Zheng Zhang, Qi Liu, Siyuan Liang, Ning Li, Zirui Hu, Weibo Gao, Rui Li, Zhenya Huang, Leszek Rutkowski, Baosheng Yu, Dacheng Tao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Computerized Adaptive Testing (CAT) offers an efficient and personalized method for assessing examinee proficiency by dynamically adjusting test questions based on individual performance. Compared to traditional, non-personalized testing methods, CAT requires fewer questions and provides more accurate assessments. As a result, CAT has been widely adopted across various fields, including education, healthcare, sports, sociology, and the evaluation of AI models. While traditional methods rely on psychometrics and statistics, the increasing complexity of large-scale testing has spurred the integration of machine learning techniques. This paper aims to provide a machine learning-focused survey on CAT, presenting a fresh perspective on this adaptive testing paradigm. We delve into measurement models, question selection algorithm, bank construction, and test control within CAT, exploring how machine learning can optimize these components. Through an analysis of current methods, strengths, limitations, and challenges, we strive to develop robust, fair, and efficient CAT systems. By bridging psychometric-driven CAT research with machine learning, this survey advocates for a more inclusive and interdisciplinary approach to the future of adaptive testing.

Isolation Levels (IL) act as correct contracts between applications and database management systems (DBMSs). The complex code logic and concurrent interactions among transactions make it a hard problem to expose violations of various ILs stated by DBMSs. With the recent proliferation of new DBMSs, especially the cloud ones, there is an urgent demand for a general way to detect bugs violating various ILs. The core challenges come from the requirements of: (a) lightweight (verifying without modifying the application logic in workloads and the source code of DBMSs), (b) generality (verifying various ILs), and (c) efficiency (performing efficient verification on a long running workload). To this end, we propose a powerful and practical bug-finding toolLeopard. For lightweight, we propose to infer transaction dependencies based on the time intervals of operations collected from the client-side, without modifying the source code of DBMSs. For generality, based on a thorough analysis of existing concurrency control protocols, we summarize and abstract four mechanisms which can implement ILs in all commercial DBMSs we have investigated. For efficiency, we design atwo-level pipelineto organize and sort massive time intervals in a time and memory conservative way; we propose amechanism-mirrored verificationto simulate the concurrency control protocols implemented in DBMSs for high throughputs; From experimental results,Leopardoutperforms existing methodsCobraandElle. In practice,Leopardhas a superpower to verify various ILs on any workload running on all commercial DBMSs. Moreover, it has successfully discovered 49 bugs that cannot be found by other existing methods.

User modeling serves as a crucial foundation for researchers to capture useful potential characteristics, playing a pivotal role in various applications such as recommender systems. One common challenge in user modeling is the cold-start problem, where interactions are notably limited for new users. To tackle this issue, the paradigm of meta-learning has been introduced to user modeling, yielding promising results. Similar to a guidebook for a new traveler, meta-learning significantly influences decision-making for new users in critical scenarios, such as career recommendations. Consequently, the issue of fairness in meta-learning has become paramount. Several methods have been proposed to mitigate unfairness in meta-learning and have shown promising results. However, a fundamental question remains unexplored: What is the critical factor leading to unfairness in meta-learned user modeling? Through theoretical analysis that integrates the meta-learning paradigm with group fairness metrics, we identify group proportion imbalance as a critical factor. Subsequently, another question arises: How can we mitigate the influence of this factor to enhance fairness while ensuring accuracy? To this end, we introduce a novel F airness-aware A daptive S ampling framework for me T a-learning, abbreviated as FAST. Its core concept involves adaptively adjusting the sampling distribution for different user groups during the interleaved training process of meta-learning. Moreover, we provide theoretical guarantees demonstrating the convergence of FAST, showcasing its potential to effectively eliminate unfairness. Furthermore, to ensure model accuracy, we enhance FAST with FAST+ by introducing a hybrid sampling strategy at an individual level. This strategy prioritizes fairness and thoroughly explores important users during the sampling process, allowing for a better accuracy-fairness tradeoff. Finally, we conduct extensive experiments on real-world datasets, which demonstrate the effectiveness of both FAST and FAST+ frameworks.

