The advancements of language language models (LLMs) have piqued growing interest in developing LLM-based language agents to automate scientific discovery end-to-end, which has sparked both excitement and skepticism about the true capabilities of such agents. In this work, we argue that for an agent to fully automate scientific discovery, it must be able to complete all essential tasks in the workflow. Thus, we call for rigorous assessment of agents on individual tasks in a scientific workflow before making bold claims on end-to-end automation. To this end, we present ScienceAgentBench, a new benchmark for evaluating language agents for data-driven scientific discovery. To ensure the scientific authenticity and real-world relevance of our benchmark, we extract 102 tasks from 44 peer-reviewed publications in four disciplines and engage nine subject matter experts to validate them. We unify the target output for every task to a self-contained Python program file and employ an array of evaluation metrics to examine the generated programs, execution results, and costs. Each task goes through multiple rounds of manual validation by annotators and subject matter experts to ensure its annotation quality and scientific plausibility. We also propose two effective strategies to mitigate data contamination concerns. Using our benchmark, we evaluate five open-weight and proprietary LLMs, each with three frameworks: direct prompting, OpenHands, and self-debug. Given three attempts for each task, the best-performing agent can only solve 32.4% of the tasks independently and 34.3% with expert-provided knowledge. These results underscore the limited capacities of current language agents in generating code for data-driven discovery, let alone end-to-end automation for scientific research.

Agentic search such as Deep Research systems-where agents autonomously browse the web, synthesize information, and return comprehensive citation-backed answers-represents a major shift in how users interact with web-scale information. While promising greater efficiency and cognitive offloading, the growing complexity and open-endedness of agentic search have outpaced existing evaluation benchmarks and methodologies, which largely assume short search horizons and static answers. In this paper, we introduce Mind2Web 2, a benchmark of 130 realistic, high-quality, and long-horizon tasks that require real-time web browsing and extensive information synthesis, constructed with over 1000 hours of human labor. To address the challenge of evaluating time-varying and complex answers, we propose a novel Agent-as-a-Judge framework. Our method constructs task-specific judge agents based on a tree-structured rubric design to automatically assess both answer correctness and source attribution. We conduct a comprehensive evaluation of ten frontier agentic search systems and human performance, along with a detailed error analysis to draw insights for future development. The best-performing system, OpenAI Deep Research, can already achieve 50-70% of human performance while spending half the time, highlighting its great potential. Altogether, Mind2Web 2 provides a rigorous foundation for developing and benchmarking the next generation of agentic search systems.

Human ability estimation is essential for educational assessment, career advancement, and professional certification. Adaptive Testing systems can improve estimation efficiency by selecting fewer, targeted questions, and are widely used in exams, e.g., GRE, GMAT, and Duolingo English Test. However, selecting an optimal subset of questions remains a challenging nested optimization problem. Existing methods rely on costly approximations or data-intensive training, making them unsuitable for today's large-scale and complex testing environments. Thus, we propose a Closed-Form solution for question subset selection in Adaptive Testing. It directly minimizes ability estimation error by reducing ability parameter's gradient bias while maintaining Hessian stability, which enables a simple greedy algorithm for question selection. Moreover, it can quantify the impact of human behavioral perturbations on ability estimation. Extensive experiments on large-scale educational datasets demonstrate that it reduces the number of required questions by 10% compared to SOTA methods, while maintaining the same estimation accuracy.

Huihan Li, Yuting Ning, Zeyi Liao, Siyuan Wang, Xiang Lorraine Li, Ximing Lu, Wenting Zhao, Faeze Brahman, Yejin Choi, Xiang Ren. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

