arXiv subjects

Wenyuan Wang

Publications and source records attributed to Wenyuan Wang.

At least 19 recordsLinked to original sources

arXiv2026

Optimal credit portfolio and consumption with regime switching and default contagion

We study an optimal portfolio and consumption problem in a regime-switching multi-name credit market with default contagion. Default events not only generate direct portfolio losses but also alter the default intensities of surviving securities. Under a Cobb-Douglas utility, the homogeneity property reduces the associated Hamilton-Jacobi-Bellman (HJB) equation to a recursive system of ordinary differential equations indexed by the default states. Solving this system backward from the all-default state, we establish existence and uniqueness of a positive classical solution, characterize the optimal feedback controls, and provide a verification theorem. Finally, numerical experiments present sensitivity analyses and comparisons across contagion settings, regimes, utilities, as well as before and after default.

q-fin.MF↗

arXiv2026

Probabilistic Residual Learning for Online Recommendations

Modern recommender systems are typically based on deep learning (DL) models, where a dense encoder learns representations of users and items. As a result, these systems often suffer from the black-box nature and computational complexity of the underlying models, making it difficult to systematically enhance their recommendation capabilities. To address this problem, we propose Probabilistic Residual Learning (PRL), a causal Bayesian recommendation model that models the residual between ground-truth and base predictions, enabling targeted refinement of existing systems. Specifically, PRL (1) probabilistically groups users for localized residual modeling, (2) models domain-level confounders that influence user and item representations, and (3) aggregates cluster-specific residual predictions over the confounders using do-calculus. Experiments demonstrate that our plug-and-play PRL is compatible with various base deep learning recommender systems, improving their performance while automatically discovering meaningful user clusters.

cs.IR↗

arXiv2026

Equilibrium analysis in a multi-agent reinsurance chain

This paper investigates a multi-layer reinsurance chain within a stochastic differential game framework involving m competing insurers and n reinsurers. Specifically, Stackelberg differential games are employed to characterize the strategic interactions between reinsurance buyers and sellers at each layer of the chain. In addition, a non-zero-sum game model is established to capture the competitive behavior among insurers. Both insurers and reinsurers are allowed to invest in a risk-free asset and a risky asset. To examine the heterogeneity of reinsurance chains under different contract types, the analysis is conducted separately for proportional reinsurance and excess-of-loss reinsurance. By combining dynamic programming and game theory, closed-form equilibrium strategies for investment and reinsurance are derived by solving the extended Hamilton-Jacobi-Bellman (HJB) systems under the mean-variance (MV) criterion. Numerical analysis is conducted to explore the impact of key parameters on the equilibrium strategies. The results indicate that intensified competition in the insurance market leads to a reduction in the safety loadings of reinsurance contracts at each layer of the reinsurance chain.

q-fin.MF↗

arXiv2026

Shortcuts to adiabaticity in five-level systems using counter-diabatic driving and time-rescaling optimization

Shortcuts to adiabaticity (STA) is a common protocol to realize high-fidelity and robust quantum control in various quantum systems. To date, STA has been widely applied in two- and three-level systems, whereas designing feasible strategies to achieve perfect quantum state engineering in multi-level systems still remains a challenging task. Here, we propose to use counterdiabatic (CD) driving and time-rescaling (TR) methods to construct multi-state stimulated Raman shortcut-to-adiabatic passage protocols for realizing robust and fast population transfer in chainwise-connected five-state systems. The first protocol is implemented by initially reducing the original five-state system to an equivalent two-state counterpart, and then designing the corresponding driving field by combining CD driving and unitary transformation. Further, we introduce the TR method to optimize the first protocol and thus offers an alternative solution. Numerical calculations show that both protocols can achieve complete population transfer and effectively suppress transient populations of all intermediate states. Compared with the first protocol, the optimized second exhibits better performance within a shorter evolution time.

quant-ph↗

arXiv2026

PhysMRV: Physical Memory Retrieval and Verification for Physics Plausibility Reasoning

