arXiv subjects

Feida Zhu

Publications and source records attributed to Feida Zhu.

At least 19 recordsLinked to original sources

arXiv2026

Omni-Decision: Evidence-Ledger Planning for Omni-Modal Agents

Omni-modal agents must seek evidence across video, audio, web pages, and computation to answer questions. Their main bottleneck is planning: noisy multimodal observations accumulate in conversation history and disrupt later decisions, while multimodal models have limited capacity for multi-step planning. Controlled backend replacements support this diagnosis: replacing the planner causes a much larger performance loss than replacing the perception backend. We present Omni-Decision, an omni-modal agent built on evidence-ledger planning: it replaces the growing dialogue history with an explicit evidence ledger that records what evidence is still missing, what has been confirmed, and where records conflict. A critic reads each noisy observation and passes only the usable content to the ledger, discarding the rest, so the planner works from a compact context throughout the task. Each run records the state, action, and verdict at every step, and supervised fine-tuning and decision-level reinforcement learning on these trajectories further improve the planner. Omni-Decision achieves state-of-the-art accuracy of 81.4% on OmniGAIA at approximately 43% of Gemini-3.1-Pro's cost per question, and 65.0% on WorldSense long-video understanding, level with the strongest end-to-end model.

cs.AI↗

arXiv2026

ProCredit: From Outcome Rewards to Progress Credit in Agentic Reinforcement Learning

Long-horizon agentic tasks require an agent to modify an environment through a sequence of tool calls, with success determined by the final state. The standard recipe assigns a single outcome reward at the end and compares trajectories sampled for the same task. As a result, a group with no successful trajectory yields no training signal, failed attempts cannot be told apart by how close they came to completion, and turns that advance the task receive the same credit as turns that only query the environment. Prior work refines the unit of comparison from the trajectory to the step, or trains a reward model to supply intermediate signal: the former still derives its signal from final success alone, and the latter estimates it with a model. We observe that the acceptance checks that decide success can also be run on intermediate states, so progress is as verifiable as the outcome. We propose ProCredit, which turns this verified progress into credit: it reruns the acceptance checks after each turn, rewards the turn by its change in progress, and uses these rewards to assign credit both across attempts at the same task and across the turns within a trajectory. Starting from Qwen3.5 base models at three scales on AppWorld, ProCredit outperforms outcome-reward baselines and progress-based baselines in task completion rate at every scale on both test sets, exceeding the strongest outcome-reward baseline by 4.1 percentage points at 4B, and results in a second environment show the same direction of improvement. Ablations show that adding the final progress to the trajectory score alone does not improve performance: the gain comes from crediting progress to the turn where it occurs.

cs.LG↗

arXiv2026

Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents

The rapid progression of large language models is extending AI from passive content generation into the active workflows of engineering and scientific discovery. This shift raises a compelling question: can AI be both the object of development and an active participant in building next-generation AI systems? We explore this question by building Qwen-Planner-Agent within a closed-loop AI-for-AI framework for scalable development and iterative improvement. Mobile planning offers a demanding test of this approach: complex, long-horizon tasks challenge agent reliability, while costly real-device interaction limits development scalability. The framework connects data production, model training, and deployment through a shared action-feedback-verification contract. (i) AI for Data builds a human-gated agentic data flywheel in which specialized agents construct tasks, collect interaction trajectories, curate and balance training data, and use training feedback to guide subsequent data generation. (ii) AI for Training combines a supervised planning cold start with hybrid-environment online agentic reinforcement learning, where we introduce Competence-Aware Reward-and-Advantage Engineering (CARE) to reduce reasoning and tool-use costs while preserving task performance. (iii) AI drives model--harness co-evolution through an execution-evidence-driven loop that orchestrates memory, skills, and tools at runtime and feeds structured action feedback and preserved failure traces back into coordinated model and harness adaptation. Qwen-Planner-Agent achieves the best overall performance among all evaluated models and systems on MobilePA-Bench, improving over its base model across tool use, memory, skills, and sub-agent coordination. Further evaluations of our model show improvements across non-mobile agentic benchmarks while largely preserving general capabilities.

