As large language models move from isolated task solving toward long-term service in human environments, they require social intelligence: the ability to infer mental states, track social relations, reason over norms, and adapt behavior under context. This report presents ZenGen, an integrated framework for measuring, internalizing, and grounding social intelligence. For measurement, we introduce SoMBench, a psychology-grounded benchmark spanning 3 primary dimensions, 17 secondary dimensions, and 71 task paradigms. It controls question format, narrative perspective, and context length across 284 shared scenarios and 3,481 expert-verified instances. Evaluation of 20 representative LLMs reveals substantial headroom: the best model achieves only 72.08

In recent years, research in the field of social bot detection has gradually evolved from individual feature analysis to group feature mining, and from traditional feature engineering to deep learning methods. Among them, graph network-based methods have shown significant advantages: they can integrate account behavior features, text semantic features, and network topology features, converting social bot detection into a graph node classification task. However, most existing detection methods adopt general models for detection and fail to consider the differences in fine-grainedfeatures among different types of social bots, which limits the detection accuracy across business scenarios. Based on this, a social bot detection method based on a fine-grained feature expert attention mechanism was proposed. The method constructed a business expert network, enabling each expert to focus on learning differentiated weight combinations of fine-grained features. Through multi-expert feature fusion and comprehensive analysis, it achieved integrated detection of potentially multi-type social bots across various business scenarios. Experimental results on a public Twitter dataset for social bot detection demonstrated that this method outperformed existing mainstream detection methods, with a relative improvement of 1.52% in the F1-score.

Retrieval-Augmented Language Models boost task performance, owing to the retriever that provides external knowledge. Although crucial, the retriever primarily focuses on semantics relevance, which may not always be effective for generation. Thus, utility-based retrieval has emerged as a promising topic, prioritizing passages that provide valid benefits for downstream tasks. However, due to insufficient understanding, capturing passage utility accurately remains unexplored. This work proposes SCARLet, a framework for training utility-based retrievers in RALMs, which incorporates two key factors, multi-task generalization and inter-passage interaction. First, SCARLet constructs shared context on which training data for various tasks is synthesized. This mitigates semantic bias from context differences, allowing retrievers to focus on learning task-specific utility and generalize across tasks. Next, SCARLet uses a perturbation-based attribution method to estimate passage-level utility for shared context, which reflects interactions between passages and provides more accurate feedback. We evaluate our approach on ten datasets across various tasks, both in-domain and out-of-domain, showing that retrievers trained by SCARLet consistently improve the overall performance of RALMs.

Large Language Models (LLMs) can enhance the credibility and verifiability by generating text with citations. However, existing tasks and evaluation methods are predominantly limited to sentence-level statement, neglecting the significance of positional fine-grained citations that can appear anywhere within sentences. To facilitate further exploration of the fine-grained citation generation, we propose ALiiCE, the first automatic evaluation framework for this task. Our framework first parses the sentence claim into atomic claims via dependency analysis and then calculates citation quality at the atomic claim level. ALiiCE introduces three novel metrics for positional fined-grained citation quality assessment, including positional fine-grained citation recall and precision, and coefficient of variation of citation positions. We evaluate the positional fine-grained citation generation performance of several LLMs on two long-form QA datasets. Our experiments and analyses demonstrate the effectiveness and reasonableness of ALiiCE. The results also indicate that existing LLMs still struggle to provide positional fine-grained citations.

Agentic search, as a more autonomous and adaptive paradigm of retrieval augmentation, is driving the evolution of intelligent search systems. However, existing evaluation frameworks fail to align well with the goals of agentic search. First, the complex queries commonly used in current benchmarks often deviate from realistic user search scenarios. Second, prior approaches tend to introduce noise when extracting ground truth for end-to-end evaluations, leading to distorted assessments at a fine-grained level. Third, most current frameworks focus solely on the quality of final answers, neglecting the evaluation of the iterative process inherent to agentic search. To address these limitations, we propose RAVine – a Reality-Aligned eValuation framework for agentic LLMs with search. RAVine targets multi-point queries and long-form answers that better reflect user intents, and introduces an attributable ground truth construction strategy to enhance the accuracy of fine-grained evaluation. Moreover, RAVine examines model's interaction with search tools throughout the iterative process, and accounts for factors of efficiency. We benchmark a series of models using RAVine and derive several insights, which we hope will contribute to advancing the development of agentic search systems. The code and datasets are available at https://github.com/SwordFaith/RAVine.

