Memory is essential for enabling large language model (LLM) agents to handle long-horizon reasoning tasks. Existing memory mechanisms are largely centralized, typically organizing retrieved information and interaction history within a single model context. This design imposes a fundamental trade-off: scaling reasoning trajectories risks context overload, whereas aggressive content pruning may result in irreversible information loss. Seeking a better trade-off, we draw inspiration from human cognitive systems, especially the functional complementarity between the prefrontal cortex (executive control) and the hippocampus (memory management), suggesting that such a trade-off need not be inherent, but may instead stem from centralized memory organization. To this end, we propose ActiveMem, a heterogeneous framework that decouples agent memory from the core reasoning process. Specifically, a high-level Planner utilizes distilled semantic gists to execute reasoning, while a lightweight, distributed memory system operates in parallel to actively accumulate and consolidate these gists throughout the task. Experiments on BrowseComp-Plus and GAIA show that ActiveMem achieves state-of-the-art accuracy with significantly reduced overhead, demonstrating the effectiveness of distributed active memory for long-horizon reasoning.

Existing reinforcement learning methods for LLM reasoning implicitly assume that the policy generating training trajectories should coincide with the one producing inference responses. We argue that this is a misleading inductive bias: the optimization-optimal trajectory distribution favors informative gradients, whereas the inference-optimal response distribution emphasizes accuracy and consistency. Forcing both into a single policy entangles their gradients and suppresses exploration. We propose R^2PO (Residual Rollout Policy Optimization), which attaches a lightweight Residual Rollout-Head atop the policy to decouple training trajectories from inference responses, diversifying rollouts during training while keeping inference generation intact. Experiments show that R^2PO consistently outperforms baselines, with average accuracy gains of 3.4

Company financial risk is widespread, and accurate prediction is critical to avoiding significant losses. Many risky companies often exhibit subtle anomalies, incomplete information, or limited interactions with others; however, the types of their interactions remains diverse and informative. Existing methods like metapath-based Graph Neural Networks effectively leverage node relationships but are constrained by manual biases, noise introduction and high computational complexity. Similarly, Graph Transformers show strong performance but suffer from prohibitively high complexity. To overcome these challenges, we propose the Graph Neural Network with Topology-Aware Augmentation (GTA). GTA adopts a dual augmentation strategy based on topology information, augmenting both topology and attributes. It first performs unification encoding on single node-type heterogeneous graphs, integrating heterogeneous topology into node representations. Expressive topology encoding is performed, followed by dual augmentation based on the learned topology embeddings. Through this approach, GTA achieves effective risk prediction. Extensive experiments on the real-world dataset demonstrate GTA's superior performance compared to state-of-the-art metapath-based and graph transformer-based methods, effectively handling sparse graphs with a single node-type and multiple edge-types. Comprehensive ablation study and visual analysis further validate the discriminative power of topology augmentation in distinguishing risky companies. And our code is publicly available at https://github.com/ckz123/GTA.

With frequent defaults occurring in the bond and financial markets in recent years, the prediction of corporate credit default risk has become increasingly crucial. However, traditional methods have primarily focused on forecasting the default probability of individual bond-issuing entities, while neglecting the interconnected structure formed by complex interentity relationships and the resulting credit risk contagion effects. This oversight has compromised the timeliness and accuracy of credit risk early warning systems. To address this issue, this study proposes a novel credit risk prediction method that integrates large language models (LLMs) with graph learning. Specifically, LLMs are employed to extract entity-interconnected networks from massive volumes of text data. Subsequently, graph attention neural networks (GATs) are utilized to further enhance the relationship density of these interconnected networks, enabling effective modeling of the transmission network of credit risk among entities. Additionally, news sentiment analysis is leveraged to calculate the credit risk score of individual entities. Through a message-passing mechanism, the comprehensive credit risk score-incorporating both the entity's own risk characteristics and the influence from the interconnected network-is computed, thereby improving the accuracy of credit risk prediction. Experimental analyses demonstrate that, compared with credit risk prediction methods focusing solely on individual entities, the entity-interconnected network construction method proposed in this study increases the predicted AUC metric by $\mathbf{4. 4 \%}$.

