As instruction-based editing models and multimodal large language models continue to advance, a wide range of image editing tasks have become feasible. However, achieving precise and consistent geometric image editing, such as translating, scaling and rotating in 3D space, remains a major challenge. In this work, we introduce BoxCtrl, a 3D-aware visual prompting framework to address this problem. Unlike text-only or coarse 2D-guided approaches, our method introduces informative RGB 3D bounding boxes projected onto 2D images as a visual prompting. Each box is outlined and its three orthogonal faces are individually painted with different RGB colors, which simultaneously encode its position, size and orientation, providing a compact and intuitive in-context visual example. The key to BoxCtrl’s success lies in its well-designed bounding boxes, which serve as visual prompting that decouple geometric control from appearance control. This design enables the model to learn consistent correspondences between faces with the same color in the latent space, leading to a more precise understanding of geometric intention and more accurate editing result. We introduce a two-stage training paradigm comprising Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL). To address the scarcity of paired training data, we first construct a large-scale synthetic dataset for the SFT stage, which equips the model with fundamental editing capabilities. However, the generalization of the SFT model is constrained by the synthetic-to-real domain gap. To bridge this, we incorporate an online RL stage leveraging unpaired real-world data. Guided by a comprehensive reward function evaluating both geometric accuracy and visual fidelity, our SFT-RL strategy significantly enhances geometric precision while maintaining photorealistic quality. Extensive experiments demonstrate that BoxCtrl achieves state-of-the-art performance across translation, rotation, scaling, and composite editing tasks. Our code is available at https://github.com/beaglew/BoxCtrl.
论文检索
输入标题、作者或关键词，从 8,741 篇学术成果中精准定位
Artistic styles often embed abstraction beyond surface appearance, involving deliberate reinterpretation of structure rather than mere changes in texture or color. Conventional style transfer methods typically preserve the input geometry and therefore struggle to capture this deeper abstraction behavior, especially for illustrative and non-photorealistic styles. In this work, we introduce Abstraction in Style (AiS), a generative framework that separates structural abstraction from visual stylization. Given a target image and a small set of style exemplars, AiS first derives an intermediate abstraction proxy that reinterprets the target’s structure in accordance with the abstraction logic exhibited by the style. The proxy captures semantic structure while relaxing geometric fidelity, enabling subsequent stylization to operate on an abstracted representation rather than the original image. In a second stage, the abstraction proxy is rendered to produce the final stylized output, preserving visual coherence with the reference style. Both stages are implemented using a shared image-space analogy, enabling transformations to be learned from visual exemplars without explicit geometric supervision. By decoupling abstraction from appearance and treating abstraction as an explicit, transferable process, AiS supports a wider range of stylistic transformations, improves controllability, and enables more expressive stylization.
Global illumination (GI) is essential for realism but remains computationally expensive. While per-scene neural methods lack generalization and screen-space approaches inherently suffer from view inconsistency, prior 3D neural rendering methods face a severe scalability barrier, restricting them to small, object-centric meshes. To overcome these trade-offs, we introduce a generalizable light transport 3D embedding that predicts global illumination directly from 3D scene configurations without rasterized or path-traced illumination cues, per-scene retraining or screen-space limitations. We employ a point-based representation to decouple our embedding from the original scene topology, then utilize a linear-complexity transformer to encode long-range light transport. This design scales to environments with millions of triangles, enabling the first generalizable GI learning on complex, high-fidelity indoor scenes, far beyond prior limits. To achieve this, we enforce a local query mechanism where rendering queries are processed independently under 3D supervision. This ensures constant complexity per pixel relative to scene size, yielding view-consistent and resolution-agnostic rendering without the memory bottlenecks typical of globally coupled attention. We further demonstrate versatility by re-targeting the encoder with limited fine-tuning, presenting preliminary results on spatial-directional radiance field prediction for glossy materials and validating transfer to downstream rendering tasks.
