arXiv subjects

Kai Yao

Publications and source records attributed to Kai Yao.

At least 19 recordsLinked to original sources

arXiv2026

Xemo-Talker: Unlock Emotions Explicitly for Audio-Driven Talking Portrait Synthesis

Precise emotion control in audio-driven talking heads remains a challenge due to the reliance on implicit emotion regulation in existing systems, which often leads to indirect and insufficient control. Additionally, training with explicit emotion-related losses across the entire motion space poses significant difficulties due to the inherent trade-off between accurate lip synchronization and fine-grained emotion control. In this paper, we reveal a key finding: although emotional cues are distributed throughout the motion space, concentrating discriminative supervision on less-principal components achieves a better emotion-lip synchronization balance, as principal components mainly encode high-energy articulation and pose variations. Building on this insight, we propose Xemo-Talker, which first learns a neutral speech-to-motion mapping for stable articulation and lip synchronization, and then introduces a lightweight emotion branch guided by less-principal subspace supervision. To enhance emotion control, we design a Tri-Loss consisting of inter-class separation, intra-class compactness, and less-principal contrastive learning. Given an audio input, a reference image, and an emotion label, Xemo-Talker achieves state-of-the-art emotion classification accuracy while maintaining competitive lip synchronization and high inference efficiency, with performance approaching that measured on real videos.The source code is publicly available at https://github.com/chaolongy/Xemo-Talker.

cs.CV↗

arXiv2026

When AI Does the Work, What Is Learning For? Post-Instrumental Learning and the Risk of Capacity Dissolution

As AI systems become capable of producing the essays, code, reports, summaries, plans, and decisions through which institutions usually recognize competence, a familiar question becomes harder to answer: what is learning for? Existing AI ethics rightly emphasizes present failures--bias, opacity, hallucination, labor extraction, privacy risk, and weak accountability. But if the case for learning rests only on those failures, then each technical improvement appears to weaken it. This article develops a different answer. Using the idealization of AI that executes specified tasks flawlessly while lacking authority over purposes, legitimacy, and responsibility, we argue for post-instrumental learning: learning that preserves the capacities people and institutions need when many useful outputs can be delegated. We analyze five such capacities--end-setting, reason-giving, contestability, refusal/revision, and participation--and name their erosion capacity dissolution. The central case is assessment under generative AI. When a polished artifact no longer reliably evidences understanding, institutions must assess the learner's accountable relation to AI-mediated work rather than the artifact alone. The takeaway is practical: AI governance should evaluate not only whether systems perform well, but also whether their deployment leaves people able to understand, challenge, revise, and share responsibility for the practices those systems mediate.

cs.CY↗

arXiv2026

Common-Neighbor-Count-Based Representative Possible World Finding on Uncertain Graphs

A representative possible world (RPW) is a deterministic graph derived from an uncertain graph $\mathcal{G}$ where a designated structural feature closely approximates its expected value in $\mathcal{G}$. Serving as a proxy for $\mathcal{G}$, the RPW allows conventional deterministic algorithms to be directly executed on it for mining tasks targeting this feature, thereby avoiding computationally expensive enumeration or sampling on $\mathcal{G}$. Existing studies on RPWs primarily focus on individual node features, e.g., degree or triangle degree. However, many mining tasks, such as link prediction, critically rely on the number of common neighbors between two nodes, which is a pairwise feature. To bridge this gap, we study the \underline{C}ommon-neighbor-count-based \underline{R}epresentative \underline{P}ossible \underline{W}orld (CRPW) problem, extending RPWs from preserving node-level statistics to preserving pairwise structural relationships. The problem seeks the possible world that best preserves the expected numbers of common neighbors between node pair, and we prove that is NP-hard. To address it, we develop a two-stage basic algorithm that quickly initializes a possible world and then refines it iteratively. We next accelerate the refinement by replacing its costly floating-point evaluation with an efficient integer counting strategy, as the refinement only requires determining whether a change is beneficial, rather than computing its exact magnitude. Moreover, we design a Beta-based adaptive termination method to automatically stop the refinement once the desired quality of the possible world is reached, preventing over- or under-execution. Extensive experiments on real-world uncertain graphs demonstrate the effectiveness of our algorithms on diverse mining tasks. Especially on common-neighbor-related tasks, we achieve the best performance among all compared methods.

cs.DB↗

arXiv2026

What Does the Credential Still Certify? Cognitive Stewardship for AI-Mediated Education

