Large language model (LLM)-based search agents synthesize open-web content into actionable recommendations on behalf of users, creating a risk that attacker-published pages are transformed into endorsed claims. We introduce SearchGEO, a controlled evaluation framework for measuring endorsement corruption in LLM-based web-search agents, combining a web-evidence manipulation pipeline, a five-mode attack taxonomy, and multiple output-level metrics. We evaluate 13 LLM backends on 308 cases each. Results show that vulnerability patterns vary across backends: overall attack success rate (ASR) ranges from 0.0% on Claude-Sonnet-4.6 to 31.4% on Gemini-3-Flash, the strongest attack mode differs by model family, and the same deployment scaffold could amplify or decrease ASR on different backends. An auxiliary agent-skill probe, where endorsement becomes an install command, exposes a sharp split among otherwise robust backends: Claude over-rejects while GPT over-trusts. These findings argue for treating recommendation reliability under adversarial search content as a first-class dimension of backend safety evaluation.

We present AnyCap, a plug-and-play framework that brings instruction alignment to omni-modal captioning. Captions offer a unified language interface for multimodal learning, and users increasingly expect instruction-driven control over their content and style. Current caption models lack explicit instruction supervision and are weak at instruction following, while directly tuning them can degrade general language ability. Achieving instruction alignment in an omni-modal setting is harder still, as each modality calls for separate models and custom designs. To address these challenges, AnyCap leverages a multimodal revision-based refinement paradigm that refines uncontrolled captions from existing models to instruction-aligned ones, without re-training base models. By processing multi-modality features in a unified framework, it enables one model to serve images, videos, and audio. To address the lack of instruction-based data, we construct AnyCapData, a large-scale, high-quality corpus spanning three modalities with 28 well-designed instruction types, each with multiple instruction samples. For evaluation, we address the limitations of current metrics for instruction-oriented captioning by designing AnyCapEval. Its key insight is to decouple evaluation into content and style for fine-grained assessment. Extensive experiments show that on AnyCapEval and diverse public benchmarks, AnyCap consistently improves both caption quality and instruction adherence for both open-source and API-based models. Notably, AnyCap-8B boosts GPT-4o’s content scores by 46% and style scores by 12%. Our code and models will be made publicly available.

LLM-based agents can interact with external environments through tool invocation, but this capability also introduces security risks such as file modification, information leakage, and unauthorized actions. Existing guardrails often evaluate completed trajectories, leaving pre-execution monitoring of step-level actions underexplored. We propose StepGuard, a step-level guard model that can audit completed agent trajectories and check tool actions before they are executed. To train StepGuard, we introduce StepGen, an automatic data engine that generates safe and unsafe trajectories with the same context but different actions at the risky step. To further reduce over-defense and under-defense, we propose Balance-GRPO, which dynamically balances learning between safe and unsafe actions based on their observed accuracy. Experiments show that StepGuard achieves the highest average accuracy among open-weight guard models, with performance comparable to GPT-5.4. When used to guard agents on AgentDojo and AgentDyn, StepGuard reduces mean ASR by 77.3% relative to the no-guard setting, while mean utility drops by only 2.8 points.

Recent advances in Large Language Models (LLMs) have significantly enhanced their capabilities, highlighting the need for comprehensive evaluation frameworks that extend beyond task-specific benchmarks.However, existing benchmarks often focus on isolated abilities, lacking a holistic framework for assessing LLM capabilities.To address this gap, we propose the Cognition-Domain-Task (CDT) framework, which comprehensively measures a model’s capabilities across three dimensions.We expand the scope of model capability definitions at the cognitive level by incorporating the Cattell-Horn-Carroll cognitive theory, refining the categorization of model capabilities.We apply CDT in two directions: dataset capability evaluation and data selection. Experiments show that our capability metrics correlate well with downstream performance and can support effective dataset analysis and construction. The experiments on data selection also show significant improvements in both general and specific benchmarks, achieving scores of 44.3 and 45.4, with an increase of 1.6 and 2.2 points over the baselines, respectively. These results validate the effectiveness and practicality of CDT. Source code and models are available at https://github.com/Alessa-mo/CDT.

