Bowen Song

Shenzhi Yang, Guangcheng Zhu, Bowen Song et al.

On-policy distillation (OPD) supervises the student only in output space by matching next-token probabilities. This output-only paradigm has two limits: (1) sampling variance from Monte Carlo KL estimates over large vocabularies (e.g., Qwen's ~150k tokens) persists throughout training, and (2) it treats the teacher as a black-box, discarding all intermediate hidden states after the LM head. We propose On-Policy Representation Distillation (OPRD), which lifts distillation into hidden-state space by aligning student and teacher representations across selected layers on the same rollouts, bypassing the LM head entirely. Theoretically, OPRD eliminates sampling variance and provides richer per-layer structural information. Empirically, OPRD closes the student-teacher gap on AIME 2024/2025 and AIMO, while output-space OPD baselines plateau below the teacher. OPRD also trains 1.44x faster and uses 54% less memory than top-k OPD. Code: https://github.com/ShenzhiYang2000/OPRD.

Guangcheng Zhu, Shenzhi Yang, Haobo Wang et al.

Reinforcement learning with verifiable rewards (RLVR) has greatly advanced large reasoning models (LRMs), but it requires timely training on a huge fully-annotated dataset. To this end, data-efficient RLVR methods have been widely studied from two perspectives: (i) data selection methods identify a small subset of "golden" samples that yield near-full-data performance, but they rely on a pre-existing pool of labeled data. (ii) unsupervised RLVR methods train the model using its own internal supervision signals on large-scale unlabeled data, yet they exhibit suboptimal performance. Accordingly, we investigate the "pick in the dark" setup for RLVR, which aims to select, without prior supervision, unlabeled samples that are most beneficial for training and worthy of annotation. Through systematic analysis, we demonstrate that smart picks hinge on a well-calibrated uncertainty estimator to enable strategic partitioning of data for adaptive training regimes. Building on this insight, we propose PivotTrace, a three-way data triage framework that leverages attention dynamics to trace metacognitive pivots during reasoning. By precisely quantifying uncertainty through pivot density, PivotTrace achieves automated data routing to synergistically maximize both annotation and training efficiency. Empirically, PivotTrace surpasses the fully supervised LRM with only 29.3% annotated samples and 2.75 faster convergence.

Guangcheng Zhu, Shenzhi Yang, Haobo Wang et al.

Reinforcement learning with verifiable rewards (RLVR) significantly advances LLM reasoning, yet it faces a dilemma: standard supervised scaling is throttled by high annotation costs, while unsupervised alternatives suffer from severe model collapse. Recent semi-supervised RLVR methods address this by using a small labeled set to guide unlabeled data, achieving a promising trade-off between training efficacy and annotation cost. However, they suffer from a severe data-efficiency bottleneck due to the reliance on coarse performance heuristics, leaving a vast majority of valuable instances underutilized. To this end, we propose GeoMin, which models global feature distributions on labeled data to decode the structural discrepancy between correct and incorrect rollouts, thereby establishing a robust prior to assess the reliability of self-reward signals and fully unleash the potential of unlabeled data. Empirically, GeoMin outperforms the strongest baselines by +4.1% and even surpasses fully supervised models with only 10% of the annotations, demonstrating remarkable data efficiency.

Xin Sun, Zhongqi Chen, Xing Zheng et al.

Knowledge Base Question Answering (KBQA) challenges models to bridge the gap between natural language and strict knowledge graph schemas by generating executable logical forms. While Large Language Models (LLMs) have advanced this field, current approaches often struggle with a dichotomy of failure: they either generate hallucinated queries without verifying schema existence or exhibit rigid, template-based reasoning that mimics synthesized traces without true comprehension of the environment. To address these limitations, we present \textbf{KBQA-R1}, a framework that shifts the paradigm from text imitation to interaction optimization via Reinforcement Learning. Treating KBQA as a multi-turn decision process, our model learns to navigate the knowledge base using a list of actions, leveraging Group Relative Policy Optimization (GRPO) to refine its strategies based on concrete execution feedback rather than static supervision. Furthermore, we introduce \textbf{Referenced Rejection Sampling (RRS)}, a data synthesis method that resolves cold-start challenges by strictly aligning reasoning traces with ground-truth action sequences. Extensive experiments on WebQSP, GrailQA, and GraphQuestions demonstrate that KBQA-R1 achieves state-of-the-art performance, effectively grounding LLM reasoning in verifiable execution.

Xin Sun, Zhongqi Chen, Qiang Liu et al.

Retrieval-Augmented Generation (RAG) has emerged as a powerful approach for enhancing large language models' question-answering capabilities through the integration of external knowledge. However, when adapting RAG systems to specialized domains, challenges arise from distribution shifts, resulting in suboptimal generalization performance. In this work, we propose TTARAG, a test-time adaptation method that dynamically updates the language model's parameters during inference to improve RAG system performance in specialized domains. Our method introduces a simple yet effective approach where the model learns to predict retrieved content, enabling automatic parameter adjustment to the target domain. Through extensive experiments across six specialized domains, we demonstrate that TTARAG achieves substantial performance improvements over baseline RAG systems. Code available at https://github.com/sunxin000/TTARAG.

