Responsible AI (RAI) tools—checklists, templates, and governance processes—often engage RAI champions, individuals intrinsically motivated to advocate ethical practices, but fail to reach non-champions, who frequently dismiss them as bureaucratic tasks. To explore this gap, we shadowed meetings and interviewed data scientists at an organization, finding that practitioners perceived RAI as irrelevant to their work. Building on these insights and theoretical foundations, we derived design principles for engaging non-champions, and introduced sticky stories—narratives of unexpected ML harms designed to be concrete, severe, surprising, diverse, and relevant, unlike widely circulated media to which practitioners are desensitized. Using a compound AI system, we generated and evaluated sticky stories through human and LLM assessments at scale, confirming they embodied the intended qualities. In a study with 29 practitioners, we found that, compared to regular stories, sticky stories significantly increased the engagement time on harm identification, broadened the range of harms recognized, and fostered deeper reflection.

Large vision-language models (LVLMs) have shown significant progress in video understanding, but they struggle to scale to long videos due to limited context windows. Existing methods reduce input dimensionality via frame sampling and feature compression, yet discard details and incur high computational cost for post-training. In contrast, retrieval-augmented generation (RAG) that indexes long videos for query retrieval and memory-based methods that maintain evolving long-term stores, offer a lighter and deployment-friendly solution. Nevertheless, they rely on shallow retrieval that selects only top-ranked segments and fails to integrate information across multiple relevant video episodes. Inspired by Multiple-Trace Theory in cognitive psychology, we revisit long video understanding from a probe-echo perspective, in which human episodic memories are activated and integrated in parallel. Building on this insight, we propose ProEchoMem, a cognitive-inspired framework that simulates the probe-echo mechanism: (1) Incremental Episodic Memory Construction builds structured knowledge graphs from video streams; (2) Probe-Driven Memory Activation generates probe signals from user queries to activate all stored traces simultaneously; (3) Memory Echo Synthesis integrates activated traces into a coherent and structured memory echo. Experiments on LongerVideos, LVBench, and cross-domain settings demonstrate the effectiveness of ProEchoMem, with multi-trace probing achieving up to 14.2% higher relevance and ablation studies validating the contribution of each module. The code is available at https://github.com/Applied-Machine-Learning-Lab/SIGIR26_ProEchoMem

Creating realistic and fully detailed 3D human models using a minimal number of views has long been a challenging goal in 3D human reconstruction. Reconstructing a realistic human model from only two images (front and back) is particularly difficult due to the limited 3D information available, leading to two major challenges: (1) it is difficult to establish spatial consistency for reconstruction due to the lack of sufficient images for reliable matching, and (2) incomplete field of view results in missing color information. To address these challenges, we propose DuetGS, a novel pipeline that divides the reconstruction process into two stages: geometry reconstruction and color reconstruction. For geometry reconstruction, we employ a data-driven neural network to recover a full-body mesh from the front and back images, providing the spatial positioning for Gaussians. For color reconstruction, we adapt Gaussian Splatting and integrate our proposed unsupervised color propagation method to establish the color details of the Gaussians. Furthermore, our Gaussians are directly mapped to the mesh, allowing us to control their rotation and translation through mesh manipulation. This mapping ensures compatibility with various animation techniques. Extensive experiments on the THUman, CustomHumans, and PeopleSnapshot datasets demonstrate that our approach outperforms existing methods in terms of reconstruction accuracy and visual quality.

Gene regulatory networks (GRNs) depict the intri-cate interactions among genes, which are essential for under-standing complex cellular biological processes. However, predicting potential links in GRNs presents substantial challenges. A key limitation of current approaches is their reliance solely on single-cell gene expression data, which neglects important regulatory signals encoded in DNA sequences. Furthermore, the sparse nature of GRNs restricts communication between gene nodes and introduces a significant imbalance between known and unknown links, further hindering the identification of novel regulatory in-teractions. To overcome these limitations, we propose GRN-LGT, a predictive model for GRN link prediction that integrates a gene language model DeepGene with a Graph Transformer. GRN-LGT employs DeepGene language model to encode gene sequences and extract sequence-level features. It then fuses these sequence-derived features with single-cell gene expression features, and feeds the combined representations into a Graph Transformer to generate gene node embeddings. Finally, the model is trained in a supervised manner with known GRN links. Compared to existing methods, GRN-LGT demonstrates greater robustness and achieves the best average performance across 44 downstream tasks encompassing multi-species datasets. The source code is available at https://github.com/ChenYanxin99/GRN-LGT.git.

In recent years, increasing evidence has demonstrated a close association between metabolites and various complex human diseases, providing valuable insights for disease diagnosis, treatment, and prevention. Although deep learning-based approaches have achieved certain success in predicting metabolic disease associations, challenges remain in enriching graph information and effectively integrating metabolic and disease features. To address these issues, this paper proposes a model named DCMDA, which extracts deep features of both metabolites and diseases using Dual-network Cross-learning for Metabolite-Disease Association prediction. DCMDA consists of three parts. The data processing module integrates similarity networks with association networks to construct a heterogeneous network. The feature extraction module extracts features from the metabolite-disease association network based on the non-negative matrix factorization method and from the heterogeneous network using graph autoencoder techniques. The feature fusion module combines the association matrix feature with the heterogeneous network feature through a Cross-Attention mechanism, thereby obtaining deep representations of metabolites and diseases. These features are then used to train the model to predict association scores between metabolites and diseases. Experimental results demonstrate that in 5-fold cross-validation, DCMDA achieves an area under the receiver operating characteristic curve (AUC) of 97.8% and an area under the precision-recall curve (AUPR) of 97.9%, outperforming state-of-the-art prediction methods.