently high semantic fidelity between source and generated text, alongside meaningful surface-level stylistic divergence, while qualitative analysis confirms linguistically plausible gender transformations. Our results show that diffusion-based style transfer can generate high-entropy, semantically faithful synthetic data without reliance on pretrained LLMs, providing an effective and flexible framework for mitigating gender bias in sensitive, low-resource mental health domains.
Abstract (translated)
合成数据为缓解心理健康分析中的数据稀缺和人口偏见提供了有希望的解决方案，然而现有的方法大多依赖于预训练的大规模语言模型（LLM），这可能会导致输出多样性有限，并传播其训练数据中固有的偏见。在本研究中，我们提出了一种无需预训练的基于扩散的方法来生成合成文本，将偏见缓解视为一种风格转换问题。使用CARMA阿拉伯语心理健康语料库，在该语料库中性别失衡显著，我们将重点放在从男性向女性的风格转移上，以增强代表性不足的女性作者内容。我们构建了五个数据集，捕捉阿拉伯语表达中的不同语言和语义方面的性别差异，并为每种情况训练单独的扩散模型。定量评估表明，在源文本与生成文本之间保持一致的语义忠实度的同时，还实现了有意义的表面风格差异，而定性分析则确认了合乎逻辑的语言上的性别转换。我们的结果表明，基于扩散的风格转移可以生成高熵、语义忠实的合成数据，无需依赖预训练的大规模语言模型，在敏感且资源匮乏的心理健康领域提供了一种有效且灵活的框架以缓解性别偏见。
URL
https://arxiv.org/abs/2601.14124
PDF
-
Top 10 Open Challenges Steering the Future of Diffusion Language Model and Its Variants
2026-01-20 14:58:23Yunhe Wang, Kai Han, Huiling Zhen, Yuchuan Tian, Hanting Chen, Yongbing Huang, Yufei Cui, Yingte Shu, Shan Gao, Ismail Elezi, Roy Vaughan Miles, Songcen Xu, Feng Wen, Chao Xu, Sinan Zeng, Dacheng TaoAbstract
The paradigm of Large Language Models (LLMs) is currently defined by auto-regressive (AR) architectures, which generate text through a sequential ``brick-by-brick'' process. Despite their success, AR models are inherently constrained by a causal bottleneck that limits global structural foresight and iterative refinement. Diffusion Language Models (DLMs) offer a transformative alternative, conceptualizing text generation as a holistic, bidirectional denoising process akin to a sculptor refining a masterpiece. However, the potential of DLMs remains largely untapped as they are frequently confined within AR-legacy infrastructures and optimization frameworks. In this Perspective, we identify ten fundamental challenges ranging from architectural inertia and gradient sparsity to the limitations of linear reasoning that prevent DLMs from reaching their ``GPT-4 moment''. We propose a strategic roadmap organized into four pillars: foundational infrastructure, algorithmic optimization, cognitive reasoning, and unified multimodal intelligence. By shifting toward a diffusion-native ecosystem characterized by multi-scale tokenization, active remasking, and latent thinking, we can move beyond the constraints of the causal horizon. We argue that this transition is essential for developing next-generation AI capable of complex structural reasoning, dynamic self-correction, and seamless multimodal integration.