Traditional knowledge graph embedding (KGE) methods typically require preserving the entire knowledge graph (KG) with significant training costs when new knowledge emerges. To address this issue, the continual knowledge graph embedding (CKGE) task has been proposed to train the KGE model by learning emerging knowledge efficiently while simultaneously preserving decent old knowledge. However, the explicit graph structure in KGs, which is critical for the above goal, has been heavily ignored by existing CKGE methods. On the one hand, existing methods usually learn new triples in a random order, destroying the inner structure of new KGs. On the other hand, old triples are preserved with equal priority, failing to alleviate catastrophic forgetting effectively. In this paper, we propose a competitive method for CKGE based on incremental distillation (IncDE), which considers the full use of the explicit graph structure in KGs. First, to optimize the learning order, we introduce a hierarchical strategy, ranking new triples for layer-by-layer learning. By employing the inter- and intra-hierarchical orders together, new triples are grouped into layers based on the graph structure features. Secondly, to preserve the old knowledge effectively, we devise a novel incremental distillation mechanism, which facilitates the seamless transfer of entity representations from the previous layer to the next one, promoting old knowledge preservation. Finally, we adopt a two-stage training paradigm to avoid the over-corruption of old knowledge influenced by under-trained new knowledge. Experimental results demonstrate the superiority of IncDE over state-of-the-art baselines. Notably, the incremental distillation mechanism contributes to improvements of 0.2%-6.5% in the mean reciprocal rank (MRR) score. More exploratory experiments validate the effectiveness of IncDE in proficiently learning new knowledge while preserving old knowledge across all time steps.

Continual Knowledge Graph Embedding (CKGE) aims to efficiently learn new knowledge and simultaneously preserve old knowledge. Dominant approaches primarily focus on alleviating catastrophic forgetting of old knowledge but neglect efficient learning for the emergence of new knowledge. However, in real-world scenarios, knowledge graphs (KGs) are continuously growing, which brings a significant challenge to fine-tuning KGE models efficiently. To address this issue, we propose a fast CKGE framework (FastKGE), incorporating an incremental low-rank adapter (IncLoRA) mechanism to efficiently acquire new knowledge while preserving old knowledge. Specifically, to mitigate catastrophic forgetting, FastKGE isolates and allocates new knowledge to specific layers based on the fine-grained influence between old and new KGs. Subsequently, to accelerate fine-tuning, FastKGE devises an efficient IncLoRA mechanism, which embeds the specific layers into incremental low-rank adapters with fewer training parameters. Moreover, IncLoRA introduces adaptive rank allocation, which makes the LoRA aware of the importance of entities and adjusts its rank scale adaptively. We conduct experiments on four public datasets and two new datasets with a larger initial scale. Experimental results demonstrate that FastKGE can reduce training time by 34%-49% while still achieving competitive link prediction performance against state-of-the-art models on four public datasets (average MRR score of 21.0% vs. 21.1%). Meanwhile, on two newly constructed datasets, FastKGE saves 51%-68% training time and improves link prediction performance by 1.5%.

Named entity recognition (NER) aims to locate and classify named entities in natural language texts. Most existing high-performance NER models employ a supervised paradigm, which requires a large quantity of high-quality annotated data during training. In order to help NER models perform well in few-shot scenarios, data augmentation approaches attempt to build extra data by means of random editing or by using end-to-end generation with PLMs. However, these methods focus on only the fluency of generated sentences, ignoring the syntactic correlation between the new and raw sentences. Such uncorrelation also brings low diversity and inconsistent labeling of synthetic samples. To fill this gap, we present SAINT (Syntactic-Aware InsertioN Transformer), a hard-constraint controlled text generation model that incorporates syntactic information. The proposed method operates by inserting new tokens between existing entities in a parallel manner. During insertion procedure, new tokens will be added taking both semantic and syntactic factors into account. Hence the resulting sentence can retain the syntactic correctness with respect to the raw data. Experimental results on two benchmark datasets, i.e., Ontonotes and Wikiann, demonstrate the comparable performance of SAINT over the state-of-the-art baselines.

Stance detection is to distinguish whether the text's author supports, opposes, or maintains a neutral stance towards a given target. In most real-world scenarios, stance detection needs to work in a zero-shot manner, i.e., predicting stances for unseen targets without labeled data. One critical challenge of zero-shot stance detection is the absence of contextual information on the targets. Current works mostly concentrate on introducing external knowledge to supplement information about targets, but the noisy schema-linking process hinders their performance in practice. To combat this issue, we argue that previous studies have ignored the extensive target -related information inhabited in the unlabeled data during the training phase, and propose a simple yet efficient Multi-Perspective Contrastive Learning Framework for zero-shot stance detection. Our framework is capable of leveraging information not only from labeled data but also from extensive unlabeled data. To this end, we design target-oriented contrastive learning and label-oriented contrastive learning to capture more comprehensive target representation and more distinguishable stance features. We conduct extensive experiments on three widely adopted datasets (from 4870 to 33,090 instances), namely SemEval-2016, WT-WT, and VAST. Our framework achieves 53.6%, 77.1%, and 72.4% macro-average F1 scores on these three datasets, showing 2.71% and 0.25% improvements over state-of-the-art baselines on the SemEval-2016 and WT-WT datasets and comparable results on the more challenging VAST dataset.