In this paper, we introduce DeAR (_Decompose-Analyze-Rethink_), a framework that iteratively builds a reasoning tree to tackle intricate problems within a single large language model (LLM). Unlike approaches that extend or search for rationales, DeAR is featured by 1) adopting a tree-based question decomposition manner to plan the organization of rationales, which mimics the logical planning inherent in human cognition; 2) globally updating the rationales at each reasoning step through natural language feedback. Specifically, the _Decompose_ stage decomposes the question into simpler sub-questions, storing them as new nodes; the _Analyze_ stage generates and self-checks rationales for sub-questions at each node evel; and the _Rethink_ stage updates parent-node rationales based on feedback from their child nodes. By generating and updating the reasoning process from a more global perspective, DeAR constructs more adaptive and accurate logical structures for complex problems, facilitating timely error correction compared to rationale-extension and search-based approaches such as Tree-of-Thoughts (ToT) and Graph-of-Thoughts (GoT). We conduct extensive experiments on three reasoning benchmarks, including ScienceQA, StrategyQA, and GSM8K, which cover a variety of reasoning tasks, demonstrating that our approach significantly reduces logical errors and enhances performance across various LLMs. Furthermore, we validate that DeAR is an efficient method that achieves a superior trade-off between accuracy and reasoning time compared to ToT and GoT.

Understanding mathematical questions effectively is a crucial task, which can benefit many applications, such as difficulty estimation. Researchers have drawn much attention to designing pre-training models for question representations due to the scarcity of human annotations (e.g., labeling difficulty). However, unlike general free-format texts (e.g., user comments), mathematical questions are generally designed with explicit purposes and mathematical logic, and usually consist of more complex content, such as formulas, and related mathematical knowledge (e.g., Function). Therefore, the problem of holistically representing mathematical questions remains underexplored. To this end, in this paper, we propose a novel contrastive pre-training approach for mathematical question representations, namely QuesCo, which attempts to bring questions with more similar purposes closer. Specifically, we first design two-level question augmentations, including content-level and structure-level, which generate literally diverse question pairs with similar purposes. Then, to fully exploit hierarchical information of knowledge concepts, we propose a knowledge hierarchy-aware rank strategy (KHAR), which ranks the similarities between questions in a fine-grained manner. Next, we adopt a ranking contrastive learning task to optimize our model based on the augmented and ranked questions. We conduct extensive experiments on two real-world mathematical datasets. The experimental results demonstrate the effectiveness of our model.

Recently, graph-based anomaly detection (GAD) has attracted rising attention due to its effectiveness in identifying anomalies in relational and structured data. Unfortunately, the performance of most existing GAD methods suffers from the inherent structural noises of graphs induced by hidden anomalies connected with considerable benign nodes. In this work, we propose SparseGAD, a novel GAD framework that sparsifies the structures of target graphs to effectively reduce noises and collaboratively learns node representations. It then robustly detects anomalies by uncovering the underlying dependency among node pairs in terms of homophily and heterophily, two essential connection properties of GAD. Extensive experiments on real-world datasets of GAD demonstrate that the proposed framework achieves significantly better detection quality compared with the state-of-the-art methods, even when the graph is heavily attacked. Code will be available at https://github.com/KellyGong/SparseGAD.git.

In recent years, due to the explosive growth of patent applications, patent mining has drawn extensive attention and interest. An important issue of patent mining is that of recognizing the technologies contained in patents, which serves as a fundamental preparation for deeper analysis. To this end, in this article, we make a focused study on constructing a technology portrait for each patent, i.e., to recognize technical phrases concerned in it, which can summarize and represent patents from a technical perspective. Along this line, a critical challenge is how to analyze the unique characteristics of technical phrases and illustrate them with definite descriptions. Therefore, we first generate the detailed descriptions about the technical phrases existing in extensive patents based on different criteria, including various previous works, practical experience, and statistical analyses. Then, considering the unique characteristics of technical phrases and the complex structure of patent documents, such as multi-aspect semantics and multi-level relevances, we further propose a novel unsupervised model, namely TechPat, which can not only automatically recognize technical phrases from massive patents but also avoid the need for expensive human labeling. After that, we evaluate the extraction results from various aspects. Specifically, we propose a novel evaluation metric called Information Retrieval Efficiency (IRE) to quantify the performance of extracted technical phrases from a new perspective. Extensive experiments on real-world patent data demonstrate that the TechPat model can effectively discriminate technical phrases in patents and greatly outperform existing methods. We further apply extracted technical phrases to two practical application tasks, namely patent search and patent classification, where the experimental results confirm the wide application prospects of technical phrases. Finally, we discuss the generalization ability of our proposed methods.

