Gaze target estimation aims to infer the position of a person's gaze within a scene. Within mainstream design logic, multi-branch methods require extra supervision and annotations, while streamlined designs prioritize low-level visual saliency over true gaze intent. The former leads to a high annotation burden and hinders domain transfer, whereas the latter causes misalignment between predicted attention and actual gaze targets. To address this issue, we propose TextGaze, a unified cross-modal architecture that leverages a Large Vision-Language Model (LVLM) as scalable semantic guidance to balance the two design paradigms. The model extracts visual features from a frozen encoder and utilizes an LVLM to obtain gaze-aligned textual cues. We design a transformer-based fusion module with hierarchical text supervision to preserve task semantics. Lightweight decoding heads enable the joint prediction of gaze heatmaps and in-/out-of-frame status. We evaluate our method on four mainstream datasets, and the results show competitive performance across key metrics with robust cross-dataset generalisation without extra fine-tuning. Overall, we provide a streamlined alternative to traditional designs and highlight the potential of LVLMs as accessible auxiliary guidance for gaze estimation. All contents are available at: https://github.com/idremo/TextGaze-IJCAI2026.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
FlashAttention improves efficiency through tiling, but its online softmax still relies on floating-point arithmetic for numerical stability, making full quantization difficult. We identify three main obstacles to integer-only FlashAttention: (1) scale explosion during tile-wise accumulation, (2) inefficient shift-based exponential operations on GPUs, and (3) quantization granularity constraints requiring uniform scales for integer comparison. To address these challenges, we propose QFlash, an end-to-end integer FlashAttention design that performs softmax entirely in the integer domain and runs as a single Triton kernel. On seven attention workloads from ViT, DeiT, and Swin models, QFlash achieves up to 6.73x speedup over I-ViT and up to 8.69x speedup on Swin, while reducing energy consumption by 18.8% compared to FP16 FlashAttention, without sacrificing Top-1 accuracy on ViT/DeiT and remaining competitive on Swin under per-tensor quantization. Our code is publicly available at https://github.com/EfficientCompLab/qflash.
Designing functional protein sequences that satisfy multiple desired properties is a core research focus of protein engineering. Prior methods struggle with inability or inefficiency when dealing with numerous, often conflicting, properties. We propose Multi-Property Protein Diffusion, (MP2D), a unified framework for multi-objective protein sequence optimization that integrates conditional discrete diffusion with constrained MCTS and global iterative refinement. MP2D formulates diffusion denoising as a constrained sequential decision-making process and employs MCTS to explore diverse denoising trajectories guided by Pareto-based rewards. A global iterative refinement strategy further enables repeated remasking and re-optimization of candidate sequences, while a dynamic Pareto constraint prevents candidate bloat and maintains balanced trade-offs across objectives. We evaluate MP2D on two challenging multi-objective protein design tasks: antimicrobial peptide and protein binder optimization, involving four to five conflicting properties. Experimental results demonstrate that MP2D consistently outperforms existing multi-objective baselines, achieving robust and balanced improvements across all objectives without retraining generative models. These results highlight MP2D as a practical and scalable solution for multi-objective functional protein design.
Low-Rank Adaptation (LoRA) is a popular approach that enables large language models (LLMs) to quickly adapt to domain-specific tasks by adding lightweight trainable adapters. Existing multi-LoRA serving systems typically exploit parameter sharing to serve hundreds of LoRA models with a single base model. However, most systems rely on first-come-first-serve (FCFS) scheduling, which can incur severe queuing delay and lead to excessive adapter memory usage, squeezing KV cache space and reducing concurrency and throughput. To address these challenges, we propose M-LoRA, a memory-aware multi-LoRA serving system that reduces queuing delay and improves throughput through efficient request scheduling guided by fine-grained memory modeling. M-LoRA consists of three components: (1) Multi-LoRA Length Predictor, which estimates the output length and KV-cache demand of each request; (2) Memory-aware Speculative Scheduler, which dispatches requests based on predicted KV-cache and adapter memory requirements to maximize throughput under a fixed GPU memory budget; and (3) Recovery Mechanism, which corrects mispredictions while preserving generation quality. Compared to state-of-the-art LoRA serving systems, M-LoRA reduces average per-token latency by 87% and improves throughput by up to 1.7x.