Unsupervised opinion summarization techniques are designed to condense the review data and summarize informative and salient opinions in the absence of golden references. Existing dominant methods generally follow a two-stage framework: first creating the synthetic "review-summary" paired datasets and then feeding them into the generative summary model for supervised training. However, these methods mainly focus on semantic similarity in synthetic dataset creation, ignoring the consistency of aspects and sentiments in synthetic pairs. Such inconsistency also brings a gap to the training and inference of the summarization model.

Nowadays, a large amount of information is stored as text, and numerous text mining techniques have been developed for various applications, such as event detection, news topic classification, public opinion detection, and sentiment analysis. Although significant progress has been achieved for short text classification, document-level text classification requires further exploration. Long documents always contain irrelevant noisy information that shelters the prominence of indicative features, limiting the interpretability of classification results. To alleviate this problem, a model called MIPELD (mining the frequent pattern of a named entity for long document classification) for long document classification is demonstrated, which mines the frequent patterns of named entities as features. Discovered patterns allow semantic generalization among documents and provide clues for verifying the results. Experiments on several datasets resulted in good accuracy and marco-F1 values, meeting the requirements for practical application. Further analysis validated the effectiveness of MIPELD in mining interpretable information in text classification.

Unsupervised opinion summarization techniques are designed to condense the review data and summarize informative and salient opinions in the absence of golden references. Existing dominant methods generally follow a two-stage framework: first creating the synthetic "review-summary" paired datasets and then feeding them into the generative summary model for supervised training. However, these methods mainly focus on semantic similarity in synthetic dataset creation, ignoring the consistency of aspects and sentiments in synthetic pairs. Such inconsistency also brings a gap to the training and inference of the summarization model. To alleviate this problem, we propose ConsistSum, an unsupervised opinion summarization method devoting to capture the consistency of aspects and sentiment between reviews and summaries. Specifically, ConsistSum first extracts the preliminary "review-summary" pairs from the raw corpus by evaluating the distance of aspect distribution and sentiment distribution. Then, we refine the preliminary summary with the constrained Metropolis-Hastings sampling to produce highly consistent synthetic datasets. In the summarization phase, we adopt the generative model T5 as the summarization model. T5 is fine-tuned for the opinion summarization task by incorporating the loss of predicting aspect and opinion distribution. Experimental results on two benchmark datasets, $i.e.$, Yelp and Amazon, demonstrate the superior performance of ConsistSum over the state-of-the-art baselines.

Predicting the popularity of online content is a fundamental problem in various applications. One practical challenge takes roots in the varying length of observation time or prediction horizon, i.e., a good model for popularity prediction is desired to handle various prediction settings. However, most existing methods adopt a separate training paradigm for each prediction setting and the obtained model for one setting is difficult to be generalized to others, causing a great waste of computational resources and a large demand for downstream labels. To solve the above issues, we propose a novel pre-training framework for popularity prediction, namely PREP, aiming to pre-train a general representation model from the readily available unlabeled diffusion data, which can be effectively transferred into various prediction settings. We design a novel pretext task for pre-training, i.e., temporal elapse inference for two randomly sampled time slices of popularity dynamics, impelling the representation model to learn intrinsic knowledge about popularity dynamics. Experimental results conducted on two real datasets demonstrate the generalization and efficiency of the pre-training framework for different popularity prediction task settings.

Stance detection aims to identify whether the author of a text is in favor of, against, or neutral to a given target. The main challenge of this task comes two-fold: few-shot learning resulting from the varying targets and the lack of contextual information of the targets. Existing works mainly focus on solving the second issue by designing attention-based models or introducing noisy external knowledge, while the first issue remains under-explored. In this paper, inspired by the potential capability of pre-trained language models (PLMs) serving as knowledge bases and few-shot learners, we propose to introduce prompt-based fine-tuning for stance detection. PLMs can provide essential contextual information for the targets and enable few-shot learning via prompts. Considering the crucial role of the target in stance detection task, we design target-aware prompts and propose a novel verbalizer. Instead of mapping each label to a concrete word, our verbalizer maps each label to a vector and picks the label that best captures the correlation between the stance and the target. Moreover, to alleviate the possible defect of dealing with varying targets with a single hand-crafted prompt, we propose to distill the information learned from multiple prompts. Experimental results show the superior performance of our proposed model in both full-data and few-shot scenarios.