Video-language models (VLMs) have achieved remarkable performance on video understanding and visual question answering, yet they remain unreliable in reasoning about physical plausibility, where understanding object interactions, causal dynamics, and fundamental physical principles is essential. This limitation is particularly evident on challenging physical reasoning benchmarks, revealing a persistent gap in physical commonsense reasoning. To address this challenge, we propose PhysMRV, a training-free physical memory and verification framework for physical plausibility reasoning. Unlike retrieval-augmented VLMs that retrieve semantically similar videos as additional context, PhysMRV transforms training videos into a Hierarchical Memory Bank of structured physical knowledge comprising three complementary levels: scene descriptions capturing visual context, physical-event graphs modeling object interactions and causal structure, and physics-rule summaries distilling reusable physical principles and cues. During inference, PhysMRV retrieves physically relevant memories and leverages their structured physical evidence to guide a frozen VLM in verifying physical plausibility, requiring neither fine-tuning nor parameter updates. We evaluate PhysMRV on three challenging physical reasoning benchmarks, ImplausiBench, IntPhys2, and GRASP Level 2, across multiple state-of-the-art VLMs. Experimental results demonstrate consistent improvements over direct prompting across diverse VLMs and evaluation benchmarks, showing that structured physical memories provide an effective and scalable means of enhancing physical plausibility reasoning without additional training.

cs.LG↗

arXiv2026

Optimal portfolio under ratio-type periodic evaluation in stochastic factor models under convex trading constraints

This paper studies a type of periodic utility maximization problem for portfolio management in incomplete stochastic factor models with convex trading constraints. The portfolio performance is periodically evaluated on the relative ratio of two adjacent wealth levels over an infinite horizon, featuring the dynamic adjustments in portfolio decision according to past achievements. Under power utility, we transform the original infinite horizon optimal control problem into an auxiliary terminal wealth optimization problem under a modified utility function. To cope with the convex trading constraints, we further introduce an auxiliary unconstrained optimization problem in a modified market model and develop the martingale duality approach to establish the existence of the dual minimizer such that the optimal unconstrained wealth process can be obtained using the dual representation. With the help of the duality results in the auxiliary problems, the relationship between the constrained and unconstrained models as well as some fixed point arguments, we derive and verify the optimal constrained portfolio process for the original problem over an infinite horizon.

q-fin.MF↗

arXiv2026

Fully Passive Quantum Conference Key Agreement

Quantum Conference Key Agreement (CKA) provides a secure method for multi-party communication. A recently developed interference-based prepare-and-measure quantum CKA possesses the advantages of measurement-device-independence, namely, being immune to side-channels from the detector side. Besides, it achieves good key rate performance, especially for high-loss channels, due to the use of single photon interference. Meanwhile, several fully passive QKD schemes have been proposed, which eliminate all side channels from the source modulation side. We extend the fully passive idea to an interference-based CKA, which has a high level of implementation security for many-user communication.

quant-ph↗

arXiv2026

Fully-Passive Twin-Field Quantum Key Distribution

We propose a fully passive twin-field quantum key distribution (QKD) setup where basis choice, decoy-state preparation and encoding are all implemented entirely by post-processing without any active modulation. Our protocol can remove the potential side-channels from both source modulators and detectors, and additionally retain the high key rate advantage offered by twin-field QKD, thus offering great implementation security and good performance. Importantly, we also propose a post-processing strategy that uses mismatched phase slices and minimizes the effect of sifting. We show with numerical simulation that the new protocol can still beat the repeaterless bound and provide satisfactory key rate.