Molecular representation learning aims to learn vector embeddings that capture molecular structure and geometry, thereby enabling property prediction and downstream scientific applications. In many AI for science tasks, labeled data are expensive to obtain and therefore limited in availability. Under the few-shot setting, models trained with scarce supervision often learn brittle structure–property relationships, resulting in substantially higher prediction errors and reduced generalization to unseen molecules. To address this limitation, we propose PCEvo, a path-consistent representation method that learns from virtual paths through dynamic structural evolution. PCEvo enumerates multiple chemically feasible edit paths between retrieved similar molecular pairs under topological dependency constraints. It transforms the labels of the two molecules into stepwise supervision along each virtual evolutionary path. It introduces a path-consistency objective that enforces prediction invariance across alternative paths connecting the same two molecules. Comprehensive experiments on the QM9 and MoleculeNet datasets demonstrate that PCEvo substantially improves the few-shot generalization performance of baseline methods. The code is available at https://github.com/DrugD/PCEvo.
Immune Checkpoint Inhibitors (ICIs) represent a cornerstone of modern cancer immunotherapy. However, their clinical application is frequently accompanied by immune-related Adverse Events (irAEs) of diverse severity. Predicting Gene-Disease Associations (GDAs) is crucial for identifying the related genes that cause irAEs but remains experimentally expensive in the context of cancer immunotherapy. While existing graph neural network-based methods provide solutions to this problem, they often overlook disease-disease and disease-gene associations that may exist but have not yet been identified, leading to contaminated representations. To address these limitations, we propose Gene-Disease Associations based on Optimal Transport (GDAs-OT), a novel framework for GDA prediction. GDAs-OT constructs gene-gene, disease-disease and gene-disease relationships as graphs. Optimal transport was utilized to identify potential disease-disease associations to expand the disease-disease graph structure, and select high-confidence negative pairs in the gene-disease graph. By mapping nodes into a refined embedding space, the processed gene-disease graph guides node representation learning, providing robust node representations. Comprehensive experiments demonstrate that GDAs-OT outperforms state-of-the-art methods across most evaluation metrics. In addition, GDAs-OT successfully identifies potential risk genes for irAEs, providing a computational foundation for understanding the mechanisms of immunotherapy toxicity. Code is available at https://github.com/RuhaoLiu/GDAs-OT.
Graph-based multi-view clustering, with its ability to mine potential associations between samples, has attracted extensive attention. To capture high-order correlations, tensor-based frameworks have been introduced to model multiple graphs jointly. Although these methods have achieved promising performance, existing methods mainly focus on stacking consistency graphs with low-rank constraints, while overlooking high-order specificity within diversity graphs, and the exploration of fine-grained diversity information among different samples across views remains insufficient. To address these issues, we propose High-Rank Tensor Specificity Induced Cooperative Multi-ViewGraph Learning (HTS-CMGL). Specifically, we first obtain the consistency graphs and diversity graphs from multi-view data, which are further reconstructed into the consistency tensor and the diversity tensor, respectively. Subsequently, a novel Enhanced Tensor Rank (ETR) is imposed on the consistency tensor, which is a tighter approximation of the tensor rank and is more noisy-robust to explore the high-order consistency. Meanwhile, we design a new Tensor High-Rank Logarithmic Norm (THLN) on the diversity tensor. By leveraging a unique high-rank constraint mechanism, THLN not only actively preserves high-rank and informative features, but also simultaneously captures view-level and sample-level diversity information. Extensive experiments on multiple datasets demonstrate the effectiveness of our proposed method on clustering multi-view data.
Standard unsupervised multi-view feature selection (UMFS) methods for large datasets exhibit limitations in modeling the competition between cross-view alignment and intra-view diversity, resulting in suboptimal solutions and expensive computational costs. This challenge is exacerbated by the diverse signals inherent in complex samples, which tend to mask shared patterns, thus complicating the pursuit of an optimal trade-off. In this work, we propose a Progressive Adversarial Feature Selection (ProAd-FS) framework for large-scale multi-view learning, which formulates this static trade-off objective as a dynamic competitive process. To be specific, ProAd-FS develops an adversarial decoupled architecture that assigns these competing objectives to distinct inter-view and intra-view games, guided by a robust gated curriculum to prioritize the learning of underlying structures. The entire framework exhibits linear computational complexity in both sample size and feature dimension, and embeds structural sparse regularization for end-to-end optimization. Extensive experimental results show that ProAd-FS achieves state-of-the-art performance while being highly scalable as well, providing an effective solution for large-scale UMFS tasks.
