Harper Hua, Zhen Han, Zhengyuan Shen, Meng-Chieh Lee, Sheng Guan, Qi Zhu, Sullam Jeoung, Yueyan Chen, Yunfei Bai, Shuai Wang, Vassilis N. Ioannidis, Huzefa Rangwala. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent AI agents, such as ChatGPT and LLaMA, primarily rely on instruction tuning and reinforcement learning to calibrate the output of large language models (LLMs) with human intentions, ensuring the outputs are harmless and helpful. Existing methods heavily depend on the manual annotation of high-quality positive samples, while contending with issues such as noisy labels and minimal distinctions between preferred and dispreferred response data. However, readily available toxic samples with clear safety distinctions are often filtered out, removing valuable negative references that could aid LLMs in safety alignment. In response, we propose Positive–Toxic Self-Alignment (PT-ALIGN), a novel safety self-alignment approach that minimizes human supervision by automatically refining positive and toxic samples and performing fine-grained dual instruction tuning. Positive samples are harmless responses, while toxic samples deliberately contain extremely harmful content, serving as a new supervisory signal. Specifically, we utilize LLM itself to iteratively generate and refine training instances by only exploring fewer than 50 human annotations. We then employ two losses, i.e., maximum likelihood estimation (MLE) and fine-grained unlikelihood training (UT), to jointly learn to enhance the LLM’s safety. The MLE loss encourages an LLM to maximize the generation of harmless content based on positive samples. Conversely, the fine-grained UT loss guides the LLM to minimize the output of harmful words based on toxic samples at the token-level, thereby guiding the model to decouple safety from effectiveness, directing it toward safer fine-tuning objectives, and increasing the likelihood of generating helpful and reliable content. Experiments on 9 popular open-source LLMs demonstrate the effectiveness of our PT-ALIGN for safety alignment, while maintaining comparable levels of helpfulness and usefulness.

Kiran Ramnath, Kang Zhou, Sheng Guan, Soumya Smruti Mishra, Xuan Qi, Zhengyuan Shen, Shuai Wang, Sangmin Woo, Sullam Jeoung, Yawei Wang, Haozhu Wang, Han Ding, Yuzhe Lu, Zhichao Xu, Yun Zhou, Balasubramaniam Srinivasan, Qiaojing Yan, Yueyan Chen, Haibo Ding, Panpan Xu, Lin Lee Cheong. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Text watermarking for Large Language Models (LLMs) has made significant progress in detecting LLM outputs and preventing misuse. Current watermarking techniques offer high detectability, minimal impact on text quality, and robustness to text editing. However, current researches lack investigation into the imperceptibility of watermarking techniques in LLM services. This is crucial as LLM providers may not want to disclose the presence of watermarks in real-world scenarios, as it could reduce user willingness to use the service and make watermarks more vulnerable to attacks. This work is the first to investigate the imperceptibility of watermarked LLMs. We design an identification algorithm called Water-Probe that detects watermarks through well-designed prompts to the LLM. Our key motivation is that current watermarked LLMs expose consistent biases under the same watermark key, resulting in similar differences across prompts under different watermark keys. Experiments show that almost all mainstream watermarking algorithms are easily identified with our well-designed prompts, while Water-Probe demonstrates a minimal false positive rate for non-watermarked LLMs. Finally, we propose that the key to enhancing the imperceptibility of watermarked LLMs is to increase the randomness of watermark key selection. Based on this, we introduce the Water-Bag strategy, which significantly improves watermark imperceptibility by merging multiple watermark keys.