Signed Distance Fields (SDFs) are a powerful volumetric representation for 3D geometry. Recent advances in surface generation from SDFs increasingly rely on learnable surface representations and direct mesh supervision. In this work, we challenge this trend and show that high-quality surface reconstruction can be achieved by learning to refine the volumetric signal itself. We present SuperSDF, a learning-based approach for sparse SDF super-resolution that operates directly in SDF space, without auxiliary surface representations or mesh-level supervision. Using a sparse voxel neural network restricted to a narrow band near the surface, our method predicts high-resolution signed-distance values from coarse inputs in a scalable and resolution-agnostic manner. Standard isosurface extraction algorithms can then process the resulting super-resolved SDFs to produce accurate, detailed surface meshes. Our results show that learning-based SDF upsampling alone is sufficient to recover fine geometric details missed by classical interpolation and prior reconstruction methods. Compared to state-of-the-art machine learning approaches, our method generates higher-fidelity surfaces at a fraction of the computational cost and scales to volumetric resolutions that were previously impractical.
Large Language Models (LLMs) are now widely deployed across modern web services, but their safe and trustworthy use in real-world settings critically depends on accurate alignment with human preferences. Preference alignment is typically achieved using methods such as reinforcement learning or direct preference optimization (DPO), whose effectiveness in practice hinges on the quality of labeled preference data. However, a fundamental practical challenge remains: preference datasets inevitably contain noise. Through a systematic analysis of mainstream preference datasets, we find that roughly 25% of preference pairs show clear inconsistencies between reward-model evaluations and human annotations. Such inconsistent examples do not convey reliable preference signals; training directly on them therefore not only fails to improve alignment but can even degrade model behavior. To address this problem, we propose Noise-Aware Preference Alignment for LLMs via Confidence and Polarity Reweighting (AlignCP), a fully automated, human-free framework for noise-aware preference alignment. AlignCP derives two interpretable metrics from reward-model outputs: Confidence, which measures the reliability of each preference judgment, and Polarity, which evaluates whether the reward-model ranking agrees with the original human label. These metrics are combined to assign a training weight to each sample—amplifying high-confidence, label-consistent pairs while down-weighting or discarding low-confidence, contradictory, or otherwise noisy pairs. Unlike methods that rely on human re-inspection, repeated relabeling, or heavy computational reruns, AlignCP performs automated data-quality control with minimal overhead and no human intervention. Empirical results show that AlignCP substantially outperforms existing data-centric alignment approaches on standard preference benchmarks and remains more robust under noisy supervision.
As large language models (LLMs) increasingly shape content generation, interaction, and decision-making across the Web, aligning them with human values has become a central objective in trustworthy AI. This challenge becomes even more pronounced when aligning multiple, potentially conflicting human values. Although recent approaches, such as reward reweighting, prompt-based supervised fine-tuning, and model merging, attempt to tackle multi-value alignment, they still face two major limitations: (1) training separate models for each value combination is prohibitively expensive; (2) value conflicts substantially degrade alignment performance. These limitations make it difficult to achieve favorable trade-offs across diverse human values. To address these challenges, we revisit multi-value alignment from the perspective of value consistency in data and propose VC-soup, a data filtering and parameter merging framework grounded in value-consistent learning. We first design a value consistency metric based on the cosine similarity between the reward-gap vector of each preference pair and an all-ones vector, which quantifies its cross-value coherence. We then filter out low-consistency preference pairs in each value dataset and train on the remaining data to obtain smooth, value-consistent policy models that better preserve linear mode connectivity. Finally, we linearly combine these policies and apply Pareto filtering across values to obtain solutions with balanced multi-value performance. Extensive experiments and theoretical analysis demonstrate that VC-soup effectively mitigates conflicts and consistently outperforms existing multi-value alignment methods.