Recent prevailing works on graph machine learning typically follow a similar methodology that involves designing advanced variants of graph neural networks (GNNs) to maintain the superior performance of GNNs on different graphs. In this paper, we aim to streamline the GNN design process and leverage the advantages of Large Language Models (LLMs) to improve the performance of GNNs on downstream tasks. We formulate a new paradigm, coined "LLMs-as-Consultants", which integrates LLMs with GNNs in an interactive manner. A framework named LOGIN (LLM cOnsulted GNN traINing) is instantiated, empowering the interactive utilization of LLMs within the GNN training process. First, we attentively craft concise prompts for spotted nodes, carrying comprehensive semantic and topological information, and serving as input to LLMs. Second, we refine GNNs by devising a complementary coping mechanism that utilizes the responses from LLMs, depending on their correctness. We empirically evaluate the effectiveness of LOGIN on node classification tasks across both homophilic and heterophilic graphs. The results illustrate that even basic GNN architectures, when employed within the proposed LLMs-as-Consultants paradigm, can achieve comparable performance to advanced GNNs with intricate designs. Our code is available at https://github.com/QiaoYRan/LOGIN.

Graphs consisting of vocal nodes ("the vocal minority") and silent nodes ("the silent majority"), namely VS-Graph, are ubiquitous in the real world. The vocal nodes tend to have abundant features and labels. In contrast, silent nodes only have incomplete features and rare labels, e.g., the description and political tendency of politicians (vocal) are abundant while not for ordinary civilians (silent) on the twitter’s social network. Predicting the silent majority remains a crucial yet challenging problem. However, most existing Graph Neural Networks (GNNs) assume that all nodes belong to the same domain, without considering the missing features and distribution-shift between domains, leading to poor ability to deal with VS-Graph. To combat the above challenges, we propose Knowledge Transferable Graph Neural Network (KTGNN), which models distribution-shifts during message passing and learns representation by transferring knowledge from vocal nodes to silent nodes. Specifically, we design the domain-adapted "feature completion and message passing mechanism" for node representation learning while preserving domain difference. And a knowledge transferable classifier based on KL-divergence is followed. Comprehensive experiments on real-world scenarios (i.e., company financial risk assessment and political elections) demonstrate the superior performance of our method. Our source code has been open-sourced1.

The data-hungry problem, characterized by insufficiency and low-quality of data, poses obstacles for deep learning models. Transfer learning has been a feasible way to transfer knowledge from high-quality external data of source domains to limited data of target domains, which follows a domain-level knowledge transfer to learn a shared posterior distribution. However, they are usually built on strong assumptions, e.g., the domain invariant posterior distribution, which is usually unsatisfied and may introduce noises, resulting in poor generalization ability on target domains. Inspired by Graph Neural Networks (GNNs) that aggregate information from neighboring nodes, we redefine the paradigm as learning a knowledge-enhanced posterior distribution for target domains, namely Knowledge Bridge Learning (KBL). KBL first learns the scope of knowledge transfer by constructing a Bridged-Graph that connects knowledgeable samples to each target sample and then performs sample-wise knowledge transfer via GNNs.KBL is free from strong assumptions and is robust to noises in the source data. Guided by KBL, we propose the Bridged-GNN including an Adaptive Knowledge Retrieval module to build Bridged-Graph and a Graph Knowledge Transfer module. Comprehensive experiments on both un-relational and relational data-hungry scenarios demonstrate the significant improvements of Bridged-GNN compared with SOTA methods