We demonstrate ModsNet , a search tool for pre-trained data science MOD el s recommendatio N using E xamplar da T aset. Given a set of pre-trained data science models, an "example" input dataset, and a user-specified performance metric, ModsNet answers the following query: "what are top-k models that have the best expected performance for the input data?" The need for searching high-quality pre-trained models is evident in data-driven analysis. Inspired by "query by example" paradigm, ModsNet does not require users to write complex queries, but only provide an "examplar" dataset, a task description, and a performance measure as input, and can automatically suggest top- k matching models that are expected to have desirable performance to perform the task over the provided sample dataset. ModsNet utilizes a knowledge graph to integrate model performances over datasets and synchronizes it with a bipartite graph neural network to estimate model performance, reduce inference cost, and promptly respond to top- k model search queries. To cope with strict cold-start (upon receiving a new dataset when no historical performance of registered models are observed), it performs a dynamic, cost-bounded "probe-and-select" strategy to incrementally identify promising models. We demonstrate the application of ModsNet in enabling efficient scientific data analysis.

Data analytical pipelines routinely involve various domain-specific data science models. Such models require expensive manual or training effort and often incur expensive validation costs (e.g., via scientific simulation analysis). Meanwhile, high-value models remain to be ad-hocly created, isolated, and underutilized for a broad community. Searching and accessing proper models for data analysis pipelines is desirable yet challenging for users without domain knowledge. This paper introduces ModsNet, a novel MODel SelectioN framework that only requires an Example daTaset. (1) We investigate the following problem: Given a library of pre-trained models, a limited amount of historical observations of their performance, and an "example" dataset as a query, return k models that are expected to perform the best over the query dataset. (2) We formulate a regression problem and introduce a knowledge-enhanced framework using a model-data interaction graph. Unlike traditional methods, (1) ModsNet uses a dynamic, cost-bounded "probe-and-select" strategy to incrementally identify promising pre-trained models in a strict cold-start scenario (when a new dataset without any interaction with existing models is given). (2) To reduce the learning cost, we develop a clustering-based sparsification strategy to prune unpromising models and their interactions. (3) We showcase of ModsNet built on top of a crowdsourced materials knowledge base platform. Our experiments verified its effectiveness, efficiency, and applications over real-world analytical pipelines.

We introduce GALE, an active adversarial learning framework to detect nodes with erroneous information in attributed graphs. GALE is empowered by a new adversarial active error detection framework, which interacts active learning with a graph generative adversarial model to best exploit limited labeled examples of erroneous nodes. It dynamically determines diversified query nodes in batches with bounded size in terms of node typicality to enrich a pool of examples, which in turn provides representative examples to best train an adversarial classifier to capture different types of errors. Moreover, GALE provides an annotation algorithm to suggest a context of possible correct attribute values and error types, to facilitate the labeling of query nodes. We show that using limited queries and examples, GALE significantly improves competing methods such as constraint-based detection, outlier detection, and Graph Neural Networks (e.g. GCNs), with 32%, 31%, and 17% gain in F-1 score on average, and is feasible in learning cost for large graphs.

Given a set of node groups in a graph (e.g., gender or race groups), how to succinctly summarize their neighbors, and meanwhile ensure a "fair" representation to mitigate under- or over-representation of a certain group? We propose a novel framework to compute concise summaries of node groups with fairness guarantees. (1) We introduce a pattern-correction structure called r-summaries. An r-summary uses a graph pattern set to specify representative nodes and an auxiliary edge correction set to losslessly describe their r-hop neighbors. (2) We formulate the fair group summarization problem, which is to compute an r-summary that can select and accurately describe high quality nodes and their neighbors with small edge corrections, and meanwhile guarantee a desirable coverage for each group. The need for generating such summaries is evident in social recommendation, healthcare and graph search. We show that the problem is $\Sigma _2^p$-complete with the verification problem already NP-complete. (3) We present approximation algorithms that can generate r-summaries with (a) guaranteed quality and coverage properties, and (b) relative approximations on optimal edge correction costs. For large groups, we introduce an efficient algorithm that interleaves node selection and localized pattern discovery to reduce unnecessary computation. In addition, we introduce an algorithm to incrementally maintain the r-summaries over dynamic graphs with evolving edges. Using real-world data, we experimentally verify the efficiency and effectiveness of our algorithms and verify their applications.