Large Language Models (LLMs) perform well across many tasks but degrade when processing large collections of repetitive or highly similar inputs, a common scenario in applications such as near-duplicate search results and large e-commerce catalogs. In these settings, concatenation-based approaches—long-context prompting and supervised fine-tuning—suffer from attention saturation and diminished signal-to-noise ratio, causing models to miss subtle but important distinctions as input size grows. We introduce SEER (Set Encoding for Efficient Representation), a framework that enables LLMs to handle massive sets of near-duplicate items through a single learned token. SEER first encodes individual items with a pretrained embedding model, then aggregates them using an adapter that captures inter-item relationships and preserves fine-grained differences while mitigating redundancy. To ensure both discriminative and generative capabilities, we propose a multi-task alignment strategy that supervises set-level descriptions across multiple semantic dimensions. Experiments on a large-scale e-commerce dataset demonstrate that SEER substantially outperforms in-context and fine-tuned LLM baselines, maintaining stable performance even when processing thousands of highly similar items. These results establish SEER as an effective and scalable approach for LLM processing of dense, redundant input sets.
Tabular Foundation Models (TFMs) have recently shown strong in-context learning capabilities on structured data, achieving zero-shot performance comparable to traditional machine learning methods. We find that zero-shot TFMs already achieve strong performance, while the benefits of fine-tuning are highly model- and data-dependent. Meta-learning and PEFT provide moderate gains under specific conditions, whereas full supervised fine-tuning often reduces accuracy or calibration quality. This work presents the first comprehensive study of fine-tuning in TFMs across benchmarks including TALENT, OpenML-CC18, and TabZilla. We compare zero-shot, meta-learning, supervised (SFT), and parameter-efficient (PEFT) approaches, analyzing how dataset factors such as imbalance, size, and dimensionality affect outcomes. Our findings cover performance, calibration, and fairness, offering practical guidelines on when fine-tuning is most beneficial and its limitations.
Accurate estimation of delivery time (EDT) is a critical factor in web e-commerce user experience. The pursuit of higher EDT accuracy has predominantly centered on designing increasingly complex model architectures. While valuable, this architecture-centric paradigm creates a tension between its high iteration costs and the industrial demand for agile deployment. This work, therefore, explores a complementary dimension: enhancing model performance by optimizing the learning process itself. We propose EDTF, a novel, plug-and-play composite learning framework that empowers existing models by augmenting their learning objective. EDTF first transforms the traditional regression problem into a structured ordinal classification task to address the training difficulties inherent in direct regression and preserve temporal order. It then introduces a cross-view consistency paradigm, decomposing the prediction task into two related views: the macroscopic end-to-end delivery time and the microscopic next-hop duration. By enforcing a self-supervised signal that aligns the sum of future next-hop durations with the overall EDT, our framework enables models to learn more robust temporal representations without extra features. Extensive experiments on a large-scale industrial dataset show that EDTF, as a plugin, consistently enhances performance and accelerates convergence across five diverse architectures. Critically, an EDTF-optimized model has been successfully deployed in a live production environment, demonstrating significant improvements over its predecessor. This work thus presents a validated and valuable new paradigm for the economical and efficient application of web services reliant on trajectory-based forecasting, from e-commerce to ride-hailing and food delivery.
Inspired by advances in LLMs, reasoning-enhanced sequential recommendation performs multi-step deliberation before making final predictions, unlocking greater potential for capturing user preferences. However, current methods are constrained by static reasoning trajectories that are ill-suited for the diverse complexity of user behaviors. They suffer from two key limitations: (1) a static reasoning direction, which uses flat supervision signals misaligned with human-like hierarchical reasoning, and (2) a fixed reasoning depth, which inefficiently applies the same computational effort to all users, regardless of pattern complexity. These rigidity lead to suboptimal performance and significant computational waste. To overcome these challenges, we propose DTRec, a novel and effective framework that explores the Dynamic reasoning Trajectory for Sequential Recommendation along both direction and depth. To guide the direction, we develop Hierarchical Process Supervision (HPS), which provides coarse-to-fine supervisory signals to emulate the natural, progressive refinement of human cognitive processes. To optimize the depth, we introduce the Adaptive Reasoning Halting (ARH) mechanism that dynamically adjusts the number of reasoning steps by jointly monitoring three indicators. Extensive experiments on three real-world datasets demonstrate the superiority of our approach, achieving up to a 24.5% performance improvement over strong baselines while simultaneously reducing computational cost by up to 41.6%.