Given a social network 𝒢 with n nodes and m edges, the opinion maximization (OM) problem aims at identifying k (k ≪ n) opinion leaders to maximize their opinion propagation in 𝒢. Despite its significance and prevalence, existing OM methods often struggle to strike a satisfactory balance between theoretical performance and practical efficiency. This paper provides an inverse optimization perspective for OM by reformulating it as an opinion minimization (OMin) problem, whose objective function holds succinct expression and desired properties. Then, we introduce a bounded estimation scheme (BES) that can estimate the objective function of OMin with bounded error in Õ(kn+m) expected time, thereby making the optimization for OMin more efficient. To further enhance the optimization efficiency, a decomposition-based decremental estimation algorithm (DDEA) with a near (1-1/e) approximation ratio in Õ(kmn) expected time is specifically designed for OMin. Extensive experiments on real-world social networks of varying scales demonstrate the effectiveness of BES and the superiority of DDEA.
We study repeated allocation of shared resources among agents with time-varying demands and capped linear utilities. In this setting, independently maximizing the minimum utility in each round satisfies sharing incentives (agents weakly prefer participating in the mechanism to not participating), strategyproofness (agents have no incentive to misreport their demands), and Pareto efficiency. However, this max-min mechanism can lead to large disparities in the total resources received by agents, even when they have the same average demand. We introduce credit fairness, a property that, together with Pareto efficiency, strengthens sharing incentives by ensuring that agents who lend resources in early rounds are able to recoup them in later rounds. Credit fairness can be achieved in conjunction with either Pareto efficiency or strategyproofness individually, but we show that, under anonymity, it cannot be achieved together with both. We propose a mechanism that is credit fair and Pareto efficient, and evaluate it in a computational resource-sharing setting.
Combinatorial optimization (CO) problems have widespread applications in science and engineering, but they present significant computational challenges. Recent advancements in generative models, particularly diffusion models, have shown promise in bypassing traditional optimization solvers by directly generating near-optimal solutions. However, existing diffusion solvers are highly sensitive to the quality of the training dataset, particularly the number of problem instances for training and the optimality of their labels. Notably, we observe an exponential scaling law between the performance of diffusion-based solvers and the number of near-optimally labeled instances needed. When training instances are scarce or sub-optimally labeled, diffusion-based solvers suffer significant performance degradation. To enhance the robustness of diffusion solvers to dataset quality, we propose a robust diffusion solver for combinatorial optimization capable of learning from sub-optimally labeled instances follows a two-stage generate-then-decode framework, integrating an objective-guided diffusion model, further reinforced by classifier-free guidance, to produce solutions that surpass the optimality of the training dataset. Experiments demonstrate the improved robustness in \myalg compared to the diffusion-based solver baseline, in a range of combinatorial optimization benchmark tasks such as TSP (Traveling Salesman Problem) and MIS (Maximum Independent Set).
Large language models (LLMs) are increasingly used for hardware and firmware code generation, but existing studies primarily evaluate functional correctness while largely overlooking security. However, LLM-generated code that appears functionally sound may embed security flaws which could induce catastrophic damages after deployment. This critical research gap motivates us to design a benchmark for assessing security awareness under realistic specifications. In this work, we introduce HardSecBench, a benchmark with 924 tasks spanning Verilog Register Transfer Level (RTL) and firmware-level C, covering 76 hardware-relevant Common Weakness Enumeration (CWE) entries. Each task includes a structured specification, a secure reference implementation, and executable tests. To automate artifact synthesis, we propose a multi-agent pipeline that decouples synthesis from verification and grounds evaluation in execution evidence, enabling reliable evaluation. We evaluate diverse LLMs and find that they often satisfy functional requirements while leaving security risks. We also find that security results vary with prompting. These findings highlight pressing challenges and offer actionable insights for future advancements in LLM-assisted hardware design. Our data and code are available at https://github.com/chenqirui2002/HardSecBench.
