Jingwei Li

Kimi Team, Tongtong Bai, Yifan Bai et al.

We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.

Bo Yang, Jingwei Li, Qiaoni Han et al.

Severe pollution induced by traditional fossil fuels arouses great attention on the usage of plug-in electric vehicles (PEVs) and renewable energy. However, large-scale penetration of PEVs combined with other kinds of appliances tends to cause excessive or even disastrous burden on the power grid, especially during peak hours. This paper focuses on the scheduling of PEVs charging process among different charging stations and each station can be supplied by both renewable energy generators and a distribution network. The distribution network also powers some uncontrollable loads. In order to minimize the on-grid energy cost with local renewable energy and non-ideal storage while avoiding the overload risk of the distribution network, an online algorithm consisting of scheduling the charging of PEVs and energy management of charging stations is developed based on Lyapunov optimization and Lagrange dual decomposition techniques. The algorithm can satisfy the random charging requests from PEVs with provable performance. Simulation results with real data demonstrate that the proposed algorithm can decrease the time-average cost of stations while avoiding overload in the distribution network in the presence of random uncontrollable loads.

Wei Fu, Weihua Du, Jingwei Li et al. · cmu

In complex reinforcement learning (RL) problems, policies with similar rewards may have substantially different behaviors. It remains a fundamental challenge to optimize rewards while also discovering as many diverse strategies as possible, which can be crucial in many practical applications. Our study examines two design choices for tackling this challenge, i.e., diversity measure and computation framework. First, we find that with existing diversity measures, visually indistinguishable policies can still yield high diversity scores. To accurately capture the behavioral difference, we propose to incorporate the state-space distance information into the diversity measure. In addition, we examine two common computation frameworks for this problem, i.e., population-based training (PBT) and iterative learning (ITR). We show that although PBT is the precise problem formulation, ITR can achieve comparable diversity scores with higher computation efficiency, leading to improved solution quality in practice. Based on our analysis, we further combine ITR with two tractable realizations of the state-distance-based diversity measures and develop a novel diversity-driven RL algorithm, State-based Intrinsic-reward Policy Optimization (SIPO), with provable convergence properties. We empirically examine SIPO across three domains from robot locomotion to multi-agent games. In all of our testing environments, SIPO consistently produces strategically diverse and human-interpretable policies that cannot be discovered by existing baselines.

Jing Dong, Jingwei Li, Baoxiang Wang et al.

Reinforcement learning (RL) has exceeded human performance in many synthetic settings such as video games and Go. However, real-world deployment of end-to-end RL models is less common, as RL models can be very sensitive to slight perturbation of the environment. The robust Markov decision process (MDP) framework -- in which the transition probabilities belong to an uncertainty set around a nominal model -- provides one way to develop robust models. While previous analysis shows RL algorithms are effective assuming access to a generative model, it remains unclear whether RL can be efficient under a more realistic online setting, which requires a careful balance between exploration and exploitation. In this work, we consider online robust MDP by interacting with an unknown nominal system. We propose a robust optimistic policy optimization algorithm that is provably efficient. To address the additional uncertainty caused by an adversarial environment, our model features a new optimistic update rule derived via Fenchel conjugates. Our analysis establishes the first regret bound for online robust MDPs.

Jingwei Li, Ruitian Wu, Tzu-liang Huang et al.

Parkinson's Disease (PD) is a progressive nervous system disorder that has affected more than 5.8 million people, especially the elderly. Due to the complexity of its symptoms and its similarity to other neurological disorders, early detection requires neurologists or PD specialists to be involved, which is not accessible to most old people. Therefore, we integrate smart mobile devices with AI technologies. In this paper, we introduce the framework of our developed PD early detection system which combines different tasks evaluating both motor and non-motor symptoms. With the developed model, we help users detect PD punctually in non-clinical settings and figure out their most severe symptoms. The results are expected to be further used for PD rehabilitation guidance and detection of other neurological disorders.

Xuanqing Yu, Wangtao Sun, Jingwei Li et al.