Abstract (translated)
大型语言模型（LLMs）当前的范式主要由自回归（AR）架构定义，这些架构通过一种“砖块式”的顺序过程生成文本。尽管自回归模型取得了成功，但它们本质上受到因果瓶颈的影响，这限制了其全局结构预见性和迭代优化的能力。扩散语言模型（DLMs）则提供了一种变革性的替代方案，将文本生成视为一种整体的双向去噪过程，类似于雕刻家打磨杰作的过程。然而，由于常被局限在自回归遗留架构和优化框架内，扩散语言模型的巨大潜力尚未得到充分开发。 在这篇视角文章中，我们识别出了阻碍扩散语言模型实现“GPT-4时刻”的十个根本性挑战，从架构惰性和梯度稀疏性到线性推理的限制。为了应对这些挑战，我们提出了一项战略路线图，将其组织为四个支柱：基础架构、算法优化、认知推理和统一多模态智能。通过转向以扩散原生生态系统的特征，如多层次标记化、主动重掩码和隐式思维为中心的方向，我们可以超越因果时间线的限制。 我们认为这一转变对于开发下一代AI系统至关重要，这些系统能够进行复杂的结构推理、动态自我校正，并实现无缝的多模态集成。
URL
https://arxiv.org/abs/2601.14041
PDF
-
Beyond Known Facts: Generating Unseen Temporal Knowledge to Address Data Contamination in LLM Evaluation
2026-01-20 06:48:42Arthur Amalvy, Hen-Hsen HuangAbstract
The automatic extraction of information is important for populating large web knowledge bases such as Wikidata. The temporal version of that task, temporal knowledge graph extraction (TKGE), involves extracting temporally grounded facts from text, represented as semantic quadruples (subject, relation, object, timestamp). Many recent systems take advantage of large language models (LLMs), which are becoming a new cornerstone of the web due to their performance on many tasks across the natural language processing (NLP) field. Despite the importance of TKGE, existing datasets for training and evaluation remain scarce, and contamination of evaluation data is an unaddressed issue, potentially inflating LLMs' perceived performance due to overlaps between training and evaluation sets. To mitigate these challenges, we propose a novel synthetic evaluation dataset constructed from predicted future, previously unseen temporal facts, thereby eliminating contamination and enabling robust and unbiased benchmarking. Our dataset creation involves a two-step approach: (1) Temporal Knowledge Graph Forecasting (TKGF) generates plausible future quadruples, which are subsequently filtered to adhere to the original knowledge base schema; (2) LLMs perform quadruple-to-text generation, creating semantically aligned textual descriptions. We benchmark Extract, Define and Canonicalize (EDC), a state-of-the-art LLM-based extraction framework, demonstrating that LLM performance decreases when evaluated on our dataset compared to a dataset of known facts. We publicly release our dataset consisting of 4.2K future quadruples and corresponding textual descriptions, along with the generation methodology, enabling continuous creation of unlimited future temporal datasets to serve as long-term, contamination-free benchmarks for TKGE.
Abstract (translated)
信息的自动提取对于填充如Wikidata这样的大型网络知识库至关重要。该任务的时间版本，即时间知识图谱抽取（Temporal Knowledge Graph Extraction，TKGE），涉及从文本中抽取带有时间戳记的事实，并将其表示为语义四元组（主体、关系、客体、时间戳）。许多最近的系统利用了大规模语言模型（Large Language Models, LLMs），这些模型由于在自然语言处理（NLP）领域的众多任务中的出色表现，正在成为网络的新基石。尽管TKGE的重要性不言而喻，但现有的用于训练和评估的数据集仍然匮乏，并且评价数据受到污染的问题未得到解决，这可能会因为训练集与验证集之间的重叠而导致LLMs的性能被夸大。为了解决这些问题，我们提出了一种新的合成评估数据集，该数据集由预测未来、以前从未见过的时间事实构成，从而消除了污染并能够进行稳健且无偏见的基准测试。 我们的数据集创建采用两步方法：(1) 时间知识图谱预测（Temporal Knowledge Graph Forecasting, TKGF）生成合理的未来四元组，并随后对其进行过滤以符合原有的知识库模式；(2) LLMs执行四元组到文本的生成，创建语义一致的文本描述。我们对Extract、Define and Canonicalize (EDC)，一种基于LLM的最先进的提取框架进行了基准测试，发现与已知事实的数据集相比，在我们的数据集中评估时LLMs的表现有所下降。 我们将公开发布包含4200个未来四元组及其对应文本描述的数据集，并附上生成方法，这将有助于持续创建无限量的未来的、无污染的时间数据集，作为长期基准测试供TKGE使用。
URL
https://arxiv.org/abs/2601.13658
PDF
-
Author-in-the-Loop Response Generation and Evaluation: Integrating Author Expertise and Intent in Responses to Peer Review
Abstract