cs.AI↗

arXiv2026

Block-Sparse Attention with Semantic-Geometric Decoupled Routing

Long-context inference has become a defining capability of large language models, but exact dense attention remains costly due to its quadratic scaling with sequence length. Block-sparse attention offers a hardware-friendly alternative by routing each query block to a small set of relevant key blocks, yet accurate training-free block routing remains difficult. Existing routers often pool post-RoPE token representations, which entangles semantic aggregation with RoPE-induced geometry and attenuates local positional cues through high-frequency phase cancellation. To resolve this mismatch, we propose \textbf{Semantic-Geometric Decoupled Routing}, a training-free block routing framework that shifts semantic aggregation to the pre-RoPE space and reconstructs geometric bias with an offline structural prior and relative block distances. This decomposition yields an explicit closed-form block routing score without token-level search or post-hoc calibration. Experiments on long-context text and video tasks show that our method approaches full-attention accuracy across 4K--128K contexts, keeps routing overhead below 3.4 ms, and achieves a 5.03$\times$ speedup over FlashAttn at a 128K context length.

cs.CL↗

arXiv2026

MobilePA-Bench: Benchmarking Mobile Planner Agents on Complex Real-World Tasks

As on-device LLM agents evolve into personal copilots, the mobile operating system has become a key testbed for this paradigm, making rigorous capability evaluation essential. Yet existing benchmarks fall into two camps, each with a critical blind spot: GUI-centric benchmarks test surface-level screen manipulation while overlooking background tool use and long-horizon planning, whereas static function-calling benchmarks rely on offline API matching that is detached from real runtime constraints. To close this gap, we present \textbf{MobilePA-Bench}, an interactive, stateful, and tool-centric benchmark for evaluating the tool-calling and planning abilities of mobile planning agents. MobilePA-Bench runs on an executable sandbox that maintains live application databases and returns structured feedback, spanning $13$ functional domains and $212$ realistic mobile tools. Beyond basic tool use, it evaluates a central planning agent along three advanced dimensions: \emph{(1)~Sub-agent Collaboration}---decomposing a complex task and delegating specialized work to capable sub-agents; \emph{(2)~Memory Usage}---recalling stored memories, user profiles, and past preferences to resolve implicit requests; and \emph{(3)~Skill Usage}---invoking pre-packaged composite skills instead of planning every step from scratch. Extensive experiments show that current frontier LLMs remain unreliable in mobile settings: performance drops sharply under strict tool ordering, permission limits, and unexpected runtime errors. By pairing an interactive function-calling sandbox with evidence-based verification, MobilePA-Bench serves as both a practical diagnostic benchmark and an interactive foundation for agentic reinforcement learning---accelerating the development of dependable mobile agents.

cs.AI↗

arXiv2026

Covering the Unseen: Information Demand Coverage Optimization for Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) typically treats context selection as ranking chunks against a single query embedding. This assumption breaks down for complex queries, such as multi-hop or ambiguous questions, where top-k selection tends to over-cover one semantic aspect while ignoring critical sub-questions. We propose GeoRAG, which recasts context selection as Information Demand Coverage Optimization. GeoRAG builds a multi-dimensional demand distribution through diverse sub-query generation and reverse-validation weighting, then selects context by minimizing the Sinkhorn-Wasserstein distance between this demand distribution and the coverage of the selected set. The resulting demand-weighted facility-location objective is monotone submodular, giving a $1-1/e$ greedy guarantee, which we approximate with a Sinkhorn-based marginal-gain surrogate. The method is unsupervised, training-free, and retrieval-agnostic. We further show that single-point, query-proximity scorers cannot cover multi-modal demands, exposing a structural limit of ranking-based selection. On six open-domain QA benchmarks, GeoRAG improves exact match (EM) by +6.5 to +7.5 points over top-k truncation (up to +9.7 on HotpotQA and ASQA) and outperforms strong baselines including MMR, DPP, BGE-Reranker, SMART-RAG, and AdaGReS, with stable gains across context budgets and sub-query generators.