Existing multimodal recommendation models using complex fusion mechanisms (e.g., attention) or multi-stage processes (e.g., early or late fusion) integrate different modalities. However, attention-based adaptive fusion is prone to shortcut learning, where dominant collaborative signals (ID) can overshadow other modalities. This dominance affects the entire fusion process: early fusion often amplifies biases driven by identity, while late fusion struggles to extract preference-relevant signals from misaligned modalities, even with alignment regularization. To address these issues, we propose Multi-Teacher Single-Student Online Distillation for Multimodal Recommendation (MTS2-4MM), which reframes the multimodal recommendation task from direct fusion to controllable knowledge transfer. Specifically, we construct multiple teachers to specialize in complementary perspectives, and a unified student distills their guidance via objectives at both the ranking and representation levels. This design explicitly controls modality contributions, improves robustness to modality noise and misalignment. Furthermore, we design a Modality-specific Preference Extractor to explicitly extract user preferences across different modalities equally. Extensive experiments across five real-world datasets demonstrate that MTS2-4MM consistently outperforms state-of-the-art baselines, achieving improvements of up to 7.22%.
Spatially resolved transcriptomics integrates gene expression with spatial coordinates to decode tissue microenvironments. Existing methods predominantly utilize graph structures to model relationships between spots. However, their performance is bottlenecked by the reliability of gene feature graph, facing the following hurdles: (1) ubiquitous housekeeping genes cause high-expression spots to densely connect with heterogeneous spots, leading to a skewed graph structure; (2) information reduction during Highly Variable Gene (HVG) selection results in the loss of intrinsic local structures. To address these challenges, we propose a Spot-Adaptive Structural Rectification method, called SASR. Specifically, SASR employs a hyperspherical expansion constraint that projects gene expression profiles onto a unit hypersphere to maximize angular distances, effectively separating spots falsely clustered by high total counts. Simultaneously, a topological consistency constraint repairs structural fractures caused by HVG selection via aligning latent embeddings with the local structures of the raw full-gene space. The complementary synergy balances angular discriminability with topological fidelity for accurate clustering. Experiments demonstrate that SASR effectively corrects structural biases and surpasses state-of-the-art methods in spatial clustering.
Cross-view geo-localization (CVGL) aims at localizing a ground-level query by retrieving its corresponding match from a database of geo-tagged satellite images. Existing multi-modal CVGL methods lack a structured design in the fusion stage, limiting their ability to fully exploit the information from multiple modalities. To overcome this limitation, we propose a Structural-Analysis-Based fusion principle that guides the design of network architecture. Following this principle, we present Decoupled Query Fusion (DQF), a novel fusion module that decouples feature interactions through role-specific learnable queries. These learnable queries aggregate features into distinct slots. Specifically, modality-specific queries capture unique information from each modality, while cross-modal queries extract redundant and synergistic information between different modalities. To better discriminate positive samples at varying geographic distances, we further propose Geo-aware Circle Loss, which adaptively weights supervision signal based on the geographical distance between samples and anchors. Extensive experiments on large-scale benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches. Comprehensive ablation studies further validate the effectiveness of each component. Our code and models will be made publicly available.
Over-parameterized neural networks, i.e., models with excess capacity that can fit training data exactly, have demonstrated superior generalization performance compared to classical models with balanced capacity. Nevertheless, their deployment in safety-critical domains re- mains severely constrained by their susceptibility to, e.g., natural per- turbations and adversarial manipulations. Verification techniques can solve such problems, but the computational cost of these methods of- ten scales poorly, specifically when applied to large models. In this work, we demonstrate that over-parameterization can be exploited not merely to enhance generalization, but also to mitigate neuron instability, one of the parameters affecting the efficiency of verification. Our experimental findings suggest that over-parameterization may serve as a crucial mech- anism for reconciling the long-standing trade-off between generalization and verifiability of neural networks.