In the realm of event prediction, temporal knowledge graph forecasting (TKGF) stands as a pivotal technique. Previous approaches face the challenges of not utilizing experience during testing and relying on a single short-term history, which limits adaptation to evolving data. In this paper, we introduce the Online Neural-Symbolic Event Prediction (ONSEP) framework, which innovates by integrating dynamic causal rule mining (DCRM) and dual history augmented generation (DHAG). DCRM dynamically constructs causal rules from real-time data, allowing for swift adaptation to new causal relationships. In parallel, DHAG merges short-term and long-term historical contexts, leveraging a bi-branch approach to enrich event prediction. Our framework demonstrates notable performance enhancements across diverse datasets, with significant Hit@k (k=1,3,10) improvements, showcasing its ability to augment large language models (LLMs) for event prediction without necessitating extensive retraining. The ONSEP framework not only advances the field of TKGF but also underscores the potential of neural-symbolic approaches in adapting to dynamic data environments.

Jingwei Li, Wei Pu

Blind image separation (BIS) refers to the inverse problem of simultaneously estimating and restoring multiple independent source images from a single observation image under conditions of unknown mixing mode and without prior knowledge of the source images. Traditional methods relying on statistical independence assumptions or CNN/GAN variants struggle to characterize complex feature distributions in real scenes, leading to estimation bias, texture distortion, and artifact residue under strong noise and nonlinear mixing. This paper innovatively introduces diffusion models into dual-channel BIS, proposing an efficient Dual-Channel Diffusion Separation Model (DCDSM). DCDSM leverages diffusion models' powerful generative capability to learn source image feature distributions and reconstruct feature structures effectively. A novel Wavelet Suppression Module (WSM) is designed within the dual-branch reverse denoising process, forming an interactive separation network that enhances detail separation by exploiting the mutual coupling noise characteristic between source images. Extensive experiments on synthetic datasets containing rain/snow and complex mixtures demonstrate that DCDSM achieves state-of-the-art performance: 1) In image restoration tasks, it obtains PSNR/SSIM values of 35.0023 dB/0.9549 and 29.8108 dB/0.9243 for rain and snow removal respectively, outperforming Histoformer and LDRCNet by 1.2570 dB/0.9272 dB (PSNR) and 0.0262/0.0289 (SSIM) on average; 2) For complex mixture separation, the restored dual-source images achieve average PSNR and SSIM of 25.0049 dB and 0.7997, surpassing comparative methods by 4.1249 dB and 0.0926. Both subjective and objective evaluations confirm DCDSM's superiority in addressing rain/snow residue removal and detail preservation challenges.

Jiayi Qin, Jingwei Li, Chuan Wu

Screen-shooting robust watermarking aims to imperceptibly embed extractable information into host images such that the watermark survives the complex distortion pipeline of screen display and camera recapture. However, achieving high extraction accuracy while maintaining satisfactory visual quality remains an open challenge, primarily because the screen-shooting channel introduces severe and entangled degradations including Moiré patterns, color-gamut shifts, perspective warping, and sensor noise. In this paper, we present an end-to-end deep learning framework that jointly optimizes watermark embedding and extraction for screen-shooting robustness. Our framework incorporates three key innovations: (i) a comprehensive noise simulation layer that faithfully models realistic screen-shooting distortions -- notably including a physically-motivated Moiré pattern generator -- enabling the network to learn robust representations against the full spectrum of capture-channel noise through adversarial training; (ii) a Just Noticeable Distortion (JND) perceptual loss function that adaptively modulates watermark embedding strength by supervising the perceptual discrepancy between the JND coefficient map and the watermark residual, thereby concentrating watermark energy in perceptually insensitive regions to maximize visual quality; and (iii) two complementary automatic localization modules -- a semantic-segmentation-based foreground extractor for captured image rectification and a symmetric noise template mechanism for anti-cropping region recovery -- that enable fully automated watermark decoding under realistic deployment conditions. Extensive experiments demonstrate that our method achieves an average PSNR of 30.94~dB and SSIM of 0.94 on watermarked images while embedding 127-bit payloads.

Anxin Guo, Jingwei Li

Large language models often hallucinate with high confidence on "random facts" that lack inferable patterns. We formalize the memorization of such facts as a membership testing problem, unifying the discrete error metrics of Bloom filters with the continuous log-loss of LLMs. By analyzing this problem in the regime where facts are sparse in the universe of plausible claims, we establish a rate-distortion theorem: the optimal memory efficiency is characterized by the minimum KL divergence between score distributions on facts and non-facts. This theoretical framework provides a distinctive explanation for hallucination: even with optimal training, perfect data, and a simplified "closed world" setting, the information-theoretically optimal strategy under limited capacity is not to abstain or forget, but to assign high confidence to some non-facts, resulting in hallucination. We validate this theory empirically on synthetic data, showing that hallucinations persist as a natural consequence of lossy compression.