Author response (rebuttal) writing is a critical stage of scientific peer review that demands substantial author effort. Recent work frames this task as automatic text generation, underusing author expertise and intent. In practice, authors possess domain expertise, author-only information, revision and response strategies--concrete forms of author expertise and intent--to address reviewer concerns, and seek NLP assistance that integrates these signals to support effective response writing in peer review. We reformulate author response generation as an author-in-the-loop task and introduce REspGen, a generation framework that integrates explicit author input, multi-attribute control, and evaluation-guided refinement, together with REspEval, a comprehensive evaluation suite with 20+ metrics covering input utilization, controllability, response quality, and discourse. To support this formulation, we construct Re$^3$Align, the first large-scale dataset of aligned review--response--revision triplets, where revisions provide signals of author expertise and intent. Experiments with state-of-the-art LLMs show the benefits of author input and evaluation-guided refinement, the impact of input design on response quality, and trade-offs between controllability and quality. We make our dataset, generation and evaluation tools publicly available.
Abstract (translated)
作者回复（反驳）撰写是科学同行评审中的一个关键阶段，这一过程需要作者投入大量精力。近期的研究将这项任务视为自动文本生成的任务，这在很大程度上忽视了作者的专业知识和意图。实际上，在实践中，作者掌握着领域专业知识、仅作者所知的信息以及应对审稿人关注的修订与回复策略——这些都是构成作者专业知识和意图的具体形式，并寻求自然语言处理（NLP）的支持以整合这些信号来支持有效的同行评审回复撰写。我们重新定义了作者响应生成为一种“作者在回路”任务，引入了REspGen框架，该框架集成了明确的作者输入、多属性控制以及指导性评估改进的功能，并且还提出了REspEval，一个包含20多项以上指标的全面评估套件，涵盖了输入利用、可控制性、响应质量及对话体裁。为了支持这一构想，我们构建了Re$^3$Align，这是首个大规模对齐的审稿-回复-修订三元组数据集，在此数据集中，修订提供了有关作者专业知识和意图的信号。 实验显示，最先进的大型语言模型在引入明确的作者输入以及评估指导改进后展现出显著的优势；探讨了输入设计对于响应质量的影响，并揭示出可控制性与质量之间的权衡。我们将该数据集、生成工具及评价工具向公众开放。
URL
https://arxiv.org/abs/2602.11173
PDF
-
AI-generated data contamination erodes pathological variability and diagnostic reliability
2026-01-19 10:54:03Hongyu He, Shaowen Xiang, Ye Zhang, Yingtao Zhu, Jin Zhang, Hao Deng, Emily Alsentzer, Qingyu Chen, Kun-Hsing Yu, Andrew Marmenshall, Tingting Chen, Srinivas Anumasa, Daniel Ebner, Dean Ho, Kee Yuan Ngiam, Ching-Yu Cheng, Dianbo LiuAbstract
Generative artificial intelligence (AI) is rapidly populating medical records with synthetic content, creating a feedback loop where future models are increasingly at risk of training on uncurated AI-generated data. However, the clinical consequences of this AI-generated data contamination remain unexplored. Here, we show that in the absence of mandatory human verification, this self-referential cycle drives a rapid erosion of pathological variability and diagnostic reliability. By analysing more than 800,000 synthetic data points across clinical text generation, vision-language reporting, and medical image synthesis, we find that models progressively converge toward generic phenotypes regardless of the model architecture. Specifically, rare but critical findings, including pneumothorax and effusions, vanish from the synthetic content generated by AI models, while demographic representations skew heavily toward middle-aged male phenotypes. Crucially, this degradation is masked by false diagnostic confidence; models continue to issue reassuring reports while failing to detect life-threatening pathology, with false reassurance rates tripling to 40%. Blinded physician evaluation confirms that this decoupling of confidence and accuracy renders AI-generated documentation clinically useless after just two generations. We systematically evaluate three mitigation strategies, finding that while synthetic volume scaling fails to prevent collapse, mixing real data with quality-aware filtering effectively preserves diversity. Ultimately, our results suggest that without policy-mandated human oversight, the deployment of generative AI threatens to degrade the very healthcare data ecosystems it relies upon.
