Yadong Zhang

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.

Jin Cui, Xinyue Long, Xunyong Zhang et al.

Multimodal Large Language Models (MLLMs) have made remarkable progress on vision-language reasoning, yet most methods still compress visual evidence into discrete textual thoughts, creating an information bottleneck for fine-grained perception. Recent latent visual reasoning methods attempt to reason in continuous hidden states, but we find that they suffer from insufficient manifold compatibility: latent trajectories drift away from pretrained reasoning circuits, collapse into instance-agnostic patterns, and are often bypassed during answer generation. To address these issues, we propose RIS (Retrieve, Integrate, and Synthesize), a spatial-semantic grounded framework that develops latent reasoning as a compatible extension of pretrained MLLM computation. We first construct a step-wise grounded reasoning dataset with bounding boxes and region-specific semantic descriptions. Built on this supervision, RIS anchors latent tokens to both spatial and semantic evidence, enforces their causal role through a progressive attention bottleneck, and introduces short language transition tokens to bridge synthesized latent states back to vocabulary-aligned decoding. Experiments on V*, HRBench4K, HRBench8K, MMVP, and BLINK show consistent improvements over closed/open-source and latent reasoning baselines. Further analyses demonstrate that RIS learns diverse, interpretable, and progressively integrated latent trajectories, offering a practical path toward faithful internal visual reasoning in MLLMs.

Yadong Zhang, Shaoguang Mao, Wenshan Wu et al.

This paper introduces BI-Directional DEliberation Reasoning (BIDDER), a novel reasoning approach to enhance the decision rationality of language models. Traditional reasoning methods typically rely on historical information and employ uni-directional (left-to-right) reasoning strategy. This lack of bi-directional deliberation reasoning results in limited awareness of potential future outcomes and insufficient integration of historical context, leading to suboptimal decisions. BIDDER addresses this gap by incorporating principles of rational decision-making, specifically managing uncertainty and predicting expected utility. Our approach involves three key processes: Inferring hidden states to represent uncertain information in the decision-making process from historical data; Using these hidden states to predict future potential states and potential outcomes; Integrating historical information (past contexts) and long-term outcomes (future contexts) to inform reasoning. By leveraging bi-directional reasoning, BIDDER ensures thorough exploration of both past and future contexts, leading to more informed and rational decisions. We tested BIDDER's effectiveness in two well-defined scenarios: Poker (Limit Texas Hold'em) and Negotiation. Our experiments demonstrate that BIDDER significantly improves the decision-making capabilities of LLMs and LLM agents.

Yadong Zhang, Pranav M Karve, Sankaran Mahadevan

We investigate the utility of graph neural networks (GNNs) as proxies of power grid operational decision-making algorithms (optimal power flow (OPF) and security-constrained unit commitment (SCUC)) to enable rigorous quantification of the operational risk. To conduct principled risk analysis, numerous Monte Carlo (MC) samples are drawn from the (foretasted) probability distributions of spatio-temporally correlated stochastic grid variables. The corresponding OPF and SCUC solutions, which are needed to quantify the risk, are generated using traditional OPF and SCUC solvers to generate data for training GNN model(s). The GNN model performance is evaluated in terms of the accuracy of predicting quantities of interests (QoIs) derived from the decision variables in OPF and SCUC. Specifically, we focus on thermal power generation and load shedding at system and individual zone level. We also perform reliability and risk quantification based on GNN predictions and compare with that obtained from OPF/SCUC solutions. Our results demonstrate that GNNs are capable of providing fast and accurate prediction of QoIs and thus can be good surrogate models for OPF and SCUC. The excellent accuracy of GNN-based reliability and risk assessment further suggests that GNN surrogate has the potential to be applied in real-time and hours-ahead risk quantification.

Yadong Zhang, Pranav M Karve, Sankaran Mahadevan

A DC OPF surrogate modeling framework is developed for Monte Carlo (MC) sampling-based risk quantification in power grid operation. MC simulation necessitates solving a large number of DC OPF problems corresponding to the samples of stochastic grid variables (power demand and renewable generation), which is computationally prohibitive. Computationally inexpensive surrogates of OPF provide an attractive alternative for expedited MC simulation. Graph neural network (GNN) surrogates of DC OPF, which are especially suitable to graph-structured data, are employed in this work. Previously developed DC OPF surrogate models have focused on accurate operational decision-making and not on risk quantification. Here, risk quantification-specific aspects of DC OPF surrogate evaluation is the main focus. To this end, the proposed GNN surrogates are evaluated using realistic joint probability distributions, quantification of their risk estimation accuracy, and investigation of their generalizability. Four synthetic grids (Case118, Case300, Case1354pegase, and Case2848rte) are used for surrogate model performance evaluation. It is shown that the GNN surrogates are sufficiently accurate for predicting the (bus-level, branch-level and system-level) grid state and enable fast as well as accurate operational risk quantification for power grids. The article thus develops tools for fast reliability and risk quantification in real-world power grids using GNN-based surrogates.