In anurans, calling behaviour is strongly seasonal and circadian. Previous studies have revealed that auditory sensitivity in frogs exhibits seasonal plasticity, and electroencephalographic signals exhibit highly correlated circadian patterns; of which, the circadian rhythm remains unknown. In this study, the circadian rhythm and intersexual differences of auditory sensitivity were tested in the Emei music frog (Nidirana daunchina). This was achieved by comparing thresholds and latencies of auditory brainstem responses (ABRs) evoked by tones and clicks stimuli between male and female frogs during the day and at night, respectively. Our results revealed that both auditory thresholds and latencies had no differences between day and night except the latencies in 3.5-4.0 kHz frequencies. However, the thresholds of tone pip evoked ABRs differed significantly between male and female frogs from 25 to 5.0 kHz. This demonstrated that the auditory sensitivity of Emei music frogs exhibits sexual dimorphism at high frequencies, with female frogs exhibiting greater auditory sensitivity than that of male frogs. Simultaneously, the power spectra of male advertisement calls are matched well with the frequency range of auditory sensitivity in male and female frogs, which supports the matched filter hypothesis. Our study enhances the understanding of circadian plasticity and sexual dimorphism of auditory sensitivity in frogs.

Company financial risk is ubiquitous and early risk assessment for listed companies can avoid considerable losses. Traditional methods mainly focus on the financial statements of companies and lack the complex relationships among them. However, the financial statements are often biased and lagged, making it difficult to identify risks accurately and timely. To address the challenges, we redefine the problem as company financial risk assessment on tribe-style graph by taking each listed company and its shareholders as a tribe and leveraging financial news to build inter-tribe connections. Such tribe-style graphs present different patterns to distinguish risky companies from normal ones. However, most nodes in the tribe-style graph lack attributes, making it difficult to directly adopt existing graph learning methods (e.g., Graph Neural Networks(GNNs)). In this paper, we propose a novel Hierarchical Graph Neural Network (TH-GNN) for Tribe-style graphs via two levels, with the first level to encode the structure pattern of the tribes with contrastive learning, and the second level to diffuse information based on the inter-tribe relations, achieving effective and efficient risk assessment. Extensive experiments on the real-world company dataset show that our method achieves significant improvements on financial risk assessment over previous competing methods. Also, the extensive ablation studies and visualization comprehensively show the effectiveness of our method.

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.

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。

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.

The Bitcoin market becomes the focus of the economic market since its birth, and it has attracted wide attention from both academia and industry. Due to the absence of regulations in the Bitcoin market, it may be easier to bring some kinds of illegal behaviors. Thus, it raises an interesting question: Is there abnormity or illegal behavior in Bitcoin platforms? To answer this question, we investigate the abnormity in five leading Bitcoin platforms. By analyzing the financial index, i.e. the normalized logarithmic price return, we find that the properties of price return in bitFlyer are completely different from others. To find the possible reasons, we find that the abnormal ask price and bid price appear simultaneously in bitFlyer, which may be potentially linked to either price manipulation or money laundering. It verifies our conjecture that there may be abnormity or price manipulation in Bitcoin platforms. Furthermore, our findings in price return could also provide an innovative and effective method to detect the abnormity in Bitcoin platforms.

Market manipulation is one of the important issues that draw much attention from academia and industry. Many efforts have been made to detect manipulation in stock market. However, with the development of technology, the means of manipulation become more and more diversified and the effective detection methods remain to be an open problem. Here, we develop a generalized method for colluded traders detection based on transaction data. We investigate the clique of trading network, and find the number and weight of clique are greater in manipulated stocks than that in non-manipulated stocks. We further propose a method to detect colluded traders based on weight of cliques. Results demonstrate that our method is effective at distinguishing the manipulated stocks and the colluded traders.

Noise is one of the main factors that can influence the processes of sound communication across a wide range of animal groups. Although the effects of ambient noise on animal communication, including anthropogenic noise, have received increasing attention, few studies have examined changes in the fine structure of acoustic signals produced by vocalizing species in constantly noisy environments. Here, we used natural recordings to determine the associations between stream noise and call parameters in the little torrent frog (Amolops torrentis). We also used playbacks of stream noise recorded in natural habitats and playbacks of white noise to examine how male vocal signals change with increasing noise levels. The results show that noise intensity has a significant effect on male call frequency, but not on call amplitude or other call characteristics. Based on this evidence, we suggest that in streamside species stream noise drives males to alter call frequency and call as loudly as possible in order to improve discriminability. These findings provide insights into the role played by ecological selection in the evolution of noise-dependent anuran vocal plasticity.