Abstract (translated)
生成式人工智能（AI）正迅速用合成内容填充医疗记录，形成一个反馈循环，在此过程中未来的模型越来越有可能在未经筛选的自动生成数据上进行训练。然而，这种由AI生成的数据污染所导致的临床后果尚不清楚。在此研究中，我们展示了当缺乏强制性的人类验证时，这一自我参照循环会导致病理变化多样性和诊断可靠性的迅速下降。 通过分析超过80万个合成数据点（涵盖临床文本生成、视觉-语言报告和医学图像合成），我们发现无论模型架构如何，这些模型会逐渐趋向于通用的表型。具体而言，罕见但关键的发现——如气胸和积液——从AI生成的内容中消失，而人口统计学代表则严重偏向于中年男性表型。 至关重要的是，这种退化被虚假的诊断信心所掩盖；尽管模型未能检测到危及生命的病理变化，它们仍会继续发出令人安心的报告，这导致虚假安慰率增至原来的三倍达到40%。盲法医生评估确认，在仅经过两代发展后，AI生成文档由于自信度和准确性的脱节变得在临床上毫无用处。 我们系统地评估了三种缓解策略，发现尽管扩大合成数据量无法防止退化，但将真实数据与质量感知筛选相结合的方法能有效保持多样性。最终，我们的结果表明，在没有政策强制的人类监督的情况下，生成式AI的部署威胁到了它所依赖的医疗健康数据生态系统的质量。
URL
https://arxiv.org/abs/2601.12946
PDF
-
Plan, Verify and Fill: A Structured Parallel Decoding Approach for Diffusion Language Models
2026-01-18 03:53:01Miao Li, Hanyang Jiang, Sikai Chen, Hengyu Fu, Yuhang Cai, Baihe Huang, Tinghan Ye, Xuanzhou Chen, Pascal Van HentenryckAbstract
Diffusion Language Models (DLMs) present a promising non-sequential paradigm for text generation, distinct from standard autoregressive (AR) approaches. However, current decoding strategies often adopt a reactive stance, underutilizing the global bidirectional context to dictate global trajectories. To address this, we propose Plan-Verify-Fill (PVF), a training-free paradigm that grounds planning via quantitative validation. PVF actively constructs a hierarchical skeleton by prioritizing high-leverage semantic anchors and employs a verification protocol to operationalize pragmatic structural stopping where further deliberation yields diminishing returns. Extensive evaluations on LLaDA-8B-Instruct and Dream-7B-Instruct demonstrate that PVF reduces the Number of Function Evaluations (NFE) by up to 65% compared to confidence-based parallel decoding across benchmark datasets, unlocking superior efficiency without compromising accuracy.