Optimizing numerical systems and mechanism design is crucial for enhancing player experience in Massively Multiplayer Online (MMO) games. Traditional optimization approaches rely on iterative online experiments or parameter tuning over abstracted statistical models, which can be inaccurate, time-consuming and potentially impair players' experience. Although simplified offline simulation systems are frequently employed as alternatives, their low fidelity constrains agents' ability to faithfully replicate real players' reasoning processes and behavioral responses to interventions. To address these limitations, we propose a generative agent-based MMO simulation system with hundreds of agents empowered by Large Language Models (LLMs). By applying Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on large-scale real player behavioral data, we adapt LLMs from general priors to game-specific domains, enabling realistic and interpretable player decision-making. In parallel, a data-driven environment model trained on real gameplay logs reconstructs dynamic in-game systems. Experiments demonstrate strong consistency with real-world player behaviors and plausible causal responses under interventions, providing a reliable, interpretable, and cost-efficient framework for data-driven numerical design optimization.
Large-scale digital platforms generate billions of timestamped user-item interactions (events) that are crucial for predicting user attributes in, e.g., fraud prevention and recommendations. While self-supervised learning (SSL) effectively models the temporal order of events, it typically overlooks the global structure of the user-item interaction graph. To bridge this gap, we propose three model-agnostic strategies for integrating this structural information into contrastive SSL: enriching event embeddings, aligning client representations with graph embeddings, and adding a structural pretext task. Experiments on four financial and e-commerce datasets demonstrate that our approach consistently improves the accuracy (up to a 2.3% AUC) and reveals that graph density is a key factor in selecting the optimal integration strategy.
Audio deepfakes generated by modern TTS and voice conversion systems are increasingly difficult to distinguish from real speech, raising serious risks for security and online trust. While state-of-the- art self-supervised models provide rich multi-layer representations, existing detectors treat layers independently and overlook temporal and hierarchical dependencies critical for identifying synthetic arte- facts. We propose HierCon, a hierarchical layer attention framework combined with margin-based contrastive learning that models de- pendencies across temporal frames, neighbouring layers, and layer groups, while encouraging domain-invariant embeddings. Evalu- ated on ASVspoof 2021 DF and In-the-Wild datasets, our method achieves state-of-the-art performance (1.93% and 6.87% EER), im- proving over independent layer weighting by 36.6% and 22.5% re- spectively. The results and attention visualisations confirm that hierarchical modelling enhances generalisation to cross-domain generation techniques and recording conditions.
Large Language Models (LLMs) have shown significant potential for improving recommendation systems through their inherent reasoning capabilities and extensive knowledge base. Yet, existing studies predominantly address warm-start scenarios with abundant user-item interaction data, leaving the more challenging cold-start scenarios, where sparse interactions hinder traditional collaborative filtering methods, underexplored. To address this limitation, we propose novel reasoning strategies designed for cold-start item recommendations within the Netflix domain. Our method utilizes the advanced reasoning capabilities of LLMs to effectively infer user preferences, particularly for newly introduced or rarely interacted items. We systematically evaluate supervised fine-tuning, reinforcement learning-based fine-tuning, and hybrid approaches that combine both methods to optimize recommendation performance. Extensive experiments on real-world data demonstrate significant improvements in both methodological efficacy and practical performance in cold-start recommendation contexts. Remarkably, our reasoning-based fine-tuned models outperform Netflix's production ranking model by up to 8% in certain cases.