Bowen Song, Soo Min Kwon, Zecheng Zhang et al.

Diffusion models have recently emerged as powerful generative priors for solving inverse problems. However, training diffusion models in the pixel space are both data-intensive and computationally demanding, which restricts their applicability as priors for high-dimensional real-world data such as medical images. Latent diffusion models, which operate in a much lower-dimensional space, offer a solution to these challenges. However, incorporating latent diffusion models to solve inverse problems remains a challenging problem due to the nonlinearity of the encoder and decoder. To address these issues, we propose \textit{ReSample}, an algorithm that can solve general inverse problems with pre-trained latent diffusion models. Our algorithm incorporates data consistency by solving an optimization problem during the reverse sampling process, a concept that we term as hard data consistency. Upon solving this optimization problem, we propose a novel resampling scheme to map the measurement-consistent sample back onto the noisy data manifold and theoretically demonstrate its benefits. Lastly, we apply our algorithm to solve a wide range of linear and nonlinear inverse problems in both natural and medical images, demonstrating that our approach outperforms existing state-of-the-art approaches, including those based on pixel-space diffusion models.

Jiankun Zhao, Bowen Song, Liyue Shen

Diffusion models have been demonstrated as strong priors for solving general inverse problems. Most existing Diffusion model-based Inverse Problem Solvers (DIS) employ a plug-and-play approach to guide the sampling trajectory with either projections or gradients. Though effective, these methods generally necessitate hundreds of sampling steps, posing a dilemma between inference time and reconstruction quality. In this work, we try to push the boundary of inference steps to 1-2 NFEs while still maintaining high reconstruction quality. To achieve this, we propose to leverage a pretrained distillation of diffusion model, namely consistency model, as the data prior. The key to achieving few-step guidance is to enforce two types of constraints during the sampling process of the consistency model: soft measurement constraint with ControlNet and hard measurement constraint via optimization. Supporting both single-step reconstruction and multistep refinement, the proposed framework further provides a way to trade image quality with additional computational cost. Within comparable NFEs, our method achieves new state-of-the-art in diffusion-based inverse problem solving, showcasing the significant potential of employing prior-based inverse problem solvers for real-world applications. Code is available at: https://github.com/BioMed-AI-Lab-U-Michgan/cosign.

Shenzhi Yang, Guangcheng Zhu, Xing Zheng et al.

Reinforcement learning with verifiable rewards (RLVR) has proven effective in training large reasoning models (LRMs) by leveraging answer-verifiable signals to guide policy optimization, which, however, suffers from high annotation costs. To alleviate this problem, recent work has explored unsupervised RLVR methods that derive rewards solely from the model's internal consistency, such as through entropy and majority voting. While seemingly promising, these methods often suffer from model collapse in the later stages of training, which may arise from the reinforcement of incorrect reasoning patterns in the absence of external supervision. In this work, we investigate a novel semi-supervised RLVR paradigm that utilizes a small labeled set to guide RLVR training on unlabeled samples. Our key insight is that supervised rewards are essential for stabilizing consistency-based training on unlabeled samples, ensuring that only reasoning patterns verified on labeled instances are incorporated into RL training. Technically, we propose an effective policy optimization algorithm, TraPO, that identifies reliable unlabeled samples by matching their learning trajectory similarity to labeled ones. Building on this, TraPO achieves remarkable data efficiency and strong generalization on six widely used mathematical reasoning benchmarks (AIME24/25, AMC, MATH-500, Minerva, and Olympiad) and three out-of-distribution tasks (ARC-c, GPQA-diamond, and MMLU-pro). With only 1K labeled and 3K unlabeled samples, TraPO reaches 42.6% average accuracy, surpassing the best unsupervised method trained on 45K unlabeled samples (38.3%). Notably, when using 4K labeled and 12K unlabeled samples, TraPO even outperforms the fully supervised model trained on the full 45K labeled samples on all benchmarks, while using only 10% of the labeled data. The code is available via https://github.com/ShenzhiYang2000/TRAPO.

Xin Sun, Jianan Xie, Zhongqi Chen et al.

Reinforcement learning (RL) is a natural fit for agentic knowledge base question answering (KBQA), where a model must issue executable actions, observe knowledge-base feedback, and eventually return an answer. However, current RL-based KBQA systems mainly optimize sparse rewards from the final answer, leaving intermediate action errors weakly supervised. This is especially limiting for logical-form annotated KBQA benchmarks: gold logical forms can be converted into executable action sequences, but existing pipelines use them mainly for warm-start data construction rather than for on-policy RL updates. We propose GAPD, a training-time Gold-Action Policy Distillation framework that adds dense token-level guidance to outcome-based RL. To align gold actions with on-policy student rollouts, GAPD uses MID-ANCHOR MATCHING: it treats the intermediate entities reached during student exploration and gold execution as state anchors, and matches student states to gold states through these explored entity sets. The current policy conditioned on this aligned gold action serves as a stop-gradient teacher, whose token distribution is distilled back to the ordinary student policy over generated action-token spans. GAPD consistently surpasses the current state of the art on WebQSP, GrailQA, and GraphQ.