cs.IR↗

arXiv2026

LineageMark: Multi-user White-box Watermarking for Contribution Tracing in Model Derivation Chains

In open large language model (LLM) ecosystems, models are frequently adapted across multiple domains and applications, forming multi-stage derivation chains. Consequently, tracking and verifying historical contributions is essential for model provenance and intellectual property protection. However, existing watermarking methods are mainly designed for single-user, one-time embeddings, often fail under repeated model derivation and incremental updates. To address this problem, we propose LineageMark, a multi-user white-box watermarking framework for model derivation chains. The framework encodes watermarks in model parameters using a projection-based approach. Stable carriers are first selected to reduce sensitivity to model changes, each watermark bit is then represented as a projection statistic over these carriers. Additional watermark insertions introduce only bounded perturbations in the projection space, and margin constraints are used to maintain signal integrity. We evaluate the effectiveness of LineageMark in multi-stage model derivation chains. Experimental results show that LineageMark preserves contributor watermarks across multi-stage derivation and supports incremental multi-user watermark insertion. Furthermore, it exhibits robustness against perturbations such as re-watermarking, fine-tuning, quantization, and pruning.

cs.CR↗

arXiv2026

Architecture-Agnostic Feature Synergy for Universal Defense Against Heterogeneous Generative Threats

Generative AI deployment poses unprecedented challenges to content safety and privacy. However, existing defense mechanisms are often tailored to specific architectures (e.g., Diffusion Models or GANs), creating fragile "defense silos" that fail against heterogeneous generative threats. This paper identifies a fundamental optimization barrier in naive pixel-space ensemble strategies: due to divergent objective functions, pixel-level gradients from heterogeneous generators become statistically orthogonal, causing destructive interference. To overcome this, we observe that despite disparate low-level mechanisms, high-level feature representations of generated content exhibit alignment across architectures. Based on this, we propose the Architecture-Agnostic Targeted Feature Synergy (ATFS) framework. By introducing a target guidance image, ATFS reformulates multi-model defense as a unified feature space alignment task, enabling intrinsic gradient alignment without complex rectification. Extensive experiments show ATFS achieves SOTA protection in heterogeneous scenarios (e.g., Diffusion+GAN). It converges rapidly, reaching over 90% performance within 40 iterations, and maintains strong attack potency even under tight perturbation budgets. The framework seamlessly extends to unseen architectures (e.g., VQ-VAE) by switching the feature extractor, and demonstrates robust resistance to JPEG compression and scaling. Being computationally efficient and lightweight, ATFS offers a viable pathway to dismantle defense silos and enable universal generative security. Code and models are open-sourced for reproducibility.

cs.CR↗

arXiv2025

Proof-of-Data: A Consensus Protocol for Collaborative Intelligence

Existing research on federated learning has been focused on the setting where learning is coordinated by a centralized entity. Yet the greatest potential of future collaborative intelligence would be unleashed in a more open and democratized setting with no central entity in a dominant role, referred to as "decentralized federated learning". New challenges arise accordingly in achieving both correct model training and fair reward allocation with collective effort among all participating nodes, especially with the threat of the Byzantine node jeopardising both tasks. In this paper, we propose a blockchain-based decentralized Byzantine fault-tolerant federated learning framework based on a novel Proof-of-Data (PoD) consensus protocol to resolve both the "trust" and "incentive" components. By decoupling model training and contribution accounting, PoD is able to enjoy not only the benefit of learning efficiency and system liveliness from asynchronous societal-scale PoW-style learning but also the finality of consensus and reward allocation from epoch-based BFT-style voting. To mitigate false reward claims by data forgery from Byzantine attacks, a privacy-aware data verification and contribution-based reward allocation mechanism is designed to complete the framework. Our evaluation results show that PoD demonstrates performance in model training close to that of the centralized counterpart while achieving trust in consensus and fairness for reward allocation with a fault tolerance ratio of 1/3.