Kuaishou serving hundreds of millions of searches daily, the quality of short-video search is paramount. However, it suffers from a severe Matthew effect on long-tail queries: sparse user behavior data causes models to amplify low-quality content such as clickbait and shallow content. The recent advancements in Large Language Models (LLMs) offer a new paradigm, as their inherent world knowledge provides a powerful mechanism to assess content quality, agnostic to sparse user interactions. To this end, we propose a LLM-driven multimodal reranking framework, which estimates user experience without real user behavior. The approach involves a two-stage training process: the first stage uses multimodal evidence to construct high-quality annotations for supervised fine-tuning, while the second stage incorporates pairwise preference optimization to help the model learn partial orderings among candidates. At inference time, the resulting experience scores are used to promote high-quality but underexposed videos in reranking, and further guide page-level optimization through reinforcement learning. Experiments show that the proposed method achieves consistent improvements over strong baselines in offline metrics including AUC, NDCG@K, and human preference judgement. An online A/B test covering 15% of traffic further demonstrates gains in both user experience and consumption metrics, confirming the practical value of the approach in long-tail video search scenarios.
Query-product relevance prediction is fundamental to e-commerce search and has become even more critical in the era of AI-powered shopping, where semantic understanding and complex reasoning directly shape the user experience and exert an indirect, yet substantial impact on business conversion. Large Language Models (LLMs) enable generative, reasoning-based approaches, typically aligned via supervised fine-tuning (SFT) or preference optimization methods like Direct Preference Optimization (DPO). However, the increasing complexity of business rules and user queries exposes the inability of existing methods to endow models with robust reasoning capacity for long-tail and challenging cases. Efforts to address this via reinforcement learning strategies like Group Relative Policy Optimization (GRPO) often suffer from sparse terminal rewards, offering insufficient guidance for multi-step reasoning, which in turn slows convergence. To address these challenges, we propose TaoSR-AGRL, an Adaptive Guided Reinforcement Learning framework for LLM-based relevance prediction in Taobao Search Relevance. TaoSR-AGRL introduces two key innovations: (1) Rule-aware Reward Shaping, which decomposes the final relevance judgment into dense, structured rewards aligned with domain-specific relevance criteria; and (2) Adaptive Guided Replay, which identifies low-accuracy rollouts during training and injects targeted ground-truth guidance to steer the policy away from stagnant, rule-violating reasoning patterns toward compliant trajectories. TaoSR-AGRL was evaluated on large-scale datasets and through online evaluations on Taobao Search. It consistently outperforms DPO and GRPO baselines, improving relevance accuracy and rule adherence, with measurable gains in user engagement and stable training dynamics. The model has been deployed on Taobao, serving hundreds of millions of users.
In e-commerce search, the diverse ways in which users express intentions lead to lexical and semantic gaps between queries and product descriptions, making query rewriting (QR) indispensable for improving matching efficiency. With the development of LLMs, QR has evolved from discriminative approaches to various LLM-based alignment methods. However, these methods typically treat all queries uniformly, without fundamentally distinguishing their rewriting difficulty or underlying linguistic issues, making the rewritten query deviate from human expectations. To address this limitation, we propose AWHCP (Aligning with Human Cognition and Preference), a novel framework that adopts a human-centric perspective and introduces the Problem–Intention–Fix–Rewrite (PIFR) paradigm. Built upon PIFR, AWHCP establishes a multi-granularity alignment training framework that simultaneously aligns with both system retrieval preferences and human rewriting behaviors. First, we construct high-quality PIFR-structured data and perform supervised fine-tuning to enable the model to learn human-like rewriting patterns. Second, we apply beam search to generate multiple candidates and leverage system-side feedback signals to conduct coarse-grained direct preference alignment, endowing the model with initial difficulty-aware reasoning capabilities. Third, we introduce a multi-dimensional rewrite quality judgment model trained via Group Relative Policy Optimization (GRPO), enabling fine-grained alignment with nuanced human rewriting preferences. Deployed on 1688's main search engine since August 2025, AWHCP has demonstrated strong effectiveness through extensive offline evaluations and large-scale online A/B tests, leading to a +3.9% gain in UV-L2O.