Cardinality Estimation (CardEst) is crucial for query optimization. Despite the remarkable achievement in DBMS, there is a pressing need to test or tune the work of CardEst. To satisfy the need, we introduce Artemis, a customizable workload generator, which can be used to generate various scenarios with the sensitive features for CardEst, including various data dependencies, complex SQL structures, and diverse cardinalities. It designs a PK-oriented deterministic data generation mechanism to plot various data characteristics; a search-based workload generation is proposed for composing queries with various complexities; it takes a constraint optimization-guided way to achieve a cost-effective cardinality calculation. In this demonstration, users can explore the core features of Artemis in generating workloads.

The large language model (LLM)-based knob tuning method has attracted considerable attention due to its excellent in-context learning ability and generalizability. However, the existing LLM-based tuning methods do not effectively harmonize multi-source external knowledge, leading to missed opportunities for enhanced knob tuning. In light of this, we propose Rabbit, a novel approach that leverages Retrieval-augmented generation to enhance database knob tuning tools, which seamlessly integrates structured historical tuning experience with graph-encoded static knowledge. First, we introduce an experience-driven knob selection strategy, enhanced by dependency-aware external knowledge integration, to systematically select key knobs. Second, we develop a cutting-edge multi-agent knob domain pruning method, which ensures the reduced search space remains compact yet effective. Finally, we leverage the few-shot capabilities of LLMs to act as surrogate models, enabling rapid exploration of the pruned search space, followed by incremental optimization that expands the search space using historical insights. Moreover, we also design an adaptive strategy to transition between these two search spaces, striking an optimal balance between exploration and exploitation. Extensive experiments on well-established bench-marks demonstrate that Rabbit outperforms the state-of-the-art methods in both effectiveness and efficiency, pointing to a new paradigm for this area.

Detecting training data for large language models (LLMs) is receiving growing attention, especially in applications requiring high reliability. While numerous efforts have been made to address this issue, they typically focus on accuracy without ensuring controllable results. To fill this gap, we propose **K**nockoff Inference-based **T**raining data **D**etector (KTD), a novel method that achieves rigorous false discovery rate (FDR) control in training data detection. Specifically, KTD generates synthetic knockoff samples that seamlessly replace original data points without compromising contextual integrity. A novel knockoff statistic, which incorporates multiple knockoff draws, is then calculated to ensure FDR control while maintaining high power. Our theoretical analysis demonstrates KTD's asymptotic optimality in terms of FDR control and power. Empirical experiments on real-world datasets such as WikiMIA, XSum and Real Time BBC News further validate KTD's superior performance compared to existing methods.

To optimize query parallelism techniques, substantial workloads are required with specific query plans and customized output size for each operator (denoted as cardinality constraint). To this end, a rich body of query-aware database generators (QAG) are proposed. However, the complex data dependencies hidden behind queries make previous QAGs suffer from deficiencies in supporting complex operators and controlling the generation errors. In this paper, we design a new generator Mirage supporting well for complex operators with low error bounds for cardinality constraints. First, Mirage leverages Query Rewriting and Set Transforming Rules to decouple dependencies between key and non-key columns, which could help generate each of them individually. Then, for the non-key columns, Mirage abstracts cardinality constraints of operators as placement requirements within each column's domain, and further models the generation problem as a classic bin packing problem. Finally, for the key columns, Mirage proposes a uniform representation of join cardinality constraints for all types of PK-FK joins and partitions the data according to the matching status between PK and F K columns. Then, it formulates the key population as a Constraint Programming problem, which can be solved by an existing CP Solver. The experiments show that Mirage conquers all previous work in either operator support or generation error.

Meta-learning has been widely employed to tackle the cold-start problem in user modeling. Similar to a guidebook for a new traveler, meta-learning significantly affects decision-making for new users in crucial scenarios, such as career recommendations. Consequently, the issue of fairness in meta-learning has gained paramount importance. Several methods have been proposed to mitigate unfairness in meta-learning and have shown promising results. However, a fundamental question remains unexplored: What is the critical factor leading to unfairness in meta-learned user modeling? Through the theoretical analysis that integrates the meta-learning paradigm with group fairness metrics, we identify group proportion imbalance as a critical factor. Subsequently, in order to mitigate the impact of this factor, we introduce a novel Fairness-aware Adaptive Sampling framework for meTa-learning, abbreviated as FAST. Its core concept involves adaptively adjusting the sampling distribution for different user groups during the interleaved training process of meta-learning. Furthermore, we provide theoretical guarantees demonstrating the convergence of FAST. Finally, empirical experiments conducted on three datasets reveal that FAST effectively enhances fairness while maintaining high accuracy. The code for FAST is available at https://github.com/zhengz99/FAST.