Modern multidisciplinary materials science routinely processes scientific workflows that integrate different data resources (e.g., X-ray data, scripts, analytical results). Most of such data resources are isolated in research labs, created ad-hocly, and remain underutilized. We demonstrate CRUX, a Crowdsourced platform for materials data ResoUrces and workflow eXploration. CRUX is empowered by coherent data-workflow modeling, knowledge-based resource assembly for workflow search, and data provenance to support workflow exploration. CRUX allows users to declare parameterized workflows as graph patterns, and automatically recommends crowdsourced resources with quality guarantees. We demonstrate the ease-of-use and the performance of CRUX with three categories of queries: data search, workflow recommendation, and resource exploration. We make case of CRUX for peak finding in X-ray Diffraction (XRD) data, a cornerstone task in materials research. We show that CRUX enables new interactive paradigms to explore and design workflows for data analysts in general.

This paper studies the problem of subgraph query generation with guarantees on both diversity and group fairness. Given a query template (with parameterized search predicates) and a set of node groups in a graph, it is to compute a set of sub-graph queries that instantiate the query template, and each query ensures diversified answers that meanwhile covers each group with a desired number of nodes. Such need is evident in web and social search with fairness constraints, query optimization, and query benchmarking. We formalize a bi-criteria optimization problem that aims to find a Pareto optimal set of query instances in terms of diversity and fairness measures. We show the problem is in Δ$P$2 and verify its hardness (NP-hard and fixed-parameter tractable). We provide (1) two efficient algorithms that can approximate Pareto optimal sets with E-dominance relations that yield representative query instances with a bounded size, and (2) an online algorithm that progressively generates and maintains fixed-size ∊-Pareto set with small delay time. We experimentally verify that our algorithms can efficiently generate queries with desired diversity and coverage properties for targeted groups.

Graph neural networks (GNNs) have emerged as powerful approaches for graph representation learning and node classification. Nevertheless, they can be vulnerable (sensitive) to link perturbations due to structural noise or adversarial attacks. This paper introduces RoboGNN, a novel framework that simultaneously robustifies an input classifier to a counterpart with certifiable robustness, and suggests desired graph representation with auxiliary links to ensure the robustness guarantee. (1) We introduce (p,θ)-robustness, which characterizes the robustness guarantee of a GNN-based classifier if its performance is insensitive for at least θ fraction of a targeted set of nodes under any perturbation of a set of vulnerable links up to a bounded size p. (2) We present a co-learning framework that interacts model learning with graph structural learning to robustify an input model M to a (p,θ)-robustness counterpart. The framework also outputs the desired graph structures that ensure the robustness. Using real-world benchmark graphs, we experimentally verify that roboGNN can effectively robustify representative GNNs with guaranteed robustness, and desirable gains on accuracy.

This paper investigates the problem of subgraph query generation with output that satisfies both diversity and fairness constraints. Given a set of groups with associated cardinality requirements, it is to compute subgraph queries with diversified output that meanwhile covers the groups with the desired cardinality. Such need is evident in web and social search with fairness constraints. We formalize subgraph query generation as a bi-criteria optimization problem on the diversity and fairness properties of queries, and verify its hardness and approximability. We show that the problem is in Σp2 , and remains NP-complete even for single-node queries. Despite the hardness, (1) we show that approximations exist whenever a corresponding subset selection process provides good solutions, and provide feasible algorithms with performance guarantees for two practical query generation scenarios. We also present a fast heuristic algorithm for the general problem, which early terminates without enumerating queries. We experimentally verify that our algorithms can efficiently generate queries with desired diversity and coverage properties for targeted groups.

Detecting nodes with erroneous values in real-world graphs remains challenging due to the lack of examples and various error scenarios. We demonstrate GEDet, an error detection engine that can detect erroneous nodes in graphs with a few examples. The GEDet framework tackles error detection as a few-shot node classification problem. We invite the attendees to experience the following unique features. (1) Few-shot detection . Users only need to provide a few examples of erroneous nodes to perform error detection with GEDet. GEDet achieves desirable accuracy with (a) a graph augmentation module, which automatically generates synthetic examples to learn the classifier, and (b) an adversarial detection module, which improves classifiers to better distinguish erroneous nodes from both cleaned nodes and synthetic examples. We show that GEDet significantly improves the state-of-the-art error detection methods. (2) Diverse error scenarios . GEDet profiles data errors with a built-in library of transformation functions from correct values to errors. Users can also easily "plug in" new error types or examples. (3) User-centric detection . GEDet supports (a) an active learning mode to engage users to verify detected results, and adapts the error detection process accordingly; and (b) visual interfaces to interpret and track detected errors.