quant-ph↗

arXiv2026

PhysProver: Advancing Automatic Theorem Proving for Physics

The combination of verifiable languages and LLMs has significantly influenced both the mathematical and computer science communities because it provides a rigorous foundation for theorem proving. Recent advancements in the field provide foundation models and sophisticated agentic systems pushing the boundaries of formal mathematical reasoning to approach the natural language capability of LLMs. However, little attention has been given to the formal physics reasoning, which also heavily relies on similar problem-solving and theorem-proving frameworks. To solve this problem, this paper presents, to the best of our knowledge, the first approach to enhance formal theorem proving in the physics domain. We compose a dedicated dataset PhysLeanData for the task. It is composed of theorems sampled from PhysLean and data generated by a conjecture-based formal data generation pipeline. In the training pipeline, we leverage DeepSeek-Prover-V2-7B, a strong open-source mathematical theorem prover, and apply Reinforcement Learning with Verifiable Rewards (RLVR) to train our model PhysProver. Comprehensive experiments demonstrate that, using only $\sim$5K training samples, PhysProver achieves an overall 2.4\% improvement in multiple sub-domains. Furthermore, after formal physics training, we observe 1.3\% gains on the MiniF2F-Test benchmark, which indicates non-trivial generalization beyond physics domains and enhancement for formal math capability as well. The results highlight the effectiveness and efficiency of our approach, which provides a paradigm for extending formal provers outside mathematical domains. To foster further research, we will release both our dataset and model to the community.

cs.AI↗

arXiv2025

De Finetti's problem with fixed transaction costs and regime switching

In this paper, we examine a modified version of de Finetti's optimal dividend problem, incorporating fixed transaction costs and altering the surplus process by introducing two-valued drift and two-valued volatility coefficients. This modification aims to capture the transitions or adjustments in the company's financial status. We identify the optimal dividend strategy, which maximizes the expected total net dividend payments (after accounting for transaction costs) until ruin, as a two-barrier impulsive dividend strategy. Notably, the optimal strategy can be explicitly determined for almost all scenarios involving different drifts and volatility coefficients. Our primary focus is on exploring how changes in drift and volatility coefficients influence the optimal dividend strategy.

q-fin.MF↗

arXiv2025

Generalizable Geometric Image Caption Synthesis

Multimodal large language models have various practical applications that demand strong reasoning abilities. Despite recent advancements, these models still struggle to solve complex geometric problems. A key challenge stems from the lack of high-quality image-text pair datasets for understanding geometric images. Furthermore, most template-based data synthesis pipelines typically fail to generalize to questions beyond their predefined templates. In this paper, we bridge this gap by introducing a complementary process of Reinforcement Learning with Verifiable Rewards (RLVR) into the data generation pipeline. By adopting RLVR to refine captions for geometric images synthesized from 50 basic geometric relations and using reward signals derived from mathematical problem-solving tasks, our pipeline successfully captures the key features of geometry problem-solving. This enables better task generalization and yields non-trivial improvements. Furthermore, even in out-of-distribution scenarios, the generated dataset enhances the general reasoning capabilities of multimodal large language models, yielding accuracy improvements of $2.8\%\text{-}4.8\%$ in statistics, arithmetic, algebraic, and numerical tasks with non-geometric input images of MathVista and MathVerse, along with $2.4\%\text{-}3.9\%$ improvements in Art, Design, Tech, and Engineering tasks in MMMU.

cs.AI↗

arXiv2025

Multimodal Needle in a Haystack: Benchmarking Long-Context Capability of Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) have shown significant promise in various applications, leading to broad interest from researchers and practitioners alike. However, a comprehensive evaluation of their long-context capabilities remains underexplored. To address these gaps, we introduce the MultiModal Needle-in-a-haystack (MMNeedle) benchmark, specifically designed to assess the long-context capabilities of MLLMs. Besides multi-image input, we employ image stitching to further increase the input context length, and develop a protocol to automatically generate labels for sub-image level retrieval. Essentially, MMNeedle evaluates MLLMs by stress-testing their capability to locate a target sub-image (needle) within a set of images (haystack) based on textual instructions and descriptions of image contents. This setup necessitates an advanced understanding of extensive visual contexts and effective information retrieval within long-context image inputs. With this benchmark, we evaluate state-of-the-art MLLMs, encompassing both API-based and open-source models. The findings reveal that GPT-4o consistently surpasses other models in long-context scenarios, but suffers from hallucination problems in negative samples, i.e., when needles are not in the haystacks. Our comprehensive long-context evaluation of MLLMs also sheds lights on the considerable performance gap between API-based and open-source models. All the code, data, and instructions required to reproduce the main results are available at https://github.com/Wang-ML-Lab/multimodal-needle-in-a-haystack.