Generative AI is changing a basic premise of educational assessment: that submitted work can reliably evidence the human capacities a credential claims to certify. The challenge is not simply whether students use AI, but what remains inferable about learning when some cognitive work has been delegated to a system. This paper develops cognitive stewardship, a framework for AI-mediated assessment that links the learning claim, delegation boundary, evidence standard, and safeguards. We then audit verified public generative AI assessment guidance from 30 universities. Using a pre-specified scoring codebook--a written, source-grounded rubric--four open-weight LLM models applied the rubric as structured coders, with scores averaged to reduce dependence on any single model's bias. The audit shows that public policies are becoming better at classifying AI use than at explaining what evidence and protections preserve credential validity. Boundaries are more visible than evidence standards; safeguards are uneven; and guidance is clearest when AI use resembles final-output substitution rather than feedback, access, verification, or professional workflow. The takeaway is that permission categories are necessary but insufficient. Universities need policies that make the certification logic visible: what learners may delegate, what they must still demonstrate, and how institutions will protect fair evidence rather than merely monitor AI use.

cs.CY↗

arXiv2026

SPRINT: Robust Model Attribution of Generated Images via Secret Pixel Reconstruction

Detecting the source model of AI-generated images is a growing accountability problem. AI fingerprinting techniques address this by detecting imperceptible patterns in the images that are unique to each model, achieving high detection accuracy under ideal conditions. However, recent research has shown that image fingerprints are extremely brittle to adaptive attacks, where knowledge of the technique can be exploited to perturb the fingerprints and evade detection. We present SPRINT (Secret Pixel Reconstruction fingerprinting), a novel model attribution method specifically designed to provide robustness to adaptive attacks. As opposed to existing fingerprinting, which focuses on publicly discoverable patterns in the image, SPRINT relies on a secret to define hidden reconstruction targets, thus keeping the verification task itself private. As a result, the attacker can no longer see the task that the verifier solves at verification time, protecting the information exploited by the attacks. Our results show that SPRINT achieves high closed-world accuracy while remaining robust to adaptive attacks: on the FFHQ dataset, SPRINT reaches 99.17% clean accuracy on a diverse 12-model pool and 98.83% on a harder pool of 6 close checkpoints of the same model architecture, while reducing adaptive removal and forgery attack success rates to 1% or below. When the same pool of close model checkpoints is considered an open world, SPRINT maintains high accuracy with an AUROC of 99.30%. These findings show that the approach of privatizing the verification task can make adaptive evasion substantially harder while maintaining performance in the clean setting.

cs.CR↗

arXiv2026

GAST: Gradient-aligned Sparse Tuning of Large Language Models with Data-layer Selection

Parameter-Efficient Fine-Tuning (PEFT) has become a key strategy for adapting large language models, with recent advances in sparse tuning reducing overhead by selectively updating key parameters or subsets of data. Existing approaches generally focus on two distinct paradigms: layer-selective methods aiming to fine-tune critical layers to minimize computational load, and data-selective methods aiming to select effective training subsets to boost training. However, current methods typically overlook the fact that different data points contribute varying degrees to distinct model layers, and they often discard potentially valuable information from data perceived as of low quality. To address these limitations, we propose Gradient-aligned Sparse Tuning (GAST), an innovative method that simultaneously performs selective fine-tuning at both data and layer dimensions as integral components of a unified optimization strategy. GAST specifically targets redundancy in information by employing a layer-sparse strategy that adaptively selects the most impactful data points for each layer, providing a more comprehensive and sophisticated solution than approaches restricted to a single dimension. Experiments demonstrate that GAST consistently outperforms baseline methods, establishing a promising direction for future research in PEFT strategies.

cs.LG↗

arXiv2026

NuNext: Reframing Nucleus Detection as Next-Point Detection

Nucleus detection in histopathology is pivotal for a wide range of clinical applications. Existing approaches either regress nuclear proxy maps that require complex post-processing, or employ dense anchors or queries that introduce severe foreground-background imbalance. In this work, we reformulate nucleus detection as next-point prediction, wherein a multimodal large language model is developed to directly output foreground nucleus centroids from the input image. The model is trained in two stages. In the supervised learning stage, we propose spatial-aware soft supervision to relax strict centroid matching and a chain-of-visual-thought strategy to incorporate visual priors that facilitate coordinate prediction. In the reinforcement fine-tuning stage, we design distribution matching reward, low-variance group filtering, and fine-grained advantage shaping to further improve the model's detection quality. Extensive experiments on nine widely used benchmarks demonstrate the superiority of our method. Code will be released soon.