Bowen Song, Marco Paolieri, Harper E. Stewart et al.

Objective: Our aim is to determine if data collected with inertial measurement units (IMUs) during steady-state running could be used to estimate ground reaction forces (GRFs) and to derive biomechanical variables (e.g., contact time, impulse, change in velocity) using lightweight machine-learning approaches. In contrast, state-of-the-art estimation using LSTMs suffers from prohibitive inference times on edge devices, requires expensive training and hyperparameter optimization, and results in black box models. Methods: We proposed a novel lightweight solution, SVD Embedding Regression (SER), using linear regression between SVD embeddings of IMU data and GRF data. We also compared lightweight solutions including SER and k-Nearest-Neighbors (KNN) regression with state-of-the-art LSTMs. Results: We performed extensive experiments to evaluate these techniques under multiple scenarios and combinations of IMU signals and quantified estimation errors for predicting GRFs and biomechanical variables. We did this using training data from different athletes, from the same athlete, or both, and we explored the use of acceleration and angular velocity data from sensors at different locations (sacrum and shanks). Conclusion: Our results illustrated that lightweight solutions such as SER and KNN can be similarly accurate or more accurate than LSTMs. The use of personal data reduced estimation errors of all methods, particularly for most biomechanical variables (as compared to GRFs); moreover, this gain was more pronounced in the lightweight methods. Significance: The study of GRFs is used to characterize the mechanical loading experienced by individuals in movements such as running, which is clinically applicable to identify athletes at risk for stress-related injuries.

Xinyu Hu, Pengfei Tang, Simiao Zuo et al.

Large language models (LLMs) have made impressive progress in natural language processing. These models rely on proper human instructions (or prompts) to generate suitable responses. However, the potential of LLMs are not fully harnessed by commonly-used prompting methods: many human-in-the-loop algorithms employ ad-hoc procedures for prompt selection; while auto prompt generation approaches are essentially searching all possible prompts randomly and inefficiently. We propose Evoke, an automatic prompt refinement framework. In Evoke, there are two instances of a same LLM: one as a reviewer (LLM-Reviewer), it scores the current prompt; the other as an author (LLM-Author), it edits the prompt by considering the edit history and the reviewer's feedback. Such an author-reviewer feedback loop ensures that the prompt is refined in each iteration. We further aggregate a data selection approach to Evoke, where only the hard samples are exposed to the LLM. The hard samples are more important because the LLM can develop deeper understanding of the tasks out of them, while the model may already know how to solve the easier cases. Experimental results show that Evoke significantly outperforms existing methods. For instance, in the challenging task of logical fallacy detection, Evoke scores above 80, while all other baseline methods struggle to reach 20.

Bowen Song, Sebastien Gros, Andrea Iannelli

Policy gradient algorithms are widely used in reinforcement learning and belong to the class of approximate dynamic programming methods. This paper studies two key policy gradient algorithms, the Natural Policy Gradient and the Gauss-Newton Method, for solving the Linear Quadratic Regulator (LQR) problem in unknown stochastic linear systems. The main challenge lies in obtaining an unbiased gradient estimate from noisy data due to errors-in-variables in linear regression. This issue is addressed by employing a primal-dual estimation procedure. Using this novel gradient estimation scheme, the paper establishes convergence guarantees with a sample complexity of order O(1/epsilon). Theoretical results are further supported by numerical experiments, which demonstrate the effectiveness of the proposed algorithms.

Bowen Song, Jianshen Zhu, Naveed Ahmed Azam et al.

In this paper, we propose a novel family of descriptors of chemical graphs, named cycle-configuration (CC), that can be used in the standard "two-layered (2L) model" of mol-infer, a molecular inference framework based on mixed integer linear programming (MILP) and machine learning (ML). Proposed descriptors capture the notion of ortho/meta/para patterns that appear in aromatic rings, which has been impossible in the framework so far. Computational experiments show that, when the new descriptors are supplied, we can construct prediction functions of similar or better performance for all of the 27 tested chemical properties. We also provide an MILP formulation that asks for a chemical graph with desired properties under the 2L model with CC descriptors (2L+CC model). We show that a chemical graph with up to 50 non-hydrogen vertices can be inferred in a practical time.

Zitao Shuai, Chenwei Wu, Zhengxu Tang et al.