cs.LG↗

arXiv2025

De Finetti's Poissonian Dividend Control Problem under Spectrally Positive Markov Additive Process

We study a De Finetti's optimal dividend and capital injection problem under a Markov additive model. The surplus process without dividend and capital injection is assumed to follow a spectrally positive Markov additive process (MAP). Dividend payments are made at the jump times of an independent Poisson process and capitals are injected to avoid bankruptcy. The aim of the paper is to characterize an optimal dividend and capital injection strategy that maximizes the expected total discounted dividends subtracted by the total discounted costs of capital injection. Applying the fluctuation and excursion theory for Levy processes and the stochastic control theory, we first address an auxiliary dividend and capital injection control problem with a terminal payoff under the spectrally positive Levy model. Using results obtained for this auxiliary problem and a fixed point argument for iterations induced by dynamic program, we characterize the optimal strategy of our prime control problem as a regime-modulated double-barrier Poissonian-continuous-reflection dividend and capital injection strategy. Besides, a numerical example is provided to illustrate the features of the optimal strategies. The impacts of model parameters are also studied.

math.OC↗

arXiv2024

Continual Learning of Large Language Models: A Comprehensive Survey

The recent success of large language models (LLMs) trained on static, pre-collected, general datasets has sparked numerous research directions and applications. One such direction addresses the non-trivial challenge of integrating pre-trained LLMs into dynamic data distributions, task structures, and user preferences. Pre-trained LLMs, when tailored for specific needs, often experience significant performance degradation in previous knowledge domains -- a phenomenon known as "catastrophic forgetting". While extensively studied in the continual learning (CL) community, it presents new manifestations in the realm of LLMs. In this survey, we provide a comprehensive overview of the current research progress on LLMs within the context of CL. This survey is structured into four main sections: we first describe an overview of continually learning LLMs, consisting of two directions of continuity: vertical continuity (or vertical continual learning), i.e., continual adaptation from general to specific capabilities, and horizontal continuity (or horizontal continual learning), i.e., continual adaptation across time and domains (Section 3). We then summarize three stages of learning LLMs in the context of modern CL: Continual Pre-Training (CPT), Domain-Adaptive Pre-training (DAP), and Continual Fine-Tuning (CFT) (Section 4). Then we provide an overview of evaluation protocols for continual learning with LLMs, along with the current available data sources (Section 5). Finally, we discuss intriguing questions pertaining to continual learning for LLMs (Section 6). The full list of papers examined in this survey is available at https://github.com/Wang-ML-Lab/llm-continual-learning-survey.

cs.LG↗

arXiv2024

Causal Priors and Their Influence on Judgements of Causality in Visualized Data

"Correlation does not imply causation" is a famous mantra in statistical and visual analysis. However, consumers of visualizations often draw causal conclusions when only correlations between variables are shown. In this paper, we investigate factors that contribute to causal relationships users perceive in visualizations. We collected a corpus of concept pairs from variables in widely used datasets and created visualizations that depict varying correlative associations using three typical statistical chart types. We conducted two MTurk studies on (1) preconceived notions on causal relations without charts, and (2) perceived causal relations with charts, for each concept pair. Our results indicate that people make assumptions about causal relationships between pairs of concepts even without seeing any visualized data. Moreover, our results suggest that these assumptions constitute causal priors that, in combination with visualized association, impact how data visualizations are interpreted. The results also suggest that causal priors may lead to over- or under-estimation in perceived causal relations in different circumstances, and that those priors can also impact users' confidence in their causal assessments. In addition, our results align with prior work, indicating that chart type may also affect causal inference. Using data from the studies, we develop a model to capture the interaction between causal priors and visualized associations as they combine to impact a user's perceived causal relations. In addition to reporting the study results and analyses, we provide an open dataset of causal priors for 56 specific concept pairs that can serve as a potential benchmark for future studies. We also suggest remaining challenges and heuristic-based guidelines to help designers improve visualization design choices to better support visual causal inference.