Large Vision-Language Models (LVLMs) often suffer from hallucinations, generating descriptions that include visual details absent from the input image. Recent preference alignment methods typically rely on supervision distilled from stronger models such as GPT. However, this offline paradigm introduces a Supervision-Perception Mismatch: the student model is forced to align with fine-grained details beyond its perceptual capacity, learning to guess rather than to see. To obtain reliable self-supervision for online learning, we identify a Generative-Discriminative Gap within LVLMs, where models exhibit higher accuracy on discriminative verification than open-ended generation. Leveraging this capability, we propose Online Self-CAlibRation (OSCAR), a framework that integrates Monte Carlo Tree Search with a Dual-Granularity Reward Mechanism to construct preference data and iteratively refines the model via Direct Preference Optimization. Extensive experiments demonstrate that OSCAR achieves state-of-the-art performance on hallucination benchmarks while preserving general multimodal capabilities.
Cognitive diagnosis (CD) aims to infer students' mastery of knowledge concepts from their response behaviors and constitutes a core component of intelligent education and personalized learning. However, existing graph-based CD models struggle to handle the pronounced long-tail distributions in educational data, where most students and concepts interact with only a limited number of exercises, resulting in suboptimal representation learning and poor generalization to low-frequency instances. To address this challenge, we propose HiCD (Hyperbolic insight for Cognitive Diagnosis), a novel hyperbolic model that embeds students, exercises, and concepts into non-Euclidean space. By exploiting the exponential representational capacity of hyperbolic geometry, HiCD naturally captures hierarchical and sparse structures, effectively alleviating long-tail bias while enhancing embedding expressiveness. A key contribution of HiCD is a hyperbolic diagnostic function that operates directly on the manifold, avoiding Euclidean approximations and preserving geometric consistency. Moreover, HiCD decomposes the educational graph into three semantically distinct subgraphs and assigns each a dedicated curvature, enabling adaptive geometric modeling of heterogeneous relations. Extensive experiments on multiple benchmark datasets demonstrate that HiCD consistently improves diagnostic accuracy and robustness, particularly under severe long-tail scenarios. The source code is available at: https://github.com/CyberXie/HiCD.
Partial multi-label learning (PML) addresses weakly-supervised scenarios where each instance is associated with a candidate label set containing both ground-truth and noisy labels. Existing PML methods primarily focus on instance-level features or pairwise label correlations for disambiguation. Building on recent insights that exploiting pairwise label correlations improves disambiguation, we observe that high-order label correlations provide even stronger disambiguation evidence in partial settings because ground-truth labels with high-order correlations frequently co-occur, while noise labels produce inconsistent combinations that rarely repeat. To exploit this property, we propose REHC-PML (Label Confidence REcovery with High-order Label Correlations in Partial Multi-label Learning), which mines frequent high-order co-occurrences from candidate sets to identify global high-order label correlations, selects instance-relevant correlations via Gumbel-Softmax pruning, and propagates their evidence to constituent labels for confidence recovery. Self-training iteratively refines pseudo-labels and trains the classifier. Extensive experiments on both real-world and UCI datasets demonstrate the effectiveness of REHC-PML.
Cross-modal hashing has been extensively adopted in cross-modal retrieval tasks owing to its high storage efficiency and fast retrieval capability. However, in practical applications, multimodal data often suffer from modality missing issues, which cause semantic incompleteness and thus severely impair both cross-modal alignment and hashing-based retrieval performance. To address these issues, this paper proposes a novel incomplete cross-modal retrieval framework, called retrieval-guided completion hashing with token–patch alignment (RGCH-TPA). Specifically, it first constructs a dynamic memory bank based on the features of complete samples and uses available modalities to retrieve and locate similar regions of incomplete samples. Thus, it achieves intra-modal reconstruction of missing data under the condition of available modalities using retrieved expert signals. Moreover, a token-patch level cross-modal alignment mechanism is introduced to model the interaction between text tokens and image patches, while global representations are jointly integrated to further enhance the discriminability and robustness of the learned cross-modal representations. Extensive evaluations on three benchmark datasets demonstrate that the proposed RGCH-TPA method consistently outperforms other state-of-the-art incomplete cross-modal hashing methods across missing-modality ratios. The source code is available at https://github.com/szq0816/RGCH-TPA.