Diffusion Transformers (DiTs) have achieved remarkable success in diverse and high-quality text-to-image(T2I) generation. However, how text and image latents individually and jointly contribute to the semantics of generated images, remain largely unexplored. Through our investigation of DiT's latent space, we have uncovered key findings that unlock the potential for zero-shot fine-grained semantic editing: (1) Both the text and image spaces in DiTs are inherently decomposable. (2) These spaces collectively form a disentangled semantic representation space, enabling precise and fine-grained semantic control. (3) Effective image editing requires the combined use of both text and image latent spaces. Leveraging these insights, we propose a simple and effective Extract-Manipulate-Sample (EMS) framework for zero-shot fine-grained image editing. Our approach first utilizes a multi-modal Large Language Model to convert input images and editing targets into text descriptions. We then linearly manipulate text embeddings based on the desired editing degree and employ constrained score distillation sampling to manipulate image embeddings. We quantify the disentanglement degree of the latent space of diffusion models by proposing a new metric. To evaluate fine-grained editing performance, we introduce a comprehensive benchmark incorporating both human annotations, manual evaluation, and automatic metrics. We have conducted extensive experimental results and in-depth analysis to thoroughly uncover the semantic disentanglement properties of the diffusion transformer, as well as the effectiveness of our proposed method. Our annotated benchmark dataset is publicly available at https://anonymous.com/anonymous/EMS-Benchmark, facilitating reproducible research in this domain.

Bowen Song, Andrea Iannelli

Policy gradient (PG) methods are the backbone of many reinforcement learning algorithms due to their good performance in policy optimization problems. As a gradient-based approach, PG methods typically rely on knowledge of the system dynamics. If this is not available, trajectory data can be utilized to approximate first-order information. When the data are noisy, gradient estimates become inaccurate and a study that investigates uncertainty estimation and the analysis of its propagation through the algorithm is currently missing. To address this, our work focuses on the Linear Quadratic Regulator (LQR) problem for systems subject to additive stochastic noise. After briefly summarizing the state of the art for cases with a known model, we focus on scenarios where the system dynamics are unknown, and approximate gradient information is obtained using zeroth-order optimization techniques. We analyze the theoretical properties by computing the error in the estimated gradient and examining how this error affects the convergence of PG algorithms. Additionally, we provide global convergence guarantees for various versions of PG methods, including those employing adaptive step sizes and variance reduction techniques, which help increase the convergence rate and reduce sample complexity. This study contributed to characterizing the robustness of model-free PG methods, aiming to identify their limitations in the presence of stochastic noise and proposing improvements to enhance their applicability.

SAM 3D Team, Xingyu Chen, Fu-Jen Chu et al.

We present SAM 3D, a generative model for visually grounded 3D object reconstruction, predicting geometry, texture, and layout from a single image. SAM 3D excels in natural images, where occlusion and scene clutter are common and visual recognition cues from context play a larger role. We achieve this with a human- and model-in-the-loop pipeline for annotating object shape, texture, and pose, providing visually grounded 3D reconstruction data at unprecedented scale. We learn from this data in a modern, multi-stage training framework that combines synthetic pretraining with real-world alignment, breaking the 3D "data barrier". We obtain significant gains over recent work, with at least a 5:1 win rate in human preference tests on real-world objects and scenes. We will release our code and model weights, an online demo, and a new challenging benchmark for in-the-wild 3D object reconstruction.

Qiang Liu, Xinlong Chen, Yue Ding et al.

Hallucination has emerged as a significant barrier to the effective application of Large Language Models (LLMs). In this work, we introduce a novel Attention-Guided SElf-Reflection (AGSER) approach for zero-shot hallucination detection in LLMs. The AGSER method utilizes attention contributions to categorize the input query into attentive and non-attentive queries. Each query is then processed separately through the LLMs, allowing us to compute consistency scores between the generated responses and the original answer. The difference between the two consistency scores serves as a hallucination estimator. In addition to its efficacy in detecting hallucinations, AGSER notably reduces computational overhead, requiring only three passes through the LLM and utilizing two sets of tokens. We have conducted extensive experiments with four widely-used LLMs across three different hallucination benchmarks, demonstrating that our approach significantly outperforms existing methods in zero-shot hallucination detection.

Jason Hu, Bowen Song, Jeffrey A. Fessler et al.

Diffusion models have achieved excellent success in solving inverse problems due to their ability to learn strong image priors, but existing approaches require a large training dataset of images that should come from the same distribution as the test dataset. When the training and test distributions are mismatched, artifacts and hallucinations can occur in reconstructed images due to the incorrect priors. In this work, we systematically study out of distribution (OOD) problems where a known training distribution is first provided. We first study the setting where only a single measurement obtained from the unknown test distribution is available. Next we study the setting where a very small sample of data belonging to the test distribution is available, and our goal is still to reconstruct an image from a measurement that came from the test distribution. In both settings, we use a patch-based diffusion prior that learns the image distribution solely from patches. Furthermore, in the first setting, we include a self-supervised loss that helps the network output maintain consistency with the measurement. Extensive experiments show that in both settings, the patch-based method can obtain high quality image reconstructions that can outperform whole-image models and can compete with methods that have access to large in-distribution training datasets. Furthermore, we show how whole-image models are prone to memorization and overfitting, leading to artifacts in the reconstructions, while a patch-based model can resolve these issues.