cs.HC↗

arXiv2024

Multi-Vehicle Trajectory Planning at V2I-enabled Intersections based on Correlated Equilibrium

Generating trajectories that ensure both vehicle safety and improve traffic efficiency remains a challenging task at intersections. Many existing works utilize Nash equilibrium (NE) for the trajectory planning at intersections. However, NE-based planning can hardly guarantee that all vehicles are in the same equilibrium, leading to a risk of collision. In this work, we propose a framework for trajectory planning based on Correlated Equilibrium (CE) when V2I communication is also enabled. The recommendation with CE allows all vehicles to reach a safe and consensual equilibrium and meanwhile keeps the rationality as NE-based methods that no vehicle has the incentive to deviate. The Intersection Manager (IM) first collects the trajectory library and the personal preference probabilities over the library from each vehicle in a low-resolution spatial-temporal grid map. Then, the IM optimizes the recommendation probability distribution for each vehicle's trajectory by minimizing overall collision probability under the CE constraint. Finally, each vehicle samples a trajectory of the low-resolution map to construct a safety corridor and derive a smooth trajectory with a local refinement optimization. We conduct comparative experiments at a crossroad intersection involving two and four vehicles, validating the effectiveness of our method in balancing vehicle safety and traffic efficiency.

cs.RO↗

arXiv2024

Resource-Efficient Real-Time Polarization Compensation for MDI-QKD with Rejected Data

Measurement-device-independent quantum key distribution (MDI-QKD) closes all the security loopholes in the detection system and is a promising solution for secret key sharing. Polarization encoding is the most common QKD encoding scheme, as it is straightforward to prepare and measure. However, implementing polarization encoding in MDI QKD imposes extra challenges, as polarization alignment must be maintained over both mutually unbiased bases and be maintained for both paths (Alice-Charlie and Bob-Charlie). Polarization alignment is usually done by interrupting the QKD process (reducing overall key generation rates) or using additional classical laser sources multiplexed with quantum channels for polarization alignment. Since low key rates and cost are the two most pressing challenges preventing wide adoption of QKD systems, using additional resources or reducing key rates runs contrary to making QKD commercially viable. Therefore, we propose and implement a novel polarization compensation scheme in the MDI-QKD system that avoids the aforementioned drawbacks by recycling part of discarded detection events. Our scheme evaluates the polarization drift in real-time based on single measurements corresponding to decoy intensities. Our fully automated experimental demonstration maintains the average polarization drift below 0.13 rad over 40 km of spooled fibre (without an insulating jacket) for at least four hours. The average quantum bit error rate is 3.8$\%$, and we achieved an average key rate of $7.45\times 10^{-6}$ bits per pulse.

quant-ph↗

arXiv2024

Multi-Tailed Vision Transformer for Efficient Inference

Recently, Vision Transformer (ViT) has achieved promising performance in image recognition and gradually serves as a powerful backbone in various vision tasks. To satisfy the sequential input of Transformer, the tail of ViT first splits each image into a sequence of visual tokens with a fixed length. Then the following self-attention layers constructs the global relationship between tokens to produce useful representation for the downstream tasks. Empirically, representing the image with more tokens leads to better performance, yet the quadratic computational complexity of self-attention layer to the number of tokens could seriously influence the efficiency of ViT's inference. For computational reduction, a few pruning methods progressively prune uninformative tokens in the Transformer encoder, while leaving the number of tokens before the Transformer untouched. In fact, fewer tokens as the input for the Transformer encoder can directly reduce the following computational cost. In this spirit, we propose a Multi-Tailed Vision Transformer (MT-ViT) in the paper. MT-ViT adopts multiple tails to produce visual sequences of different lengths for the following Transformer encoder. A tail predictor is introduced to decide which tail is the most efficient for the image to produce accurate prediction. Both modules are optimized in an end-to-end fashion, with the Gumbel-Softmax trick. Experiments on ImageNet-1K demonstrate that MT-ViT can achieve a significant reduction on FLOPs with no degradation of the accuracy and outperform other compared methods in both accuracy and FLOPs.

cs.CV↗