cs.CR↗

arXiv2025

New Recipe for Semi-supervised Community Detection: Clique Annealing under Crystallization Kinetics

Semi-supervised community detection methods are widely used for identifying specific communities due to the label scarcity. Existing semi-supervised community detection methods typically involve two learning stages learning in both initial identification and subsequent adjustment, which often starts from an unreasonable community core candidate. Moreover, these methods encounter scalability issues because they depend on reinforcement learning and generative adversarial networks, leading to higher computational costs and restricting the selection of candidates. To address these limitations, we draw a parallel between crystallization kinetics and community detection to integrate the spontaneity of the annealing process into community detection. Specifically, we liken community detection to identifying a crystal subgrain (core) that expands into a complete grain (community) through a process similar to annealing. Based on this finding, we propose CLique ANNealing (CLANN), which applies kinetics concepts to community detection by integrating these principles into the optimization process to strengthen the consistency of the community core. Subsequently, a learning-free Transitive Annealer was employed to refine the first-stage candidates by merging neighboring cliques and repositioning the community core, enabling a spontaneous growth process that enhances scalability. Extensive experiments on \textbf{43} different network settings demonstrate that CLANN outperforms state-of-the-art methods across multiple real-world datasets, showcasing its exceptional efficacy and efficiency in community detection.

cs.SI↗

arXiv2025

Seed1.5-VL Technical Report

We present Seed1.5-VL, a vision-language foundation model designed to advance general-purpose multimodal understanding and reasoning. Seed1.5-VL is composed with a 532M-parameter vision encoder and a Mixture-of-Experts (MoE) LLM of 20B active parameters. Despite its relatively compact architecture, it delivers strong performance across a wide spectrum of public VLM benchmarks and internal evaluation suites, achieving the state-of-the-art performance on 38 out of 60 public benchmarks. Moreover, in agent-centric tasks such as GUI control and gameplay, Seed1.5-VL outperforms leading multimodal systems, including OpenAI CUA and Claude 3.7. Beyond visual and video understanding, it also demonstrates strong reasoning abilities, making it particularly effective for multimodal reasoning challenges such as visual puzzles. We believe these capabilities will empower broader applications across diverse tasks. In this report, we mainly provide a comprehensive review of our experiences in building Seed1.5-VL across model design, data construction, and training at various stages, hoping that this report can inspire further research. Seed1.5-VL is now accessible at https://www.volcengine.com/ (Volcano Engine Model ID: doubao-1-5-thinking-vision-pro-250428)

cs.CV↗

arXiv2025

Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled yet Hard-to-Learn Samples in Noisy Data

We propose a novel sample selection method for image classification in the presence of noisy labels. Existing methods typically consider small-loss samples as correctly labeled. However, some correctly labeled samples are inherently difficult for the model to learn and can exhibit high loss similar to mislabeled samples in the early stages of training. Consequently, setting a threshold on per-sample loss to select correct labels results in a trade-off between precision and recall in sample selection: a lower threshold may miss many correctly labeled hard-to-learn samples (low recall), while a higher threshold may include many mislabeled samples (low precision). To address this issue, our goal is to accurately distinguish correctly labeled yet hard-to-learn samples from mislabeled ones, thus alleviating the trade-off dilemma. We achieve this by considering the trends in model prediction confidence rather than relying solely on loss values. Empirical observations show that only for correctly labeled samples, the model's prediction confidence for the annotated labels typically increases faster than for any other classes. Based on this insight, we propose tracking the confidence gaps between the annotated labels and other classes during training and evaluating their trends using the Mann-Kendall Test. A sample is considered potentially correctly labeled if all its confidence gaps tend to increase. Our method functions as a plug-and-play component that can be seamlessly integrated into existing sample selection techniques. Experiments on several standard benchmarks and real-world datasets demonstrate that our method enhances the performance of existing methods for learning with noisy labels.