Zheqin Yin, Yupei Ren, Yadong Zhang et al.

Argumentative essays serve as a vital medium for assessing critical thinking and reasoning skills, yet there is limited works on accurately understanding and evaluating such texts via prompt. In this work, we propose TIDE, a novel framework designed to improve criteria-based prompt optimization for argument-related tasks by integrating TrIal and DEbate mechanism. Our method addresses key limitations of criteria-based prompt optimizing by mitigating the influence of noisy training data and enhancing optimization stability. We evaluate TIDE on three core tasks: Automated Essay Scoring, Argument Component Detection, and Argument Relation Identification. Results demonstrate that our framework improves performance across tasks. These findings underscore the potential of combining prompt-based methods for advanced argument understanding.

Haiduo Huang, Jiangcheng Song, Yadong Zhang et al.

Recent advances in knowledge distillation have emphasized the importance of decoupling different knowledge components. While existing methods utilize momentum mechanisms to separate task-oriented and distillation gradients, they overlook the inherent conflict between target-class and non-target-class knowledge flows. Furthermore, low-confidence dark knowledge in non-target classes introduces noisy signals that hinder effective knowledge transfer. To address these limitations, we propose DeepKD, a novel training framework that integrates dual-level decoupling with adaptive denoising. First, through theoretical analysis of gradient signal-to-noise ratio (GSNR) characteristics in task-oriented and non-task-oriented knowledge distillation, we design independent momentum updaters for each component to prevent mutual interference. We observe that the optimal momentum coefficients for task-oriented gradient (TOG), target-class gradient (TCG), and non-target-class gradient (NCG) should be positively related to their GSNR. Second, we introduce a dynamic top-k mask (DTM) mechanism that gradually increases K from a small initial value to incorporate more non-target classes as training progresses, following curriculum learning principles. The DTM jointly filters low-confidence logits from both teacher and student models, effectively purifying dark knowledge during early training. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO demonstrate DeepKD's effectiveness. Our code is available at https://github.com/haiduo/DeepKD.

Huajian Feng, Guoxiao Zhang, Yadong Zhang et al.

Accurate post-click conversion rate (CVR) estimation is crucial for online advertising systems. Despite significant advances in causal approaches designed to address the Sample Selection Bias problem, CVR estimation still faces challenges due to Covariate Shift. Given the intrinsic connection between the distribution of covariates in the click and non-click spaces, this study proposes an Exposure-Guided Embedding Alignment Network (EGEAN) to address estimation bias caused by covariate shift. Additionally, we propose a Parameter Varying Doubly Robust Estimator with steady-state control to handle small propensities better. Online A/B tests conducted on the Meituan advertising system demonstrate that our method significantly outperforms baseline models with respect to CVR and GMV, validating its effectiveness. Code is available: https://github.com/hydrogen-maker/EGEAN.

Yadong Zhang, Shaoguang Mao, Tao Ge et al.

This paper presents a comprehensive survey of the current status and opportunities for Large Language Models (LLMs) in strategic reasoning, a sophisticated form of reasoning that necessitates understanding and predicting adversary actions in multi-agent settings while adjusting strategies accordingly. Strategic reasoning is distinguished by its focus on the dynamic and uncertain nature of interactions among multi-agents, where comprehending the environment and anticipating the behavior of others is crucial. We explore the scopes, applications, methodologies, and evaluation metrics related to strategic reasoning with LLMs, highlighting the burgeoning development in this area and the interdisciplinary approaches enhancing their decision-making performance. It aims to systematize and clarify the scattered literature on this subject, providing a systematic review that underscores the importance of strategic reasoning as a critical cognitive capability and offers insights into future research directions and potential improvements.