Yiming Zhang, Yingfan Ma, Yanmei Gu et al.

Large Language Models (LLMs) have shown impressive performance in domains such as mathematics and programming, yet their capabilities in physics remain underexplored and poorly understood. Physics poses unique challenges that demand not only precise computation but also deep conceptual understanding and physical modeling skills. Existing benchmarks often fall short due to limited difficulty, multiple-choice formats, and static evaluation settings that fail to capture physical modeling ability. In this paper, we introduce ABench-Physics, a novel benchmark designed to rigorously evaluate LLMs' physical reasoning and generalization capabilities. ABench-Physics consists of two components: Phy_A, a static set of 400 graduate- or Olympiad-level problems; and Phy_B, a dynamic subset of 100 problems equipped with an automatic variation engine to test model robustness across changing conditions. All questions require precise numerical answers, with strict formatting and tolerance constraints. Our evaluation of several state-of-the-art LLMs reveals substantial performance gaps, highlighting persistent limitations in physical reasoning, especially in generalization to dynamic variants. ABench-Physics provides a challenging and diagnostic framework for advancing scientific reasoning in LLMs.

Shenzhi Yang, Guangcheng Zhu, Bowen Song et al.

Reinforcement Learning with Verifiable Rewards (RLVR) effectively trains reasoning models that rely on abundant perfect labels, but its vulnerability to unavoidable noisy labels due to expert scarcity remains critically underexplored. In this work, we take the first step toward a systematic analysis of noisy label mechanisms in RLVR. In contrast to supervised classification, most RLVR algorithms incorporate a rollout-based condition: a label's influence on training is contingent on whether the current policy can generate rollouts that realize it, a property that naturally extends to noisy labels. Based on this observation, we distinguish two types of noise: inactive noisy labels, which reduce data efficiency, and active noisy labels, which are reinforced and risk skewing the model toward incorrect distributions. From experiments on training with noisy samples, we identify an Early Correctness Coherence phenomenon: although noisy samples begin to lag behind in later stages, accuracy on both clean and noisy samples increases similarly in early training. Motivated by this dynamic, we propose Online Label Refinement (OLR), which progressively corrects potentially noisy labels with majority-voted answers when two conditions hold: a positive slope in the majority answer's rollout pass rate and stable historical consistency across updates, enabling gradual self-correction as the policy improves. We evaluate OLR on six in-distribution mathematical reasoning benchmarks (AIME24/25, AMC, MATH-500, Minerva, and Olympiad) and three out-of-distribution tasks (ARC-c, GPQA-diamond, and MMLU-pro). Across noise ratios from 0.1 to 0.9, OLR consistently improves robustness under both inactive and active noisy-label settings, achieving average gains of 3.6% to 3.9% on in-distribution benchmarks and 3.3% to 4.6% on out-of-distribution evaluations.

Xin Sun, Jianan Xie, Zhongqi Chen et al.

Large language models (LLMs) augmented with retrieval systems have significantly advanced natural language processing tasks by integrating external knowledge sources, enabling more accurate and contextually rich responses. To improve the robustness of such systems against noisy retrievals, Retrieval-Augmented Fine-Tuning (RAFT) has emerged as a widely adopted method. However, RAFT conditions models to generate answers even in the absence of reliable knowledge. This behavior undermines their reliability in high-stakes domains, where acknowledging uncertainty is critical. To address this issue, we propose Divide-Then-Align (DTA), a post-training approach designed to endow RAG systems with the ability to respond with "I don't know" when the query is out of the knowledge boundary of both the retrieved passages and the model's internal knowledge. DTA divides data samples into four knowledge quadrants and constructs tailored preference data for each quadrant, resulting in a curated dataset for Direct Preference Optimization (DPO). Experimental results on three benchmark datasets demonstrate that DTA effectively balances accuracy with appropriate abstention, enhancing the reliability and trustworthiness of retrieval-augmented systems.

Zihan Wang, Songlin Li, Lingyan Hao et al.

As video generation models advance rapidly, assessing the quality of generated videos has become increasingly critical. Existing metrics, such as Fréchet Video Distance (FVD), Inception Score (IS), and ClipSim, measure quality primarily in latent space rather than from a human visual perspective, often overlooking key aspects like appearance and motion consistency to physical laws. In this paper, we propose a novel metric, VAMP (Visual Appearance and Motion Plausibility), that evaluates both the visual appearance and physical plausibility of generated videos. VAMP is composed of two main components: an appearance score, which assesses color, shape, and texture consistency across frames, and a motion score, which evaluates the realism of object movements. We validate VAMP through two experiments: corrupted video evaluation and generated video evaluation. In the corrupted video evaluation, we introduce various types of corruptions into real videos and measure the correlation between corruption severity and VAMP scores. In the generated video evaluation, we use state-of-the-art models to generate videos from carefully designed prompts and compare VAMP's performance to human evaluators' rankings. Our results demonstrate that VAMP effectively captures both visual fidelity and temporal consistency, offering a more comprehensive evaluation of video quality than traditional methods.