cs.CV↗

arXiv2025

Data Assetization via Resources-decoupled Federated Learning

With the development of the digital economy, data is increasingly recognized as an essential resource for both work and life. However, due to privacy concerns, data owners tend to maximize the value of data through the circulation of information rather than direct data transfer. Federated learning (FL) provides an effective approach to collaborative training models while preserving privacy. However, as model parameters and training data grow, there are not only real differences in data resources between different data owners, but also mismatches between data and computing resources. These challenges lead to inadequate collaboration among data owners, compute centers, and model owners, reducing the global utility of the three parties and the effectiveness of data assetization. In this work, we first propose a framework for resource-decoupled FL involving three parties. Then, we design a Tripartite Stackelberg Model and theoretically analyze the Stackelberg-Nash equilibrium (SNE) for participants to optimize global utility. Next, we propose the Quality-aware Dynamic Resources-decoupled FL algorithm (QD-RDFL), in which we derive and solve the optimal strategies of all parties to achieve SNE using backward induction. We also design a dynamic optimization mechanism to improve the optimal strategy profile by evaluating the contribution of data quality from data owners to the global model during real training. Finally, our extensive experiments demonstrate that our method effectively encourages the linkage of the three parties involved, maximizing the global utility and value of data assets.

cs.LG↗

arXiv2024

PonziLens+: Visualizing Bytecode Actions for Smart Ponzi Scheme Identification

With the prevalence of smart contracts, smart Ponzi schemes have become a common fraud on blockchain and have caused significant financial loss to cryptocurrency investors in the past few years. Despite the critical importance of detecting smart Ponzi schemes, a reliable and transparent identification approach adaptive to various smart Ponzi schemes is still missing. To fill the research gap, we first extract semantic-meaningful actions to represent the execution behaviors specified in smart contract bytecodes, which are derived from a literature review and in-depth interviews with domain experts. We then propose PonziLens+, a novel visual analytic approach that provides an intuitive and reliable analysis of Ponzi-scheme-related features within these execution behaviors. PonziLens+ has three visualization modules that intuitively reveal all potential behaviors of a smart contract, highlighting fraudulent features across three levels of detail. It can help smart contract investors and auditors achieve confident identification of any smart Ponzi schemes. We conducted two case studies and in-depth user interviews with 12 domain experts and common investors to evaluate PonziLens+. The results demonstrate the effectiveness and usability of PonziLens+ in achieving an effective identification of smart Ponzi schemes.

cs.HC↗

arXiv2024

A Full-History Network Dataset for BTC Asset Decentralization Profiling

Since its advent in 2009, Bitcoin (BTC) has garnered increasing attention from both academia and industry. However, due to the massive transaction volume, no systematic study has quantitatively measured the asset decentralization degree specifically from a network perspective. In this paper, by conducting a thorough analysis of the BTC transaction network, we first address the significant gap in the availability of full-history BTC graph and network property dataset, which spans over 15 years from the genesis block (1st March, 2009) to the 845651-th block (29, May 2024). We then present the first systematic investigation to profile BTC's asset decentralization and design several decentralization degrees for quantification. Through extensive experiments, we emphasize the significant role of network properties and our network-based decentralization degree in enhancing Bitcoin analysis. Our findings demonstrate the importance of our comprehensive dataset and analysis in advancing research on Bitcoin's transaction dynamics and decentralization, providing valuable insights into the network's structure and its implications.