Manipulation is an important issue for both developed and emerging stock markets. Many efforts have been made to detect manipulation in stock market. However, it is still an open problem to identify the fraudulent traders, especially when they collude with each other. In this paper, we focus on the problem of identifying anomalous traders using the transaction data of 8 manipulated stocks and 42 non-manipulated stocks during a one-year period. For each stock, we construct a multi-slice trading network to characterize the daily trading behavior and the cross-day participation of each trader. Comparing the multi-slice trading network of manipulated stocks and non-manipulated stocks with their randomized version, we find that manipulated stocks exhibit high number of trader pairs that trade with each other in multiple days and high deviation from randomized network at correlation between trading frequency and trading activity. These findings are effective at distinguishing manipulated stocks from non-manipulated ones and at identifying anomalous traders.

Stock price prediction is an important and challenging problem in stock market analysis. Existing prediction methods either exploit autocorrelation of stock price and its correlation with the supply and demand of stock, or explore predictive indictors exogenous to stock market. In this paper, using transaction record of stocks with identifier of traders, we introduce an index to characterize market confidence, i.e., the ratio of the number of traders who is active in two successive trading days to the number of active traders in a certain trading day. Strong Granger causality is found between the index of market confidence and stock price. We further predict stock price by incorporating the index of market confidence into a neural network based on time series of stock price. Experimental results on 50 stocks in two Chinese Stock Exchanges demonstrate that the accuracy of stock price prediction is significantly improved by the inclusion of the market confidence index. This study sheds light on using cross-day trading behavior to characterize market confidence and to predict stock price.

Stock price prediction is an important and challenging problem for studying financial markets. Existing studies are mainly based on the time series of stock price or the operation performance of listed company. In this paper, we propose to predict stock price based on investors' trading behavior. For each stock, we characterize the daily trading relationship among its investors using a trading network. We then classify the nodes of trading network into three roles according to their connectivity pattern. Strong Granger causality is found between stock price and trading relationship indices, i.e., the fraction of trading relationship among nodes with different roles. We further predict stock price by incorporating these trading relationship indices into a neural network based on time series of stock price. Experimental results on 51 stocks in two Chinese Stock Exchanges demonstrate the accuracy of stock price prediction is significantly improved by the inclusion of trading relationship indices.

Manipulation is an important issue for both developed and emerging stock markets. Many efforts have been made to detect manipulation in stock markets. However, it is still an open problem to identify the fraudulent traders, especially when they collude with each other. In this paper, we focus on the problem of identifying the anomalous traders using the transaction data of eight manipulated stocks and forty-four non-manipulated stocks during a one-year period. By analyzing the trading networks of stocks, we find that the trading networks of manipulated stocks exhibit significantly higher degree-strength correlation than the trading networks of non-manipulated stocks and the randomized trading networks. We further propose a method to detect anomalous traders of manipulated stocks based on statistical significance analysis of degree-strength correlation. Experimental results demonstrate that our method is effective at distinguishing the manipulated stocks from non-manipulated ones. Our method outperforms the traditional weight-threshold method at identifying the anomalous traders in manipulated stocks. More importantly, our method is difficult to be fooled by colluded traders.

Manipulation is an important issue for both developed and emerging stock markets. For the study of manipulation, it is critical to analyze investor behavior in the stock market. In this paper, an analysis of the full transaction records of over a hundred stocks in a one-year period is conducted. For each stock, a trading network is constructed to characterize the relations among its investors. In trading networks, nodes represent investors and a directed link connects a stock seller to a buyer with the total trade size as the weight of the link, and the node strength is the sum of all edge weights of a node. For all these trading networks, we find that the node degree and node strength both have tails following a power-law distribution. Compared with non-manipulated stocks, manipulated stocks have a high lower bound of the power-law tail, a high average degree of the trading network and a low correlation between the price return and the seller–buyer ratio. These findings may help us to detect manipulated stocks.