cs.CV↗

arXiv2026

Smudged Fingerprints: A Systematic Evaluation of the Robustness of AI Image Fingerprints

Model fingerprint detection has shown promise to trace the provenance of AI-generated images in forensic applications. However, despite the inherent adversarial nature of these applications, existing evaluations rarely consider adversarial settings. We present the first systematic security evaluation of these techniques, formalizing threat models that encompass both white- and black-box access and two attack goals: fingerprint removal, which erases identifying traces to evade attribution, and fingerprint forgery, which seeks to cause misattribution to a target model. We implement five attack strategies and evaluate 14 representative fingerprinting methods across RGB, frequency, and learned-feature domains on 12 state-of-the-art image generators. Our experiments reveal a pronounced gap between clean and adversarial performance. Removal attacks are highly effective, often achieving success rates above 80% in white-box settings and over 50% under black-box access. While forgery is more challenging than removal, its success varies significantly across targeted models. We also observe a utility-robustness trade-off: accurate attribution methods are often vulnerable to attacks and, although some techniques are robust in specific settings, none achieves robustness and accuracy across all evaluated threat models. These findings highlight the need for techniques that balance robustness and accuracy, and we identify the most promising approaches toward this goal. Code available at: https://github.com/kaikaiyao/SmudgedFingerprints.

cs.CV↗

arXiv2025

Towards Instance-wise Personalized Federated Learning via Semi-Implicit Bayesian Prompt Tuning

Federated learning (FL) is a privacy-preserving machine learning paradigm that enables collaborative model training across multiple distributed clients without disclosing their raw data. Personalized federated learning (pFL) has gained increasing attention for its ability to address data heterogeneity. However, most existing pFL methods assume that each client's data follows a single distribution and learn one client-level personalized model for each client. This assumption often fails in practice, where a single client may possess data from multiple sources or domains, resulting in significant intra-client heterogeneity and suboptimal performance. To tackle this challenge, we propose pFedBayesPT, a fine-grained instance-wise pFL framework based on visual prompt tuning. Specifically, we formulate instance-wise prompt generation from a Bayesian perspective and model the prompt posterior as an implicit distribution to capture diverse visual semantics. We derive a variational training objective under the semi-implicit variational inference framework. Extensive experiments on benchmark datasets demonstrate that pFedBayesPT consistently outperforms existing pFL methods under both feature and label heterogeneity settings.

cs.LG↗

arXiv2025

Exploiting Layer Normalization Fine-tuning in Visual Transformer Foundation Models for Classification

LayerNorm is pivotal in Vision Transformers (ViTs), yet its fine-tuning dynamics under data scarcity and domain shifts remain underexplored. This paper shows that shifts in LayerNorm parameters after fine-tuning (LayerNorm shifts) are indicative of the transitions between source and target domains; its efficacy is contingent upon the degree to which the target training samples accurately represent the target domain, as quantified by our proposed Fine-tuning Shift Ratio ($FSR$). Building on this, we propose a simple yet effective rescaling mechanism using a scalar $λ$ that is negatively correlated to $FSR$ to align learned LayerNorm shifts with those ideal shifts achieved under fully representative data, combined with a cyclic framework that further enhances the LayerNorm fine-tuning. Extensive experiments across natural and pathological images, in both in-distribution (ID) and out-of-distribution (OOD) settings, and various target training sample regimes validate our framework. Notably, OOD tasks tend to yield lower $FSR$ and higher $λ$ in comparison to ID cases, especially with scarce data, indicating under-represented target training samples. Moreover, ViTFs fine-tuned on pathological data behave more like ID settings, favoring conservative LayerNorm updates. Our findings illuminate the underexplored dynamics of LayerNorm in transfer learning and provide practical strategies for LayerNorm fine-tuning.

cs.CV↗

arXiv2025

GradOT: Training-free Gradient-preserving Offsite-tuning for Large Language Models

The rapid growth of large language models (LLMs) with traditional centralized fine-tuning emerges as a key technique for adapting these models to domain-specific challenges, yielding privacy risks for both model and data owners. One promising solution, called offsite-tuning (OT), is proposed to address these challenges, where a weaker emulator is compressed from the original model and further fine-tuned with adapter to enhance privacy. However, the existing OT-based methods require high computational costs and lack theoretical analysis. This paper introduces a novel OT approach based on gradient-preserving compression, named GradOT. By analyzing the OT problem through the lens of optimization, we propose a method that selectively applies compression techniques such as rank compression and channel pruning, preserving the gradients of fine-tuned adapters while ensuring privacy. Extensive experiments demonstrate that our approach surpasses existing OT methods, both in terms of privacy protection and model performance. Our method provides a theoretical foundation for OT and offers a practical, training-free solution for offsite-tuning of large-scale LLMs.