Traditional search engines return uniform results for identical queries, overlooking users' personalized intents. While personalized search has been extensively studied, research on query rewriting for personalized intents has been constrained by traditional approaches like statistical co-occurrence and synonym expansion. Current work primarily addresses multi-turn dialogue scenarios in Conversational AI rather than exploring applications in large-scale search engines. This typically stems from the absence of real search scenario data and the difficulty of inferring users' intents through personalized reasoning. While Large Language Models (LLMs) combined with Chain-of-Thought (CoT) capabilities provide possibilities for personalized reasoning, CoT introduces additional reasoning overhead that is difficult to accept in online scenarios requiring low latency. To address this, this paper proposes PicQue (Personalized Efficient Query Rewrite), a personalized query rewriting model training pipeline aimed at achieving high accuracy with low latency. PicQue contains a two-stage training that first employs a novel Hybrid Supervised Fine-Tuning strategy to retain the model's reasoning capabilities while allowing the decoding process to skip CoT, thereby obtaining CoT's accuracy gains without increasing latency. Building on this foundation, PicQue conducts second-stage reinforcement learning using Group Relative Policy Optimization (GRPO) to further improve rewriting accuracy and reduce the risk of over-rewriting. We also propose a Guided Search strategy to optimize GRPO training, alleviating the reduction in training sample utilization when all sampling rollouts are wrong. Extensive offline and online experiments demonstrate PicQue's effectiveness. In offline metrics, PicQue achieves over 7% improvement in rewriting accuracy compared to baseline methods and compresses up to 95% of decoding tokens. In online A/B tests, user satisfaction increases by 1.78% and query change rate decreases by 0.71%, achieving significant online gains.
Time series forecasting underpins many real-world services. Recent trends have focused on foundation models inspired by the paradigm of large language models, which rely on large volumes of centralized time-series data across diverse domains. However, such approaches raise significant concerns regarding data privacy. Federated learning (FL) has emerged as a promising paradigm for training unified time-series models using isolated datasets distributed across multiple clients. Nevertheless, existing FL methods face two critical challenges: heterogeneous variables and heterogeneous temporal correlations. To address these issues, we propose FedRMamba, a personalized federated forecasting framework built entirely from Mamba state-space blocks. Each client adopts a residual-coupled architecture, where a global frequency-aware Mamba module captures the common low-frequency structures shared across different variables, while a local patch-wise Mamba module learns personalized high-frequency patterns within the multivariate context. To clearly separate these responsibilities, we introduce a frequency-aware supervision that aligns the global path with low-frequency components and the local path with high-frequency residuals. Additionally, we design a gated fusion mechanism that dynamically combines the low-frequency and high-frequency components for improved prediction. We conduct extensive experiments to evaluate the performance of our proposed framework, demonstrating its effectiveness in handling heterogeneous data in federated settings.
As a key to accessing research impact, citation dynamics underpins research evaluation, scholarly recommendation, and the study of knowledge diffusion. Citation prediction is particularly critical for newborn papers, where early assessment must be performed without citation signals and under highly long-tailed distributions. We identify two key research gaps: (i) insufficient modeling of implicit factors of scientific impact, leading to reliance on coarse proxies; and (ii) a lack of bias-aware learning that can deliver stable predictions on lowly cited papers. We address these gaps by proposing a Bias-Aware Citation Prediction Framework, which combines multi-agent feature extraction with robust graph representation learning. First, a multi-agent × graph co-learning module derives fine-grained, interpretable signals, such as reproducibility, collaboration network, and text quality, from metadata and external resources, and fuses them with heterogeneous-network embeddings to provide rich supervision even in the absence of early citation signals. Second, we incorporate a set of robust mechanisms: a two-stage forward process that routes explicit factors through an intermediate exposure estimate, GroupDRO to optimize worst-case group risk across environments, and a regularization head that performs what-if analyses on controllable factors under monotonicity and smoothness constraints. Comprehensive experiments on two real-world datasets demonstrate the effectiveness of our proposed model. Specifically, our model achieves around a 13% reduction in error metrics (MALE and RMSLE) and a notable 5.5% improvement in the ranking metric (NDCG) over the baseline methods.