Bonan Zhang, Zhongqi Chen, Bowen Song et al.

Reinforcement learning (RL) has become a standard paradigm for refining large language models (LLMs) beyond pre-training and instruction tuning. A prominent line of work is RL with verifiable rewards (RLVR), which leverages automatically verifiable outcomes (e.g., correctness or executability) to generate reward signals. While efficient, this framework faces two key limitations: First, its binary feedback is too sparse to capture the quality of the reasoning process. Second, its coarse-grained rewards potentially lead to vanishing gradients. Inspired by observations from human learning, we introduce a RL technique that integrates verifiable outcomes with the model's own confidence estimates. This joint design enriches the reward signal, providing finer-grained feedback and implicitly supervising the reasoning process. Experimental results demonstrate that our proposed method enhances RL performance across multiple datasets and reduces token consumption during inference, while incurring negligible additional training cost. Moreover, it can be used as a plug-in module to enhance other state-of-the-art RL methods.

Bowen Song, Zecheng Zhang, Zhaoxu Luo et al.

Diffusion models have emerged as powerful tools for generative tasks, producing high-quality outputs across diverse domains. However, how the generated data responds to the initial noise perturbation in diffusion models remains under-explored, which hinders understanding the controllability of the sampling process. In this work, we first observe an interesting phenomenon: the relationship between the change of generation outputs and the scale of initial noise perturbation is highly linear through the diffusion ODE sampling. Then we provide both theoretical and empirical study to justify this linearity property of this input-output (noise-generation data) relationship. Inspired by these new insights, we propose a novel Controllable and Constrained Sampling method (CCS) together with a new controller algorithm for diffusion models to sample with desired statistical properties while preserving good sample quality. We perform extensive experiments to compare our proposed sampling approach with other methods on both sampling controllability and sampled data quality. Results show that our CCS method achieves more precisely controlled sampling while maintaining superior sample quality and diversity.

Zitao Shuai, Chenwei Wu, Zhengxu Tang et al.

Diffusion Transformers (DiTs) have recently achieved remarkable success in text-guided image generation. In image editing, DiTs project text and image inputs to a joint latent space, from which they decode and synthesize new images. However, it remains largely unexplored how multimodal information collectively forms this joint space and how they guide the semantics of the synthesized images. In this paper, we investigate the latent space of DiT models and uncover two key properties: First, DiT's latent space is inherently semantically disentangled, where different semantic attributes can be controlled by specific editing directions. Second, consistent semantic editing requires utilizing the entire joint latent space, as neither encoded image nor text alone contains enough semantic information. We show that these editing directions can be obtained directly from text prompts, enabling precise semantic control without additional training or mask annotations. Based on these insights, we propose a simple yet effective Encode-Identify-Manipulate (EIM) framework for zero-shot fine-grained image editing. Specifically, we first encode both the given source image and the text prompt that describes the image, to obtain the joint latent embedding. Then, using our proposed Hessian Score Distillation Sampling (HSDS) method, we identify editing directions that control specific target attributes while preserving other image features. These directions are guided by text prompts and used to manipulate the latent embeddings. Moreover, we propose a new metric to quantify the disentanglement degree of the latent space of diffusion models. Extensive experiment results on our new curated benchmark dataset and analysis demonstrate DiT's disentanglement properties and effectiveness of the EIM framework.

Dan Lin, Shunfeng Lu, Ziyan Liu et al.

Cross-chain bridges play a vital role in enabling blockchain interoperability. However, due to the inherent design flaws and the enormous value they hold, they have become prime targets for hacker attacks. Existing detection methods show progress yet remain limited, as they mainly address single-chain behaviors and fail to capture cross-chain semantics. To address this gap, we leverage heterogeneous graph attention networks, which are well-suited for modeling multi-typed entities and relations, to capture the complex execution semantics of cross-chain behaviors. We propose BridgeShield, a detection framework that jointly models the source chain, off-chain coordination, and destination chain within a unified heterogeneous graph representation. BridgeShield incorporates intra-meta-path attention to learn fine-grained dependencies within cross-chain paths and inter-meta-path attention to highlight discriminative cross-chain patterns, thereby enabling precise identification of attack behaviors. Extensive experiments on 51 real-world cross-chain attack events demonstrate that BridgeShield achieves an average F1-score of 92.58%, representing a 24.39% improvement over state-of-the-art baselines. These results validate the effectiveness of BridgeShield as a practical solution for securing cross-chain bridges and enhancing the resilience of multi-chain ecosystems.