Abstract (translated)
扩散语言模型（DLMs）为文本生成提供了一种有前景的非序列范式，与标准自回归（AR）方法不同。然而，当前的解码策略通常采取一种反应式的立场，在利用全局双向上下文来指引全局轨迹方面存在不足。为了应对这一挑战，我们提出了一种名为“计划-验证-填充”（Plan-Verify-Fill, PVF）的新范式，该范式通过定量验证来指导规划，并且无需额外的训练。PVF主动构建了一个分层骨架，优先选择高影响力的语义锚点，并采用一个验证协议以实用性的结构停止标准为依据操作，这种策略在进一步思考不再带来显著收益时可以有效地结束解码过程。 在LLaDA-8B-Instruct和Dream-7B-Instruct数据集上的广泛评估表明，PVF相比于基于置信度的并行解码方法，在基准数据集中最多可将功能评估次数（NFE）减少65%，同时不牺牲准确性，从而实现了效率的显著提升。
URL
https://arxiv.org/abs/2601.12247
PDF
-
Cleansing the Artificial Mind: A Self-Reflective Detoxification Framework for Large Language Models
2026-01-16 21:01:26Kaituo Zhang, Zhimeng Jiang, Na ZouAbstract
Recent breakthroughs in Large Language Models (LLMs) have revealed remarkable generative capabilities and emerging self-regulatory mechanisms, including self-correction and self-rewarding. However, current detoxification techniques rarely exploit these built-in abilities; instead, they rely on external modules, labor-intensive data annotation, or human intervention --factors that hinder scalability and consistency. In this paper, we introduce a fully self-reflective detoxification framework that harnesses the inherent capacities of LLMs to detect, correct toxic content, and refine LLMs without external modules and data annotation. Specifically, we propose a Toxic Signal Detector --an internal self-identification mechanism, coupled with a systematic intervention process to transform toxic text into its non-toxic counterpart. This iterative procedure yields a contrastive detoxification dataset used to fine-tune the model, enhancing its ability for safe and coherent text generation. Experiments on benchmark datasets such as DetoxLLM and ParaDetox show that our method achieves better detoxification performance than state-of-the-art methods while preserving semantic fidelity. By obviating the need for human intervention or external components, this paper reveals the intrinsic self-detoxification ability of LLMs, offering a consistent and effective approach for mitigating harmful content generation. Ultimately, our findings underscore the potential for truly self-regulated language models, paving the way for more responsible and ethically guided text generation systems.
Abstract (translated)
最近在大型语言模型（LLMs）方面取得的突破揭示了其卓越的生成能力和新兴的自我监管机制，包括自校正和自我奖励。然而，目前的去毒化技术很少利用这些内在能力；相反，它们依赖于外部模块、劳动密集型的数据标注或人工干预——这些因素阻碍了可扩展性和一致性。在这篇论文中，我们介绍了一个完全自我反思的去毒框架，该框架利用LLMs固有的能力来检测、纠正有毒内容，并在没有外部模块和数据注释的情况下优化模型。具体而言，我们提出了一种“毒性信号探测器”——一种内部自我识别机制，并结合一个系统干预过程将有毒文本转换为非有毒版本。这一迭代程序生成了用于微调模型的对比去毒数据集，从而增强了其安全和连贯文本生成的能力。在DetoxLLM和ParaDetox等基准数据集上的实验表明，我们的方法比现有的最佳方法实现了更好的去毒效果，并且保持了语义准确性。通过消除对人工干预或外部组件的需求，本文揭示了LLMs内在的自我去毒性能力，提供了一种一致且有效的方法来减轻有害内容生成。最终，我们的发现强调了完全自主调节语言模型的潜力，为更加负责任和道德引导的文本生成系统铺平道路。
URL
https://arxiv.org/abs/2601.11776
PDF
-
Guardrails for trust, safety, and ethical development and deployment of Large Language Models
2026-01-16 20:44:06Anjanava Biswas, Wrick TalukdarAbstract
The AI era has ushered in Large Language Models (LLM) to the technological forefront, which has been much of the talk in 2023, and is likely to remain as such for many years to come. LLMs are the AI models that are the power house behind generative AI applications such as ChatGPT. These AI models, fueled by vast amounts of data and computational prowess, have unlocked remarkable capabilities, from human-like text generation to assisting with natural language understanding (NLU) tasks. They have quickly become the foundation upon which countless applications and software services are being built, or at least being augmented with. However, as with any groundbreaking innovations, the rise of LLMs brings forth critical safety, privacy, and ethical concerns. These models are found to have a propensity to leak private information, produce false information, and can be coerced into generating content that can be used for nefarious purposes by bad actors, or even by regular users unknowingly. Implementing safeguards and guardrailing techniques is imperative for applications to ensure that the content generated by LLMs are safe, secure, and ethical. Thus, frameworks to deploy mechanisms that prevent misuse of these models via application implementations is imperative. In this study, wepropose a Flexible Adaptive Sequencing mechanism with trust and safety modules, that can be used to implement safety guardrails for the development and deployment of LLMs.