Detecting nodes with erroneous information in graphs is important yet challenging, due to the lack of examples and the diversified s cenarios o f e rrors. W e i ntroduce GEDet, a few-shot learning based framework to detect erroneous nodes in graphs. GEDet consists of two novel components, each addresses a unique challenge. (1) To cope with the lack of examples, we introduce a graph augmentation module to enrich training labels. The module not only generates additional synthetic training labels by simulating different erroneous scenarios, but also exploits non-local relations to enrich neighborhood information. (2) To further improve the accuracy, we introduce an adversarially learned module that can better detect erroneous nodes by distinguishing nodes with synthetic and real labels encoded by graph autoencoders. Unlike conventional error detection models, GEDet yields effective classifiers that are optimized for a few yet diversified examples in the presence of multiple error scenarios. We show that using only a small number of examples, GEDet significantly improves the competing methods such as constraint-based detection and anomaly detection, with a gain of 35% on recall, and 30% on precision.

Backbones refer to critical tree structures that span a set of nodes of interests in networks. This paper introduces a novel class of attributed backbones and detection algorithms in richly attributed networks. Unlike conventional models, attributed backbones capture dynamics in edge cost model: it specifies affinitive attributes for each edge, and the cost of each edge is dynamically determined by the selection of its associated affinitive attributes and the closeness of their values at its end nodes. The backbone discovery is to compute an attributed backbone that covers interested nodes with smallest connection cost dynamically determined by selected affinitive attributes. While this problem is hard to approximate, we develop feasible algorithms within practical reach for large attributed networks. (1) We show that this problem is fixed-parameter approximable parameterized by the number of affinitive attributes, by providing a Lagrangean-preserving 2-approximation. (2) When the attribute number is large and specifying closeness function is difficult, we provide a fast heuristic, which learns an edge-generative model, and applies the model to infer best backbones, without the need of specifying closeness functions. Using real-world networks, we verify the effectiveness and efficiency of our algorithms and show their applications in collaboration recommendation.

If an object is interreflected between two planar mirrors, we may take an image containing both the object and its multiple reflections, i.e., simultaneously imaging multiple views of an object by a single pinhole camera. This paper emphasizes the problem of recovering both the intrinsic and extrinsic parameters of the camera using multiple silhouettes from one single image. View pairs among views in a single image can be divided into two kinds by the relationship between the two views in the pair: reflected by some mirror (real or virtual) and in a circular motion. Epipoles in the first kind of pairs can be easily determined from intersections of common tangent lines of silhouettes. Based on the projective properties of these epipoles, efficient methods are proposed to recover both the imaged circular points and the included angle between two mirrors. Epipoles in the second kind of pairs can be recovered simultaneously with the projection of intersection line between two mirrors by solving a simple 1D optimization problem using the consistency constraint of epipolar tangent lines. Fundamental matrices among views in a single image are all recovered. Using the estimated intrinsic and extrinsic parameters of the camera, a euclidean reconstruction can be obtained. Experiments validate the proposed approach.

The Dempster-Shafer theory of evidence has been found to be promising for a variety of inexact reasoning applications. Many attempts have been made to generalize the theory to problem spaces which are not just the power sets of finite sets, which the original Dempster-Shafer theory addresses, but general Boolean algebras. However, some most important structures in applications to expert systems such as Gordon-Shortliffe's tree hierarchy (J. Gordon & E.H. Shortliffe 1985), are lattices rather than Boolean algebras. It is interesting to generalize evidence theory to general lattices. The authors generalize the theory by reworking some of the conventional theorems in evidence theory and establish the relationships between weaker forms of the familiar evidential functions.