Large language models (LLMs) excel in problem-solving but require training data with diverse reasoning processes. Existing methods mainly optimize instruction-response pairs but lack a systematic design for the underlying reasoning structure. This paper proposes RSS: a Reasoning Structure driven data Synthesis method. We first proactively develop a hierarchical GFlowNet to construct reasoning structures efficiently through a coarse-to-fine directed acyclic graph (DAG) growth process. Then reasoning DAGs are leveraged to actively guide the instruction generation via an iterative suggester-editor workflow and enhance response quality using a structure-aware strategy. Experiments show that LLMs trained on our synthetic datasets achieve 48.50%, 84.00%, 79.90% for AlpacaEval2, GSM8K and HumanEval, outperforming existing data synthesis methods.

With the widespread adoption of Large Language Models (LLMs), jailbreak attacks have become an increasingly pressing safety concern. While safety-aligned LLMs can effectively defend against normal harmful queries, they remain vulnerable to such attacks. Existing defense methods primarily rely on fine-tuning or input modification, which often suffer from limited generalization and reduced utility. To address this, we introduce DeTAM, a finetuning-free defense approach that improves the defensive capabilities against jailbreak attacks of LLMs via targeted attention modification. Specifically, we analyze the differences in attention scores between successful and unsuccessful defenses to identify the attention heads sensitive to jailbreak attacks. During inference, we reallocate attention to emphasize users’ core intentions, minimizing interference from attack tokens. Our experimental results demonstrate that DeTAM outperforms various baselines in jailbreak defense and exhibits robust generalization across different attacks and models, maintaining its effectiveness even on in-the-wild jailbreak data. Furthermore, we compare DeTAM with the baselines on over-defense datasets, further validating its superior balance between helpfulness and harmlessness.

With the development of large language models, their ability to follow simple instructions has significantly improved. However, adhering to complex instructions remains a major challenge. Current approaches to generating complex instructions are often irrelevant to the current instruction requirements or suffer from limited scalability and diversity. Moreover, methods such as back-translation, while effective for simple instruction generation, fail to leverage the rich knowledge and formatting in human written documents. In this paper, we propose a novel Automatic Iterative Refinement (AIR) framework to generate complex instructions with constraints, which not only better reflects the requirements of real scenarios but also significantly enhances LLMs’ ability to follow complex instructions. The AIR framework consists of two stages: 1) Generate an initial instruction from a document; 2) Iteratively refine instructions with LLM-as-judge guidance by comparing the model’s output with the document to incorporate valuable constraints. Finally, we construct the AIR-10K dataset with 10K complex instructions and demonstrate that instructions generated with our approach significantly improve the model’s ability to follow complex instructions, outperforming existing methods for instruction generation.

The rapid evolution of artificial intelligence in drug discovery encounters challenges with generalization and extensive training, yet Large Language Models (LLMs) offer promise in reshaping interactions with complex molecular data. Our novel contribution, InstructMol, a multi-modal LLM, effectively aligns molecular structures with natural language via an instruction-tuning approach, utilizing a two-stage training strategy that adeptly combines limited domain-specific data with molecular and textual information. InstructMol showcases substantial performance improvements in drug discovery-related molecular tasks, surpassing leading LLMs and significantly reducing the gap with specialists, thereby establishing a robust foundation for a versatile and dependable drug discovery assistant.

The proliferation of jailbreak attacks against large language models (LLMs) highlights the need for robust security measures. However, in multi-round dialogues, malicious intentions may be hidden in interactions, leading LLMs to be more prone to produce harmful responses. In this paper, we propose the Multi-Turn Safety Alignment (MTSA) framework, to address the challenge of securing LLMs in multi-round interactions. It consists of two stages: In the thought-guided attack learning stage, the red-team model learns about thought-guided multi-round jailbreak attacks to generate adversarial prompts. In the adversarial iterative optimization stage, the red-team model and the target model continuously improve their respective capabilities in interaction. Furthermore, we introduce a multi-turn reinforcement learning algorithm based on future rewards to enhance the robustness of safety alignment. Experimental results show that the red-team model exhibits state-of-the-art attack capabilities, while the target model significantly improves its performance on safety benchmarks.

Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks. Existing defense mechanisms have not fully deleted harmful knowledge in LLMs, which allows such attacks to bypass safeguards and produce harmful outputs. To address this challenge, we propose a novel safety alignment strategy, Constrained Knowledge Unlearning (CKU), which focuses on two primary objectives: knowledge localization and retention, and unlearning harmful knowledge. CKU works by scoring neurons in specific multilayer perceptron (MLP) layers to identify a subset U of neurons associated with useful knowledge. During the unlearning process, CKU prunes the gradients of neurons in U to preserve valuable knowledge while effectively mitigating harmful content. Experimental results demonstrate that CKU significantly enhances model safety without compromising overall performance, offering a superior balance between safety and utility compared to existing methods. Additionally, our analysis of neuron knowledge sensitivity across various MLP layers provides valuable insights into the mechanics of safety alignment and model knowledge editing.

Chatbots have become popular in educational settings, revolutionizing how students interact with material and how teachers teach. We present Curriculum-Driven EduBot, a framework for developing a chatbot that combines the interactive features of chatbots with the systematic material of English textbooks to assist students in enhancing their conversational skills. We begin by extracting pertinent topics from textbooks and using large language models to generate dialogues related to these topics. We then fine-tune an open-source LLM using our generated conversational data to create our curriculum-driven chatbot. User studies demonstrate that EduBot outperforms ChatGPT in leading curriculum-based dialogues and adapting its dialogue to match the user’s English proficiency level. By combining traditional textbook methodologies with conversational AI, our approach offers learners an interactive tool that aligns with their curriculum and provides user-tailored conversation practice. This facilitates meaningful student-bot dialogues and enriches the overall learning experience within the curriculum’s pedagogical framework.

Augmenting Large Language Models (LLMs) for Question Answering (QA) with domain specific data has attracted wide attention. However, domain data often exists in a hybrid format, including text and semi-structured tables, posing challenges for the seamless integration of information. Table-to-Text Generation is a promising solution by facilitating the transformation of hybrid data into a uniformly text-formatted corpus. Although this technique has been widely studied by the NLP community, there is currently no comparative analysis on how corpora generated by different table-to-text methods affect the performance of QA systems.In this paper, we address this research gap in two steps. First, we innovatively integrate table-to-text generation into the framework of enhancing LLM-based QA systems with domain hybrid data. Then, we utilize this framework in real-world industrial data to conduct extensive experiments on two types of QA systems (DSFT and RAG frameworks) with four representative methods: Markdown format, Template serialization, TPLM-based method, and LLM-based method. Based on the experimental results, we draw some empirical findings and explore the underlying reasons behind the success of some methods. We hope the findings of this work will provide a valuable reference for the academic and industrial communities in developing robust QA systems.

“在机器阅读理解领域,处理和分析多方对话一直是一项具有挑战性的研究任务。鉴于中文语境下相关数据资源的缺乏,本研究构建了DialogueMRC数据集,旨在促进该领域的研究进展。DialogueMRC数据集作为首个面向中文多方对话的机器阅读理解数据集,包含705个多方对话实例,涵盖24451个话语单元以及8305个问答对。区别于以往的MRC数据集,DialogueMRC数据集强调深入理解动态的对话过程,对模型应对多方对话中的复杂性及篇章解析能力提出了更高的要求。为应对中文多方对话机器阅读理解的挑战,本研究提出了融合篇章结构感知能力的中文多方对话问答模型(DiscourseStructure-aware QA Model for Chinese Multi-party Dialogue,DSQA-CMD),该模型融合了问答和篇章解析任务,以提升对话上下文的理解能力。实验结果表明,相较于典型的基于微调的预训练语言模型,DSQA-CMD模型表现出明显优势,对比基于Longformer的方法,DSQA-CMD模型在MRC任务的F1和EM评价指标上分别提升了5.4%和10.0%;与当前主流的大型语言模型相比,本模型也展现了更佳的性能,表明了本文所提出方法的有效性。”

Opinion summarization is expected to digest larger review sets and provide summaries from different perspectives. However, most existing solutions are deficient in epitomizing extensive reviews and offering opinion summaries from various angles due to the lack of designs for information selection. To this end, we propose SubSumm, a supervised summarization framework for large-scale multi-perspective opinion summarization. SubSumm consists of a review sampling strategy set and a two-stage training scheme. The sampling strategies take sentiment orientation and contrastive information value into consideration, with which the review subsets from different perspectives and quality levels can be selected. Subsequently, the summarizer is encouraged to learn from the sub-optimal and optimal subsets successively in order to capitalize on the massive input. Experimental results on AmaSum and Rotten Tomatoes datasets demonstrate that SubSumm is adept at generating pros, cons, and verdict summaries from hundreds of input reviews. Furthermore, our in-depth analysis verifies that the advanced selection of review subsets and the two-stage training scheme are vital to boosting the summarization performance.