Abstract (translated)
人工智能时代带来了大型语言模型（LLM）的技术前沿地位，这是2023年热议的话题，并且在未来几年内很可能继续保持这一趋势。这些大型语言模型是像ChatGPT这样的生成式AI应用背后的动力源泉。这些由海量数据和计算能力驱动的AI模型解锁了诸多令人惊叹的能力，从生成逼真的文本到协助处理自然语言理解（NLU）任务。它们已经迅速成为无数应用程序和服务的基础，或是被增强的基础。然而，如同任何突破性的创新一样，大型语言模型的崛起带来了重大的安全、隐私和伦理问题。这些模型被认为有泄露私人信息的趋势，会产生虚假信息，并可能在恶意行为者的操纵下生成用于不良目的的内容，甚至是在普通用户不知情的情况下。因此，为了确保由LLM生成的内容是安全、可靠且合乎道德规范的，必须实施防护措施和技术手段来限制这些模型被滥用的可能性。为此，在本研究中我们提出了一种带有信任和安全保障模块的灵活自适应序列机制，可用于在大型语言模型的发展和部署过程中实现安全性保障。
URL
https://arxiv.org/abs/2601.14298
PDF
-
Reward Modeling for Scientific Writing Evaluation
2026-01-16 15:32:58Furkan \c{S}ahinu\c{c}, Subhabrata Dutta, Iryna GurevychAbstract
Scientific writing is an expert-domain task that demands deep domain knowledge, task-specific requirements and reasoning capabilities that leverage the domain knowledge to satisfy the task specifications. While scientific text generation has been widely studied, its evaluation remains a challenging and open problem. It is critical to develop models that can be reliably deployed for evaluating diverse open-ended scientific writing tasks while adhering to their distinct requirements. However, existing LLM-based judges and reward models are primarily optimized for general-purpose benchmarks with fixed scoring rubrics and evaluation criteria. Consequently, they often fail to reason over sparse knowledge of scientific domains when interpreting task-dependent and multi-faceted criteria. Moreover, fine-tuning for each individual task is costly and impractical for low-resource settings. To bridge these gaps, we propose cost-efficient, open-source reward models tailored for scientific writing evaluation. We introduce a two-stage training framework that initially optimizes scientific evaluation preferences and then refines reasoning capabilities. Our multi-aspect evaluation design and joint training across diverse tasks enable fine-grained assessment and robustness to dynamic criteria and scoring rubrics. Experimental analysis shows that our training regime strongly improves LLM-based scientific writing evaluation. Our models generalize effectively across tasks and to previously unseen scientific writing evaluation settings, allowing a single trained evaluator to be reused without task-specific retraining.