The generation of large amounts of personal data provides data centers with sufficient resources to mine idiosyncrasy from private records. User modeling has long been a fundamental task with the goal of capturing the latent characteristics of users from their behaviors. However, centralized user modeling on collected data has raised concerns about the risk of data misuse and privacy leakage. As a result, federated user modeling has come into favor, since it expects to provide secure multi-client collaboration for user modeling through federated learning. Unfortunately, to the best of our knowledge, existing federated learning methods that ignore the inconsistency among clients cannot be applied directly to practical user modeling scenarios, and moreover, they meet the following critical challenges: 1) Statistical heterogeneity . The distributions of user data in different clients are not always independently identically distributed (IID), which leads to unique clients with needful personalized information; 2) Privacy heterogeneity . User data contains both public and private information, which have different levels of privacy, indicating that we should balance different information shared and protected; 3) Model heterogeneity . The local user models trained with client records are heterogeneous, and thus require a flexible aggregation in the server; 4) Quality heterogeneity . Low-quality information from inconsistent clients poisons the reliability of user models and offsets the benefit from high-quality ones, meaning that we should augment the high-quality information during the process. To address the challenges, in this paper, we first propose a novel client-server architecture framework, namely Hierarchical Personalized Federated Learning (HPFL), with a primary goal of serving federated learning for user modeling in inconsistent clients. More specifically, the client train and deliver the local user model via the hierarchical components containing hierarchical information from privacy heterogeneity to join collaboration in federated learning. Moreover, the client updates the personalized user model with a fine-grained personalized update strategy for statistical heterogeneity. Correspondingly, the server flexibly aggregates hierarchical components from heterogeneous user models in the case of privacy and model heterogeneity with a differentiated component aggregation strategy. In order to augment high-quality information and generate high-quality user models, we expand HPFL to the Augmented-HPFL (AHPFL) framework by incorporating the augmented mechanisms, which filters out low-quality information such as noise, sparse information and redundant information. Specially, we construct two implementations of AHPFL, i.e., AHPFL-SVD and AHPFL-AE, where the augmented mechanisms follow SVD (singular value decomposition) and AE (autoencoder), respectively. Finally, we conduct extensive experiments on real-world datasets, which demonstrate the effectiveness of both HPFL and AHPFL frameworks.

User modeling aims to capture the latent characteristics of users from their behaviors, and is widely applied in numerous applications. Usually, centralized user modeling suffers from the risk of privacy leakage. Instead, federated user modeling expects to provide a secure multi-client collaboration for user modeling through federated learning. Existing federated learning methods are mainly designed for consistent clients, which cannot be directly applied to practical scenarios, where different clients usually store inconsistent user data. Therefore, it is a crucial demand to design an appropriate federated solution that can better adapt to user modeling tasks, and however, meets following critical challenges: 1) Statistical heterogeneity. The distributions of user data in different clients are not always independently identically distributed which leads to personalized clients; 2) Privacy heterogeneity. User data contains both public and private information, which have different levels of privacy. It means we should balance different information to be shared and protected; 3) Model heterogeneity. The local user models trained with client records are heterogeneous which need flexible aggregation in the server. In this paper, we propose a novel client-server architecture framework, namely Hierarchical Personalized Federated Learning (HPFL) to serve federated learning in user modeling with inconsistent clients. In the framework, we first define hierarchical information to finely partition the data with privacy heterogeneity. On this basis, the client trains a user model which contains different components designed for hierarchical information. Moreover, client processes a fine-grained personalized update strategy to update personalized user model for statistical heterogeneity. Correspondingly, the server completes a differentiated component aggregation strategy to flexibly aggregate heterogeneous user models in the case of privacy and model heterogeneity. Finally, we conduct extensive experiments on real-world datasets, which demonstrate the effectiveness of the HPFL framework.