cs.CL↗

arXiv2025

A Survey of Test-Time Compute: From Intuitive Inference to Deliberate Reasoning

The remarkable performance of the o1 model in complex reasoning demonstrates that test-time compute scaling can further unlock the model's potential, enabling powerful System-2 thinking. However, there is still a lack of comprehensive surveys for test-time compute scaling. We trace the concept of test-time compute back to System-1 models. In System-1 models, test-time compute addresses distribution shifts and improves robustness and generalization through parameter updating, input modification, representation editing, and output calibration. In System-2 models, it enhances the model's reasoning ability to solve complex problems through repeated sampling, self-correction, and tree search. We organize this survey according to the trend of System-1 to System-2 thinking, highlighting the key role of test-time compute in the transition from System-1 models to weak System-2 models, and then to strong System-2 models. We also point out advanced topics and future directions.

cs.AI↗

arXiv2025

Unlock Pose Diversity: Accurate and Efficient Implicit Keypoint-based Spatiotemporal Diffusion for Audio-driven Talking Portrait

Audio-driven single-image talking portrait generation plays a crucial role in virtual reality, digital human creation, and filmmaking. Existing approaches are generally categorized into keypoint-based and image-based methods. Keypoint-based methods effectively preserve character identity but struggle to capture fine facial details due to the fixed points limitation of the 3D Morphable Model. Moreover, traditional generative networks face challenges in establishing causality between audio and keypoints on limited datasets, resulting in low pose diversity. In contrast, image-based approaches produce high-quality portraits with diverse details using the diffusion network but incur identity distortion and expensive computational costs. In this work, we propose KDTalker, the first framework to combine unsupervised implicit 3D keypoint with a spatiotemporal diffusion model. Leveraging unsupervised implicit 3D keypoints, KDTalker adapts facial information densities, allowing the diffusion process to model diverse head poses and capture fine facial details flexibly. The custom-designed spatiotemporal attention mechanism ensures accurate lip synchronization, producing temporally consistent, high-quality animations while enhancing computational efficiency. Experimental results demonstrate that KDTalker achieves state-of-the-art performance regarding lip synchronization accuracy, head pose diversity, and execution efficiency.Our codes are available at https://github.com/chaolongy/KDTalker.

cs.CV↗

arXiv2025

SoK: What Makes Private Learning Unfair?

Differential privacy has emerged as the most studied framework for privacy-preserving machine learning. However, recent studies show that enforcing differential privacy guarantees can not only significantly degrade the utility of the model, but also amplify existing disparities in its predictive performance across demographic groups. Although there is extensive research on the identification of factors that contribute to this phenomenon, we still lack a complete understanding of the mechanisms through which differential privacy exacerbates disparities. The literature on this problem is muddled by varying definitions of fairness, differential privacy mechanisms, and inconsistent experimental settings, often leading to seemingly contradictory results. This survey provides the first comprehensive overview of the factors that contribute to the disparate effect of training models with differential privacy guarantees. We discuss their impact and analyze their causal role in such a disparate effect. Our analysis is guided by a taxonomy that categorizes these factors by their position within the machine learning pipeline, allowing us to draw conclusions about their interaction and the feasibility of potential mitigation strategies. We find that factors related to the training dataset and the underlying distribution play a decisive role in the occurrence of disparate impact, highlighting the need for research on these factors to address the issue.

cs.LG↗

arXiv2024

ScaleOT: Privacy-utility-scalable Offsite-tuning with Dynamic LayerReplace and Selective Rank Compression

Offsite-tuning is a privacy-preserving method for tuning large language models (LLMs) by sharing a lossy compressed emulator from the LLM owners with data owners for downstream task tuning. This approach protects the privacy of both the model and data owners. However, current offsite tuning methods often suffer from adaptation degradation, high computational costs, and limited protection strength due to uniformly dropping LLM layers or relying on expensive knowledge distillation. To address these issues, we propose ScaleOT, a novel privacy-utility-scalable offsite-tuning framework that effectively balances privacy and utility. ScaleOT introduces a novel layerwise lossy compression algorithm that uses reinforcement learning to obtain the importance of each layer. It employs lightweight networks, termed harmonizers, to replace the raw LLM layers. By combining important original LLM layers and harmonizers in different ratios, ScaleOT generates emulators tailored for optimal performance with various model scales for enhanced privacy protection. Additionally, we present a rank reduction method to further compress the original LLM layers, significantly enhancing privacy with negligible impact on utility. Comprehensive experiments show that ScaleOT can achieve nearly lossless offsite tuning performance compared with full fine-tuning while obtaining better model privacy.