Dialogue summarization has recently garnered significant attention due to its wide range of applications. However, existing methods for summarizing dialogues have limitations because they do not take into account the inherent structure of dialogue and rely heavily on labeled data, which can lead to poor performance in new domains. In this work, we propose DIONYSUS (dynamic input optimization in pre-training for dialogue summarization), a pre-trained encoder-decoder model for summarizing dialogues in any new domain. To pre-train DIONYSUS, we create two pseudo summaries for each dialogue example: one from a fine-tuned summarization model and the other from important dialogue turns. We then choose one of these pseudo summaries based on information distribution differences in different types of dialogues. This selected pseudo summary serves as the objective for pre-training DIONYSUS using a self-supervised approach on a large dialogue corpus. Our experiments show that DIONYSUS outperforms existing methods on six datasets, as demonstrated by its ROUGE scores in zero-shot and few-shot settings

We describe a new freely available Chinese multi-party dialogue dataset for automatic extraction of dialogue-based character relationships. The data has been extracted from the original TV scripts of a Chinese sitcom called “I Love My Home” with complex family-based human daily spoken conversations in Chinese. First, we introduced human annotation scheme for both global Character relationship map and character reference relationship. And then we generated the dialogue-based character relationship triples. The corpus annotates relationships between 140 entities in total. We also carried out a data exploration experiment by deploying a BERT-based model to extract character relationships on the CRECIL corpus and another existing relation extraction corpus (DialogRE (CITATION)).The results demonstrate that extracting character relationships is more challenging in CRECIL than in DialogRE.

Text-video retrieval focuses on two aspects: cross-modality interaction and video-language encoding. Currently, the mainstream approach is to train a joint embedding space for multimodal interactions. However, there are structural and semantic differences between text and video, making this approach challenging for fine-grained understanding. In order to solve this, we propose an end-to-end graph-based hierarchical aggregation network for text-video retrieval according to the hierarchy possessed by text and video. We design a token-level weighted network to refine intra-modality representations and construct a graph-based message passing attention network for global-local alignment across modality. We conduct experiments on the public datasets MSR-VTT-9K, MSR-VTT-7K and MSVD, and achieve Recall@1 of 73.0%, 65.6%, and 64.0% , which is 25.7%, 16.5%, and 14.2% better than the current state-of-the-art model.

Dialog systems are often designed or trained to output human-like responses. However, some responses may be impossible for a machine to truthfully say (e.g. “that movie made me cry”). Highly anthropomorphic responses might make users uncomfortable or implicitly deceive them into thinking they are interacting with a human. We collect human ratings on the feasibility of approximately 900 two-turn dialogs sampled from 9 diverse data sources. Ratings are for two hypothetical machine embodiments: a futuristic humanoid robot and a digital assistant. We find that for some data-sources commonly used to train dialog systems, 20-30% of utterances are not viewed as possible for a machine. Rating is marginally affected by machine embodiment. We explore qualitative and quantitative reasons for these ratings. Finally, we build classifiers and explore how modeling configuration might affect output permissibly, and discuss implications for building less falsely anthropomorphic dialog systems.

Persuasion dialogue system reflects the machine’s ability to make strategic moves beyond verbal communication, and therefore differentiates itself from task-oriented or open-domain dialogues and has its own unique values. However, the repetition and inconsistency problems still persist in dialogue response generation and could substantially impact user experience and impede the persuasion outcome. Besides, although reinforcement learning (RL) approaches have achieved big success in strategic tasks such as games, it requires a sophisticated user simulator to provide real-time feedback to the dialogue system, which limits the application of RL on persuasion dialogues. To address these issues towards a better persuasion dialogue system, we apply RL to refine a language model baseline without user simulators, and distill sentence-level information about repetition, inconsistency, and task relevance through rewards. Moreover, to better accomplish the persuasion task, the model learns from human demonstration to imitate human persuasion behavior and selects the most persuasive responses. Experiments show that our model outperforms previous state-of-the-art dialogue models on both automatic metrics and human evaluation results on a donation persuasion task, and generates more diverse, consistent and persuasive conversations according to the user feedback. We will make the code and model publicly available.