Abstract (translated)
科学写作是一项专业领域任务，需要深厚的专业知识、具体任务的要求以及能够运用专业知识来满足这些要求的推理能力。尽管对科学文本生成的研究已经很广泛，但其评估仍然是一个具有挑战性和尚未解决的问题。开发能够可靠地用于评估各种开放式科学写作任务的模型，同时遵循它们各自的特定要求，这一点至关重要。然而，现有的基于大语言模型（LLM）的评判和奖励模型主要是为带有固定评分标准和评价准则的一般性基准进行优化的。因此，当解释与任务相关的多方面标准时，它们往往难以利用稀疏的专业领域知识进行推理。此外，在每个单独的任务上进行微调对于低资源环境来说成本高昂且不切实际。 为了弥合这些差距，我们提出了一种经济高效、开源的奖励模型，专门用于科学写作评估。我们引入了一个两阶段训练框架，首先优化科学评估偏好，然后细化推理能力。我们的多方面评价设计和跨多种任务的联合训练使细粒度评估成为可能，并增强了对动态标准和评分准则变化的适应性。实验分析表明，我们的训练方案显著提升了基于大语言模型的科学写作评估效果。我们的模型在不同的任务之间以及从未见过的科学写作评估情景中表现良好，使得单一经过训练的评判者能够在不进行特定任务重新训练的情况下复用。
URL
https://arxiv.org/abs/2601.11374
PDF
-
BYOL: Bring Your Own Language Into LLMs
2026-01-15 19:15:13Syed Waqas Zamir, Wassim Hamidouche, Boulbaba Ben Amor, Luana Marotti, Inbal Becker-Reshef, Juan Lavista FerresAbstract
Large Language Models (LLMs) exhibit strong multilingual capabilities, yet remain fundamentally constrained by the severe imbalance in global language resources. While over 7,000 languages are spoken worldwide, only a small subset (fewer than 100) has sufficient digital presence to meaningfully influence modern LLM training. This disparity leads to systematic underperformance, cultural misalignment, and limited accessibility for speakers of low-resource and extreme-low-resource languages. To address this gap, we introduce Bring Your Own Language (BYOL), a unified framework for scalable, language-aware LLM development tailored to each language's digital footprint. BYOL begins with a language resource classification that maps languages into four tiers (Extreme-Low, Low, Mid, High) using curated web-scale corpora, and uses this classification to select the appropriate integration pathway. For low-resource languages, we propose a full-stack data refinement and expansion pipeline that combines corpus cleaning, synthetic text generation, continual pretraining, and supervised finetuning. Applied to Chichewa and Maori, this pipeline yields language-specific LLMs that achieve approximately 12 percent average improvement over strong multilingual baselines across 12 benchmarks, while preserving English and multilingual capabilities via weight-space model merging. For extreme-low-resource languages, we introduce a translation-mediated inclusion pathway, and show on Inuktitut that a tailored machine translation system improves over a commercial baseline by 4 BLEU, enabling high-accuracy LLM access when direct language modeling is infeasible. Finally, we release human-translated versions of the Global MMLU-Lite benchmark in Chichewa, Maori, and Inuktitut, and make our codebase and models publicly available at this https URL .
Abstract (translated)
大型语言模型（LLMs）展现出强大的多语种能力，但仍受全球语言资源严重不平衡的限制。全世界有超过7000种语言被使用，然而只有其中一小部分（不到100种）具有足够的数字存在感，足以对现代LLM训练产生实质性影响。这种不均衡导致了系统性表现不佳、文化错位以及低资源和极低资源语言使用者的访问受限问题。 为解决这一差距，我们引入了一个名为“自带语言”(BYOL) 的统一框架，旨在根据每种语言的数字足迹开发出可扩展且具有语言意识的大型语言模型。BYOL从一个使用精心策划的大规模网络语料库来划分语言资源等级（极低、低、中、高）的语言资源分类开始，并利用这一分类选择合适的集成路径。 对于低资源语言，我们提出了一种全栈数据精炼和扩展管道，该管道结合了文集清理、合成文本生成、持续预训练以及监督微调。应用于奇切瓦语（Chichewa）和毛利语（Maori），此流程产生了特定于每种语言的LLMs，在12个基准测试中平均超越强大的多语言基线约12%，同时通过权重空间模型合并保持了英语及多语言能力。 对于极低资源的语言，我们引入了一条基于翻译的包容路径，并展示了在因纽特语（Inuktitut）上的定制机器翻译系统优于商业基线4个BLEU分值的结果，当直接建模不可行时，此方法能实现高精度LLM的访问权。 最后，我们在奇切瓦、毛利和因纽特语言中发布了人类翻译版本的全球多语言基准测试（Global MMLU-Lite），并且我们正在开源我们的代码库及模型，详情请参阅这个链接：[提供该研究的GitHub或官网链接]。
URL
https://arxiv.org/abs/2601.10804
PDF