Jingwei Li

In this paper, we study two problems: determining action model equivalence and minimizing the event space of an action model under certain structural relationships. The Kripke model equivalence is perfectly caught by the structural relationship called bisimulation. In this paper, we propose the generalized action emulation perfectly catching the action model equivalence. Previous structural relationships sufficient for the action model equivalence, i.e. the bisimulation, the propositional action emulation, the action emulation, and the action emulation of canonical action models, can be described by various restricted versions of the generalized action emulation. We summarize four critical properties of the atom set over preconditions, and prove that any formula set satisfying these properties can be used to restrict the generalized action emulation to determine the action model equivalence by an iteration algorithm. We also construct a new formula set with these four properties, which is generally more efficient than the atom set. The technique of the partition refinement has been used to minimize the world space of a Kripke model under the bisimulation. Applying the partition refinement to action models allows one to minimize their event spaces under the bisimulation. The propositional action emulation is weaker than bismulation but still sufficient for the action model equivalence. We prove that it is PSPACE-complete to minimize the event space of an action model under the propositional action emulation, and provide a PSPACE algorithm for it. Finally, we prove that minimize the event space under the action model equivalence is PSPACE-hard, and propose a computable method based on the canonical formulas of modal logics to solve this problem.

Jingwei Li, Xinran Gu, Jingzhao Zhang

A data mixture refers to how different data sources are combined to train large language models, and selecting an effective mixture is crucial for optimal downstream performance. Existing methods either conduct costly searches directly on the target model or rely on mixture scaling laws that fail to extrapolate well to large model sizes. We address these limitations by introducing a compute-efficient pipeline for data mixture scaling. First, we propose CAMEL, a capacity-aware mixture law that models validation loss with the nonlinear interplay between model size and mixture. We also introduce a loss-to-benchmark prediction law that estimates benchmark accuracy from validation loss, enabling end-to-end performance prediction for the target model. Next, we study how to allocate a fixed compute budget across model scales to fit the law and reduce prediction error. Finally, we apply our method to Mixture-of-Experts models with up to 7B-A150M parameters to fit the law, and verify the optimal mixture derived from the law by extrapolating to a 55B-A1.2B target model. Compared to prior methods, we reduces mixture optimization costs by 50\% and improves downstream benchmark performance by up to 3\%.

Lei Deng, Wenhan Xu, Jingwei Li et al.

Real-time network traffic forecasting is crucial for network management and early resource allocation. Existing network traffic forecasting approaches operate under the assumption that the network traffic data is fully observed. However, in practical scenarios, the collected data are often incomplete due to various human and natural factors. In this paper, we propose a generative model approach for real-time network traffic forecasting with missing data. Firstly, we model the network traffic forecasting task as a tensor completion problem. Secondly, we incorporate a pre-trained generative model to achieve the low-rank structure commonly associated with tensor completion. The generative model effectively captures the intrinsic low-rank structure of network traffic data during pre-training and enables the mapping from a compact latent representation to the tensor space. Thirdly, rather than directly optimizing the high-dimensional tensor, we optimize its latent representation, which simplifies the optimization process and enables real-time forecasting. We also establish a theoretical recovery guarantee that quantifies the error bound of the proposed approach. Experiments on real-world datasets demonstrate that our approach achieves accurate network traffic forecasting within 100 ms, with a mean absolute error (MAE) below 0.002, as validated on the Abilene dataset.

Xujun Wei, Feng Zhang, Renhe Zhang et al.

In the past few years, Artificial Intelligence (AI)-based weather forecasting methods have widely demonstrated strong competitiveness among the weather forecasting systems. However, these methods are insufficient for high-spatial-resolution short-term nowcasting within 6 hours, which is crucial for warning short-duration, mesoscale and small-scale weather events. Geostationary satellite remote sensing provides detailed, high spatio-temporal and all-day observations, which can address the above limitations of existing methods. Therefore, this paper proposed an advanced data-driven thermal infrared cloud images forecasting model, "DaYu." Unlike existing data-driven weather forecasting models, DaYu is specifically designed for geostationary satellite observations, with a temporal resolution of 0.5 hours and a spatial resolution of ${0.05}^\circ$ $\times$ ${0.05}^\circ$. DaYu is based on a large-scale transformer architecture, which enables it to capture fine-grained cloud structures and learn fast-changing spatio-temporal evolution features effectively. Moreover, its attention mechanism design achieves a balance in computational complexity, making it practical for applications. DaYu not only achieves accurate forecasts up to 3 hours with a correlation coefficient higher than 0.9, 6 hours higher than 0.8, and 12 hours higher than 0.7, but also detects short-duration, mesoscale, and small-scale weather events with enhanced detail, effectively addressing the shortcomings of existing methods in providing detailed short-term nowcasting within 6 hours. Furthermore, DaYu has significant potential in short-term climate disaster prevention and mitigation.