In natural language processing (NLP), state-of-the-art (SOTA) semi-supervised learning (SSL) frameworks have shown great performance on deep pre-trained language models such as BERT, and are expected to significantly reduce the demand for manual labeling. However, our empirical studies indicate that these frameworks are not suitable for lightweight models such as TextCNN, LSTM and etc. In this work, we develop a new SSL framework called FLiText, which stands for Faster and Lighter semi-supervised Text classification. FLiText introduces an inspirer network together with the consistency regularization framework, which leverages a generalized regular constraint on the lightweight models for efficient SSL. As a result, FLiText obtains new SOTA performance for lightweight models across multiple SSL benchmarks on text classification. Compared with existing SOTA SSL methods on TextCNN, FLiText improves the accuracy of lightweight model TextCNN from 51.00% to 90.49% on IMDb, 39.8% to 58.06% on Yelp-5, and from 55.3% to 65.08% on Yahoo! Answer. In addition, compared with the fully supervised method on the full dataset, FLiText just uses less than 1% of labeled data to improve the accuracy by 6.59%, 3.94%, and 3.22% on the datasets of IMDb, Yelp-5, and Yahoo! Answer respectively.

Linguistically informed analyses of language models (LMs) contribute to the understanding and improvement of such models. Here, we introduce the corpus of Chinese linguistic minimal pairs (CLiMP) to investigate what knowledge Chinese LMs acquire. CLiMP consists of sets of 1000 minimal pairs (MPs) for 16 syntactic contrasts in Chinese, covering 9 major Chinese linguistic phenomena. The MPs are semi-automatically generated, and human agreement with the labels in CLiMP is 95.8%. We evaluate 11 different LMs on CLiMP, covering n-grams, LSTMs, and Chinese BERT. We find that classifier–noun agreement and verb complement selection are the phenomena that models generally perform best at. However, models struggle the most with the ba construction, binding, and filler-gap dependencies. Overall, Chinese BERT achieves an 81.8% average accuracy, while the performances of LSTMs and 5-grams are only moderately above chance level.

Existing dialog system models require extensive human annotations and are difficult to generalize to different tasks. The recent success of large pre-trained language models such as BERT and GPT-2 (Devlin et al., 2019; Radford et al., 2019) have suggested the effectiveness of incorporating language priors in down-stream NLP tasks. However, how much pre-trained language models can help dialog response generation is still under exploration. In this paper, we propose a simple, general, and effective framework: Alternating Recurrent Dialog Model (ARDM). ARDM models each speaker separately and takes advantage of the large pre-trained language model. It requires no supervision from human annotations such as belief states or dialog acts to achieve effective conversations. ARDM outperforms or is on par with state-of-the-art methods on two popular task-oriented dialog datasets: CamRest676 and MultiWOZ. Moreover, we can generalize ARDM to more challenging, non-collaborative tasks such as persuasion. In persuasion tasks, ARDM is capable of generating human-like responses to persuade people to donate to a charity.

Humans are increasingly interacting with machines through language, sometimes in contexts where the user may not know they are talking to a machine (like over the phone or a text chatbot). We aim to understand how system designers and researchers might allow their systems to confirm its non-human identity. We collect over 2,500 phrasings related to the intent of “Are you a robot?”. This is paired with over 2,500 adversarially selected utterances where only confirming the system is non-human would be insufficient or disfluent. We compare classifiers to recognize the intent and discuss the precision/recall and model complexity tradeoffs. Such classifiers could be integrated into dialog systems to avoid undesired deception. We then explore how both a generative research model (Blender) as well as two deployed systems (Amazon Alexa, Google Assistant) handle this intent, finding that systems often fail to confirm their non-human identity. Finally, we try to understand what a good response to the intent would be, and conduct a user study to compare the important aspects when responding to this intent.