q-fin.ST↗

arXiv2024

Knowledge Enhanced Multi-intent Transformer Network for Recommendation

Incorporating Knowledge Graphs into Recommendation has attracted growing attention in industry, due to the great potential of KG in providing abundant supplementary information and interpretability for the underlying models. However, simply integrating KG into recommendation usually brings in negative feedback in industry, due to the ignorance of the following two factors: i) users' multiple intents, which involve diverse nodes in KG. For example, in e-commerce scenarios, users may exhibit preferences for specific styles, brands, or colors. ii) knowledge noise, which is a prevalent issue in Knowledge Enhanced Recommendation (KGR) and even more severe in industry scenarios. The irrelevant knowledge properties of items may result in inferior model performance compared to approaches that do not incorporate knowledge. To tackle these challenges, we propose a novel approach named Knowledge Enhanced Multi-intent Transformer Network for Recommendation (KGTN), comprising two primary modules: Global Intents Modeling with Graph Transformer, and Knowledge Contrastive Denoising under Intents. Specifically, Global Intents with Graph Transformer focuses on capturing learnable user intents, by incorporating global signals from user-item-relation-entity interactions with a graph transformer, meanwhile learning intent-aware user/item representations. Knowledge Contrastive Denoising under Intents is dedicated to learning precise and robust representations. It leverages intent-aware representations to sample relevant knowledge, and proposes a local-global contrastive mechanism to enhance noise-irrelevant representation learning. Extensive experiments conducted on benchmark datasets show the superior performance of our proposed method over the state-of-the-arts. And online A/B testing results on Alibaba large-scale industrial recommendation platform also indicate the real-scenario effectiveness of KGTN.

cs.IR↗

arXiv2023

WarpDiffusion: Efficient Diffusion Model for High-Fidelity Virtual Try-on

Image-based Virtual Try-On (VITON) aims to transfer an in-shop garment image onto a target person. While existing methods focus on warping the garment to fit the body pose, they often overlook the synthesis quality around the garment-skin boundary and realistic effects like wrinkles and shadows on the warped garments. These limitations greatly reduce the realism of the generated results and hinder the practical application of VITON techniques. Leveraging the notable success of diffusion-based models in cross-modal image synthesis, some recent diffusion-based methods have ventured to tackle this issue. However, they tend to either consume a significant amount of training resources or struggle to achieve realistic try-on effects and retain garment details. For efficient and high-fidelity VITON, we propose WarpDiffusion, which bridges the warping-based and diffusion-based paradigms via a novel informative and local garment feature attention mechanism. Specifically, WarpDiffusion incorporates local texture attention to reduce resource consumption and uses a novel auto-mask module that effectively retains only the critical areas of the warped garment while disregarding unrealistic or erroneous portions. Notably, WarpDiffusion can be integrated as a plug-and-play component into existing VITON methodologies, elevating their synthesis quality. Extensive experiments on high-resolution VITON benchmarks and an in-the-wild test set demonstrate the superiority of WarpDiffusion, surpassing state-of-the-art methods both qualitatively and quantitatively.

cs.CV↗

arXiv2023

From Asset Flow to Status, Action and Intention Discovery: Early Malice Detection in Cryptocurrency

Cryptocurrency has been subject to illicit activities probably more often than traditional financial assets due to the pseudo-anonymous nature of its transacting entities. An ideal detection model is expected to achieve all three critical properties of (I) early detection, (II) good interpretability, and (III) versatility for various illicit activities. However, existing solutions cannot meet all these requirements, as most of them heavily rely on deep learning without interpretability and are only available for retrospective analysis of a specific illicit type. To tackle all these challenges, we propose Intention-Monitor for early malice detection in Bitcoin (BTC), where the on-chain record data for a certain address are much scarcer than other cryptocurrency platforms. We first define asset transfer paths with the Decision-Tree based feature Selection and Complement (DT-SC) to build different feature sets for different malice types. Then, the Status/Action Proposal Module (S/A-PM) and the Intention-VAE module generate the status, action, intent-snippet, and hidden intent-snippet embedding. With all these modules, our model is highly interpretable and can detect various illegal activities. Moreover, well-designed loss functions further enhance the prediction speed and model's interpretability. Extensive experiments on three real-world datasets demonstrate that our proposed algorithm outperforms the state-of-the-art methods. Furthermore, additional case studies justify our model can not only explain existing illicit patterns but can also find new suspicious characters.

cs.LG↗