Jingwei Li

Statistic modeling and data-driven learning are the two vital fields that attract many attentions. Statistic models intend to capture and interpret the relationships among variables, while data-based learning attempt to extract information directly from the data without pre-processing through complex models. Given the extensive studies in both fields, a subtle issue is how to properly integrate data based methods with existing knowledge or models. In this paper, based on the time series data, we propose two different directions to integrate the two, a decomposition-based method and a method exploiting the statistic extraction of data features. The first one decomposes the data into linear stable, nonlinear stable and unstable parts, where suitable statistical models are used for the linear stable and nonlinear stable parts while the appropriate machine learning tools are used for the unstable parts. The second one applies statistic models to extract statistics features of data and feed them as additional inputs into the machine learning platform for training. The most critical and challenging thing is how to determine and extract the valuable information from mathematical or statistical models to boost the performance of machine learning algorithms. We evaluate the proposal using time series data with varying degrees of stability. Performance results show that both methods can outperform existing schemes that use models and learning separately, and the improvements can be over 60%. Both our proposed methods are promising in bridging the gap between model-based and data-driven schemes and integrating the two to provide an overall higher learning performance.

Jingwei Li, Patrick P. C. Lee, Chufeng Tan et al.

Encrypted deduplication combines encryption and deduplication to simultaneously achieve both data security and storage efficiency. State-of-the-art encrypted deduplication systems mainly build on deterministic encryption to preserve deduplication effectiveness. However, such deterministic encryption reveals the underlying frequency distribution of the original plaintext chunks. This allows an adversary to launch frequency analysis against the ciphertext chunks and infer the content of the original plaintext chunks. In this paper, we study how frequency analysis affects information leakage in encrypted deduplication storage, from both attack and defense perspectives. Specifically, we target backup workloads, and propose a new inference attack that exploits chunk locality to increase the coverage of inferred chunks. We further combine the new inference attack with the knowledge of chunk sizes and show its attack effectiveness against variable-size chunks. We conduct trace-driven evaluation on both real-world and synthetic datasets and show that our proposed attacks infer a significant fraction of plaintext chunks under backup workloads. To defend against frequency analysis, we present two defense approaches, namely MinHash encryption and scrambling. Our trace-driven evaluation shows that our combined MinHash encryption and scrambling scheme effectively mitigates the severity of the inference attacks, while maintaining high storage efficiency and incurring limited metadata access overhead.

Chuan Qin, Jingwei Li, Patrick P. C. Lee

Rekeying refers to an operation of replacing an existing key with a new key for encryption. It renews security protection, so as to protect against key compromise and enable dynamic access control in cryptographic storage. However, it is non-trivial to realize efficient rekeying in encrypted deduplication storage systems, which use deterministic content-derived encryption keys to allow deduplication on ciphertexts. We design and implement REED, a rekeying-aware encrypted deduplication storage system. REED builds on a deterministic version of all-or-nothing transform (AONT), such that it enables secure and lightweight rekeying, while preserving the deduplication capability. We propose two REED encryption schemes that trade between performance and security, and extend REED for dynamic access control. We implement a REED prototype with various performance optimization techniques and demonstrate how we can exploit similarity to mitigate key generation overhead. Our trace-driven testbed evaluation shows that our REED prototype maintains high performance and storage efficiency.

Ezilda Almeida, Pedro Ferreira, Tiago Vinhoza et al.

Bayesian network structures are usually built using only the data and starting from an empty network or from a naive Bayes structure. Very often, in some domains, like medicine, a prior structure knowledge is already known. This structure can be automatically or manually refined in search for better performance models. In this work, we take Bayesian networks built by specialists and show that minor perturbations to this original network can yield better classifiers with a very small computational cost, while maintaining most of the intended meaning of the original model.