Understanding the emergence of universal features such as the stylized facts in markets is a long-standing challenge that has drawn much attention from economists and physicists. Most existing models, such as stochastic volatility models, focus mainly on price changes, neglecting the complex trading dynamics. Recently, there are increasing studies on order books, thanks to the availability of large-scale trading datasets, aiming to understand the underlying mechanisms governing the market dynamics. In this paper, we collect order-book datasets of Bitcoin platforms across three countries over millions of users and billions of daily turnovers. We find a 1+1D field theory, govern by a set of KPZ-like stochastic equations, predicts precisely the order book dynamics observed in empirical data. Despite the microscopic difference of markets, we argue the proposed effective field theory captures the correct universality class of market dynamics. We also show that the model agrees with the existing stochastic volatility models at the long-wavelength limit.

The key challenge of cross-domain context-dependent text-to-SQL generation tasks lies in capturing the relation of natural language utterance and SQL queries in different turns. A line of works attempt to combat this challenge by capturing the overlaps among consecutively generated SQL queries. Existing models sequentially generate the SQL query for a single turn and model the SQL overlaps via copying tokens or segments generated in previous turns. However, they are not flexible enough to capture various overlapping granularities, e.g., columns, filters, or even the whole query, as they neglect the intrinsic structures inhabited in SQL queries. In this paper, we employ tree-structured intermediate representations of SQL queries, i.e., SemQL, for SQL generation and propose a novel subtree-copy mechanism to characterize the SQL overlaps. At each turn, we encode the interaction questions and previously generated trees as context and decode the SemQL tree in a top-down fashion. Each node is either generated according to SemQL grammar or copied from previously generated SemQL subtrees. Our model can capture various overlapping granularities by copying nodes at different levels of SemQL trees. We evaluate our approach on the SParC dataset and the experimental results show the superior performance of our model compared with state-of-the-art baselines.

语义解析的目标是将自然语言表达映射为机器可理解的逻辑表达,该任务的关键挑战在于难以刻画自然语言中蕴含的组合语义。目前,结合深度神经网络模型的语义解析方法已经成为该领域的主流方法,该类方法通常采用编码器—解码器框架,通过设计树形结构的解码器或者在解码器中添加语法限制,从语法层面上提升逻辑表达生成的准确率。与现有的神经语义解析方法不同,该文从语义建模角度出发,以语义框架作为中间形式,通过自顶向下的生成方式,显式地建模自然语言表达中蕴含的层次化语义结构。模型先根据自然语言输入,自顶向下地生成语义框架,再将语义框架表示融入到逻辑表达的生成过程中。三个数据集上的实验结果表明,该文提出的模型能更准确地生成语义框架,并且在语义解析任务中取得更好的效果。

Predicting the popularity of online content is a fundamental problem in various application areas. One practical challenge for popularity prediction takes roots in the different settings of popularity prediction tasks in different situations, e.g., the varying lengths of the observation time window or prediction horizon. In other words, a good model for popularity prediction is desired to handle various tasks with different settings. However, the conventional paradigm for popularity prediction is training a separate prediction model for each prediction task, and thus the obtained model for one task is difficult to be generalized to other tasks, causing a great waste of training time and computational resources. To solve this issue, in this paper, we propose a novel pre-training framework for popularity prediction, aiming to pre-train a general deep representation model by learning intrinsic knowledge about popularity dynamics from the readily available diffusion cascades. We design a novel pretext task for pre-training, i.e., temporal context prediction for two randomly sampled time slices of popularity dynamics, impelling the deep prediction model to effectively capture the characteristics of popularity dynamics. Taking the state-of-the-art deep model, i.e., temporal convolutional neural network, as an instantiation of our proposed framework, experimental results conducted on both Sina Weibo and Twitter datasets demonstrate both the effectiveness and efficiency of the proposed pre-training framework for multiple popularity prediction tasks.

在线社交网络中的消息流行度预测研究,对推荐、广告、检索等应用场景都具有非常重要的作用。近年来,深度学习的蓬勃发展和消息传播数据的积累,为基于深度学习的流行度预测研究提供了坚实的发展基础。现有的流行度预测研究综述,主要是围绕传统的流行度预测方法展开的,而基于深度学习的流行度预测方法目前仍未得到系统性地归纳和梳理,不利于流行度预测领域的持续发展。鉴于此,该文重点论述和分析现有的基于深度学习的流行度预测相关研究,对近年来基于深度学习的流行度预测研究进行了归纳梳理,将其分为基于深度表示和基于深度融合的流行度预测方法,并对该研究方向的发展现状和未来趋势进行了分析展望。

Risk prediction is one of the important issues that draws much attention from academia and industry. And the fluctuation—absolute value of the change of price, is one of the indexes of risk. In this paper, we focus on the relationship between fluctuation and order volume. Based on the observation that the price would move when the volume of order changes, the prediction of price fluctuation can be converted into the prediction of order volume. Modelling the trader’s behaviours—order placement and order cancellation, we propose an order-based fluctuation prediction model. And our model outperforms better than baseline in OKCoin and BTC-e datasets.