Emotional support is a crucial ability for many conversation scenarios, including social interactions, mental health support, and customer service chats. Following reasonable procedures and using various support skills can help to effectively provide support. However, due to the lack of a well-designed task and corpora of effective emotional support conversations, research on building emotional support into dialog systems remains lacking. In this paper, we define the Emotional Support Conversation (ESC) task and propose an ESC Framework, which is grounded on the Helping Skills Theory. We construct an Emotion Support Conversation dataset (ESConv) with rich annotation (especially support strategy) in a help-seeker and supporter mode. To ensure a corpus of high-quality conversations that provide examples of effective emotional support, we take extensive effort to design training tutorials for supporters and several mechanisms for quality control during data collection. Finally, we evaluate state-of-the-art dialog models with respect to the ability to provide emotional support. Our results show the importance of support strategies in providing effective emotional support and the utility of ESConv in training more emotional support systems.

We present LEGOEval, an open-source toolkit that enables researchers to easily evaluate dialogue systems in a few lines of code using the online crowdsource platform, Amazon Mechanical Turk. Compared to existing toolkits, LEGOEval features a flexible task design by providing a Python API that maps to commonly used React.js interface components. Researchers can personalize their evaluation procedures easily with our built-in pages as if playing with LEGO blocks. Thus, LEGOEval provides a fast, consistent method for reproducing human evaluation results. Besides the flexible task design, LEGOEval also offers an easy API to review collected data.

Suicide prevention hotline counselors aid individuals during difficult times through millions of calls and chats. A chatbot cannot safely replace a counselor, but we explore whether a chatbot can be developed to help train human counselors. Such a system needs to simulate intimate situations across multiple practice sessions. Open-domain dialogue systems frequently suffer from generic responses that do not characterize personal stories, so we look to infuse conversations with persona information by mimicking prototype conversations. Towards building a “Crisisbot” hotline visitor simulation, we propose a counseling strategy annotation scheme and a multi-task framework that leverages these counselor strategies to retrieve similar examples, generate diverse sub-utterances, and interleave prototype and generated sub-utterances into complex responses. We evaluate this framework with crowdworkers and experienced hotline counselors. The framework considerably increases response diversity and specificity, with limited impact to coherence. Our results also show a considerable discrepancy between crowdworker and counselor judgements, which emphasizes the importance of including target populations in system development and evaluation.

Image paragraph captioning (IPC) aims to generate a fine-grained paragraph to describe the visual content of an image. Significant progress has been made by deep neural networks, in which the attention mechanism plays an essential role. However, conventional attention mechanisms tend to ignore the past alignment information, which often results in problems of repetitive captioning and incomplete captioning. In this paper, we propose an Interactive key-value Memory- augmented Attention model for image Paragraph captioning (IMAP) to keep track of the attention history (salient objects coverage information) along with the update-chain of the decoder state and therefore avoid generating repetitive or incomplete image descriptions. In addition, we employ an adaptive attention mechanism to realize adaptive alignment from image regions to caption words, where an image region can be mapped to an arbitrary number of caption words while a caption word can also attend to an arbitrary number of image regions. Extensive experiments on a benchmark dataset (i.e., Stanford) demonstrate the effectiveness of our IMAP model.

面向知识库的自动问答是自然语言处理的一项重要任务,它旨在对用户提出的自然语言形式问题给出精炼、准确的回复。目前由于缺少数据集、特征不一致等因素,导致难以使用通用的数据和方法实现领域知识库问答。因此,本文将“问题意图”视作不同领域问答可能存在的共同特征,将“问题”与三元组知识库中“关系谓词”的映射过程作为问答核心工作。为了考虑多种层次的语义避免重要信息的损失,本文分别将“基于门控卷积的深层语义”和“基于交互注意力机制的浅层语义”两个方面通过门控感知机制相融合。我们在NLPCC-ICCPOL 2016 KBQA数据集上的实验表明,本文提出的方法与现有的基于CDSSM和BDSSM相比,效能有明显的提升。此外,本文通过构造天文常识知识库,将问题与关系谓词映射模型移植到特定领域,结合Bi-LSTM-CRF模型构建了天文常识自动问答系统。