Kossi Amouzouvi, Bowen Song, Andrea Coletta et al.

Knowledge graph representation learning approaches provide a mapping between symbolic knowledge in the form of triples in a knowledge graph (KG) and their feature vectors. Knowledge graph embedding (KGE) models often represent relations in a KG as geometric transformations. Most state-of-the-art (SOTA) KGE models are derived from elementary geometric transformations (EGTs), such as translation, scaling, rotation, and reflection, or their combinations. These geometric transformations enable the models to effectively preserve specific structural and relational patterns of the KG. However, the current use of EGTs by KGEs remains insufficient without considering relation-specific transformations. Although recent models attempted to address this problem by ensembling SOTA baseline models in different ways, only a single or composite version of geometric transformations are used by such baselines to represent all the relations. In this paper, we propose a framework that evaluates how well each relation fits with different geometric transformations. Based on this ranking, the model can: (1) assign the best-matching transformation to each relation, or (2) use majority voting to choose one transformation type to apply across all relations. That is, the model learns a single relation-specific EGT in low dimensional vector space through an attention mechanism. Furthermore, we use the correlation between relations and EGTs, which are learned in a low dimension, for relation embeddings in a high dimensional vector space. The effectiveness of our models is demonstrated through comprehensive evaluations on three benchmark KGs as well as a real-world financial KG, witnessing a performance comparable to leading models

Zhaoxu Luo, Bowen Song, Liyue Shen

During the acquisition of satellite images, there is generally a trade-off between spatial resolution and temporal resolution (acquisition frequency) due to the onboard sensors of satellite imaging systems. High-resolution satellite images are very important for land crop monitoring, urban planning, wildfire management and a variety of applications. It is a significant yet challenging task to achieve high spatial-temporal resolution in satellite imaging. With the advent of diffusion models, we can now learn strong generative priors to generate realistic satellite images with high resolution, which can be utilized to promote the super-resolution task as well. In this work, we propose a novel diffusion-based fusion algorithm called \textbf{SatDiffMoE} that can take an arbitrary number of sequential low-resolution satellite images at the same location as inputs, and fuse them into one high-resolution reconstructed image with more fine details, by leveraging and fusing the complementary information from different time points. Our algorithm is highly flexible and allows training and inference on arbitrary number of low-resolution images. Experimental results show that our proposed SatDiffMoE method not only achieves superior performance for the satellite image super-resolution tasks on a variety of datasets, but also gets an improved computational efficiency with reduced model parameters, compared with previous methods.

Bowen Song, Jason Hu, Zhaoxu Luo et al.

Diffusion models face significant challenges when employed for large-scale medical image reconstruction in real practice such as 3D Computed Tomography (CT). Due to the demanding memory, time, and data requirements, it is difficult to train a diffusion model directly on the entire volume of high-dimensional data to obtain an efficient 3D diffusion prior. Existing works utilizing diffusion priors on single 2D image slice with hand-crafted cross-slice regularization would sacrifice the z-axis consistency, which results in severe artifacts along the z-axis. In this work, we propose a novel framework that enables learning the 3D image prior through position-aware 3D-patch diffusion score blending for reconstructing large-scale 3D medical images. To the best of our knowledge, we are the first to utilize a 3D-patch diffusion prior for 3D medical image reconstruction. Extensive experiments on sparse view and limited angle CT reconstruction show that our DiffusionBlend method significantly outperforms previous methods and achieves state-of-the-art performance on real-world CT reconstruction problems with high-dimensional 3D image (i.e., $256 \times 256 \times 500$). Our algorithm also comes with better or comparable computational efficiency than previous state-of-the-art methods.

Jason Hu, Bowen Song, Xiaojian Xu et al.

Diffusion models can learn strong image priors from underlying data distribution and use them to solve inverse problems, but the training process is computationally expensive and requires lots of data. Such bottlenecks prevent most existing works from being feasible for high-dimensional and high-resolution data such as 3D images. This paper proposes a method to learn an efficient data prior for the entire image by training diffusion models only on patches of images. Specifically, we propose a patch-based position-aware diffusion inverse solver, called PaDIS, where we obtain the score function of the whole image through scores of patches and their positional encoding and utilize this as the prior for solving inverse problems. First of all, we show that this diffusion model achieves an improved memory efficiency and data efficiency while still maintaining the capability to generate entire images via positional encoding. Additionally, the proposed PaDIS model is highly flexible and can be plugged in with different diffusion inverse solvers (DIS). We demonstrate that the proposed PaDIS approach enables solving various inverse problems in both natural and medical image domains, including CT reconstruction, deblurring, and superresolution, given only patch-based priors. Notably, PaDIS outperforms previous DIS methods trained on entire image priors in the case of limited training data, demonstrating the data efficiency of our proposed approach by learning patch-based prior.