cs.CL↗

arXiv2024

Layer-wise Importance Matters: Less Memory for Better Performance in Parameter-efficient Fine-tuning of Large Language Models

Parameter-Efficient Fine-Tuning (PEFT) methods have gained significant popularity for adapting pre-trained Large Language Models (LLMs) to downstream tasks, primarily due to their potential to significantly reduce memory and computational overheads. However, a common limitation in most PEFT approaches is their application of a uniform architectural design across all layers. This uniformity involves identical trainable modules and ignores the varying importance of each layer, leading to sub-optimal fine-tuning results. To overcome the above limitation and obtain better performance, we develop a novel approach, Importance-aware Sparse Tuning (IST), to fully utilize the inherent sparsity and select the most important subset of full layers with effective layer-wise importance scoring. The proposed IST is a versatile and plug-and-play technique compatible with various PEFT methods that operate on a per-layer basis. By leveraging the estimated importance scores, IST dynamically updates these selected layers in PEFT modules, leading to reduced memory demands. We further provide theoretical proof of convergence and empirical evidence of superior performance to demonstrate the advantages of IST over uniform updating strategies. Extensive experiments on a range of LLMs, PEFTs, and downstream tasks substantiate the effectiveness of our proposed method, showcasing IST's capacity to enhance existing layer-based PEFT methods. Our code is available at https://github.com/Kaiseem/IST.

cs.CL↗

arXiv2024

OPa-Ma: Text Guided Mamba for 360-degree Image Out-painting

In this paper, we tackle the recently popular topic of generating 360-degree images given the conventional narrow field of view (NFoV) images that could be taken from a single camera or cellphone. This task aims to predict the reasonable and consistent surroundings from the NFoV images. Existing methods for feature extraction and fusion, often built with transformer-based architectures, incur substantial memory usage and computational expense. They also have limitations in maintaining visual continuity across the entire 360-degree images, which could cause inconsistent texture and style generation. To solve the aforementioned issues, we propose a novel text-guided out-painting framework equipped with a State-Space Model called Mamba to utilize its long-sequence modelling and spatial continuity. Furthermore, incorporating textual information is an effective strategy for guiding image generation, enriching the process with detailed context and increasing diversity. Efficiently extracting textual features and integrating them with image attributes presents a significant challenge for 360-degree image out-painting. To address this, we develop two modules, Visual-textual Consistency Refiner (VCR) and Global-local Mamba Adapter (GMA). VCR enhances contextual richness by fusing the modified text features with the image features, while GMA provides adaptive state-selective conditions by capturing the information flow from global to local representations. Our proposed method achieves state-of-the-art performance with extensive experiments on two broadly used 360-degree image datasets, including indoor and outdoor settings.

cs.CV↗

arXiv2024

SCMix: Stochastic Compound Mixing for Open Compound Domain Adaptation in Semantic Segmentation

Open compound domain adaptation (OCDA) aims to transfer knowledge from a labeled source domain to a mix of unlabeled homogeneous compound target domains while generalizing to open unseen domains. Existing OCDA methods solve the intra-domain gaps by a divide-and-conquer strategy, which divides the problem into several individual and parallel domain adaptation (DA) tasks. Such approaches often contain multiple sub-networks or stages, which may constrain the model's performance. In this work, starting from the general DA theory, we establish the generalization bound for the setting of OCDA. Built upon this, we argue that conventional OCDA approaches may substantially underestimate the inherent variance inside the compound target domains for model generalization. We subsequently present Stochastic Compound Mixing (SCMix), an augmentation strategy with the primary objective of mitigating the divergence between source and mixed target distributions. We provide theoretical analysis to substantiate the superiority of SCMix and prove that the previous methods are sub-groups of our methods. Extensive experiments show that our method attains a lower empirical risk on OCDA semantic segmentation tasks, thus supporting our theories. Combining the transformer architecture, SCMix achieves a notable performance boost compared to the SoTA results.

cs.CV↗