Sheng Tian, Jihai Dong, Jintang Li et al.

Anomaly detection aims to distinguish abnormal instances that deviate significantly from the majority of benign ones. As instances that appear in the real world are naturally connected and can be represented with graphs, graph neural networks become increasingly popular in tackling the anomaly detection problem. Despite the promising results, research on anomaly detection has almost exclusively focused on static graphs while the mining of anomalous patterns from dynamic graphs is rarely studied but has significant application value. In addition, anomaly detection is typically tackled from semi-supervised perspectives due to the lack of sufficient labeled data. However, most proposed methods are limited to merely exploiting labeled data, leaving a large number of unlabeled samples unexplored. In this work, we present semi-supervised anomaly detection (SAD), an end-to-end framework for anomaly detection on dynamic graphs. By a combination of a time-equipped memory bank and a pseudo-label contrastive learning module, SAD is able to fully exploit the potential of large unlabeled samples and uncover underlying anomalies on evolving graph streams. Extensive experiments on four real-world datasets demonstrate that SAD efficiently discovers anomalies from dynamic graphs and outperforms existing advanced methods even when provided with only little labeled data.

Yongchun Zhu, Dongbo Xi, Bowen Song et al.

With the explosive growth of the e-commerce industry, detecting online transaction fraud in real-world applications has become increasingly important to the development of e-commerce platforms. The sequential behavior history of users provides useful information in differentiating fraudulent payments from regular ones. Recently, some approaches have been proposed to solve this sequence-based fraud detection problem. However, these methods usually suffer from two problems: the prediction results are difficult to explain and the exploitation of the internal information of behaviors is insufficient. To tackle the above two problems, we propose a Hierarchical Explainable Network (HEN) to model users' behavior sequences, which could not only improve the performance of fraud detection but also make the inference process interpretable. Meanwhile, as e-commerce business expands to new domains, e.g., new countries or new markets, one major problem for modeling user behavior in fraud detection systems is the limitation of data collection, e.g., very few data/labels available. Thus, in this paper, we further propose a transfer framework to tackle the cross-domain fraud detection problem, which aims to transfer knowledge from existing domains (source domains) with enough and mature data to improve the performance in the new domain (target domain). Our proposed method is a general transfer framework that could not only be applied upon HEN but also various existing models in the Embedding & MLP paradigm. Based on 90 transfer task experiments, we also demonstrate that our transfer framework could not only contribute to the cross-domain fraud detection task with HEN, but also be universal and expandable for various existing models.

Dongbo Xi, Fuzhen Zhuang, Bowen Song et al.

Many prediction tasks of real-world applications need to model multi-order feature interactions in user's event sequence for better detection performance. However, existing popular solutions usually suffer two key issues: 1) only focusing on feature interactions and failing to capture the sequence influence; 2) only focusing on sequence information, but ignoring internal feature relations of each event, thus failing to extract a better event representation. In this paper, we consider a two-level structure for capturing the hierarchical information over user's event sequence: 1) learning effective feature interactions based event representation; 2) modeling the sequence representation of user's historical events. Experimental results on both industrial and public datasets clearly demonstrate that our model achieves significantly better performance compared with state-of-the-art baselines.

Dongbo Xi, Bowen Song, Fuzhen Zhuang et al.

With the explosive growth of e-commerce, online transaction fraud has become one of the biggest challenges for e-commerce platforms. The historical behaviors of users provide rich information for digging into the users' fraud risk. While considerable efforts have been made in this direction, a long-standing challenge is how to effectively exploit internal user information and provide explainable prediction results. In fact, the value variations of same field from different events and the interactions of different fields inside one event have proven to be strong indicators for fraudulent behaviors. In this paper, we propose the Dual Importance-aware Factorization Machines (DIFM), which exploits the internal field information among users' behavior sequence from dual perspectives, i.e., field value variations and field interactions simultaneously for fraud detection. The proposed model is deployed in the risk management system of one of the world's largest e-commerce platforms, which utilize it to provide real-time transaction fraud detection. Experimental results on real industrial data from different regions in the platform clearly demonstrate that our model achieves significant improvements compared with various state-of-the-art baseline models. Moreover, the DIFM could also give an insight into the explanation of the prediction results from dual perspectives.

Wenjing Fang, Chaochao Chen, Bowen Song et al.

Secure online transaction is an essential task for e-commerce platforms. Alipay, one of the world's leading cashless payment platform, provides the payment service to both merchants and individual customers. The fraud detection models are built to protect the customers, but stronger demands are raised by the new scenes, which are lacking in training data and labels. The proposed model makes a difference by utilizing the data under similar old scenes and the data under a new scene is treated as the target domain to be promoted. Inspired by this real case in Alipay, we view the problem as a transfer learning problem and design a set of revise strategies to transfer the source domain models to the target domain under the framework of gradient boosting tree models. This work provides an option for the cold-starting and data-sharing problems.