Wenshan Wu, Shaoguang Mao, Yadong Zhang et al. · microsoft-research

Large language models (LLMs) have exhibited impressive performance in language comprehension and various reasoning tasks. However, their abilities in spatial reasoning, a crucial aspect of human cognition, remain relatively unexplored. Human possess a remarkable ability to create mental images of unseen objects and actions through a process known as the Mind's Eye, enabling the imagination of the unseen world. Inspired by this cognitive capacity, we propose Visualization-of-Thought (VoT) prompting. VoT aims to elicit spatial reasoning of LLMs by visualizing their reasoning traces, thereby guiding subsequent reasoning steps. We employed VoT for multi-hop spatial reasoning tasks, including natural language navigation, visual navigation, and visual tiling in 2D grid worlds. Experimental results demonstrated that VoT significantly enhances the spatial reasoning abilities of LLMs. Notably, VoT outperformed existing multimodal large language models (MLLMs) in these tasks. While VoT works surprisingly well on LLMs, the ability to generate mental images to facilitate spatial reasoning resembles the mind's eye process, suggesting its potential viability in MLLMs. Please find the dataset and codes at https://microsoft.github.io/visualization-of-thought

Yadong Zhang, Shaoguang Mao, Tao Ge et al.

Strategic reasoning is a complex yet essential capability for intelligent agents. It requires Large Language Model (LLM) agents to adapt their strategies dynamically in multi-agent environments. Unlike static reasoning tasks, success in these contexts depends on anticipating other agents' beliefs and actions while continuously adjusting strategies to achieve individual goals. LLMs and LLM agents often struggle with strategic reasoning due to the absence of a reasoning framework that enables them to dynamically infer others' perspectives and adapt to changing environments. Inspired by the Level-K framework from game theory and behavioral economics, which extends reasoning from simple reactions to structured strategic depth, we propose a novel framework: "K-Level Reasoning with Large Language Models (K-R)." This framework employs recursive mechanisms to enable LLMs to achieve varying levels of strategic depth, allowing agents to form higher order beliefs - beliefs about others' beliefs. We validate this framework through rigorous testing on four testbeds: two classical game theory problems and two social intelligence tasks. The results demonstrate the advantages of K-R in strategic reasoning. Our work presents the first recursive implementation of strategic depth in large language models (LLMs). It establishes a foundation for future research into theory of mind and strategic reasoning in LLMs.

Guoxiao Zhang, Yi Wei, Yadong Zhang et al.

Click-Through Rate (CTR) prediction is essential in online advertising, where semantic information plays a pivotal role in shaping user decisions and enhancing CTR effectiveness. Capturing and modeling deep semantic information, such as a user's preference for "Häagen-Dazs' HEAVEN strawberry light ice cream" due to its health-conscious and premium attributes, is challenging. Traditional semantic modeling often overlooks these intricate details at the user and item levels. To bridge this gap, we introduce a novel approach that models deep semantic information end-to-end, leveraging the comprehensive world knowledge capabilities of Large Language Models (LLMs). Our proposed LLM-infused CTR prediction framework(Multi-level Deep Semantic Information Infused CTR model via Distillation, MSD) is designed to uncover deep semantic insights by utilizing LLMs to extract and distill critical information into a smaller, more efficient model, enabling seamless end-to-end training and inference. Importantly, our framework is carefully designed to balance efficiency and effectiveness, ensuring that the model not only achieves high performance but also operates with optimal resource utilization. Online A/B tests conducted on the Meituan sponsored-search system demonstrate that our method significantly outperforms baseline models in terms of Cost Per Mile (CPM) and CTR, validating its effectiveness, scalability, and balanced approach in real-world applications.

Yadong Zhang, Pranav M Karve, Sankaran Mahadevan

This article investigates the ability of graph neural networks (GNNs) to identify risky conditions in a power grid over the subsequent few hours, without explicit, high-resolution information regarding future generator on/off status (grid topology) or power dispatch decisions. The GNNs are trained using supervised learning, to predict the power grid's aggregated bus-level (either zonal or system-level) or individual branch-level state under different power supply and demand conditions. The variability of the stochastic grid variables (wind/solar generation and load demand), and their statistical correlations, are rigorously considered while generating the inputs for the training data. The outputs in the training data, obtained by solving numerous mixed-integer linear programming (MILP) optimal power flow problems, correspond to system-level, zonal and transmission line-level quantities of interest (QoIs). The QoIs predicted by the GNNs are used to conduct hours-ahead, sampling-based reliability and risk assessment w.r.t. zonal and system-level (load shedding) as well as branch-level (overloading) failure events. The proposed methodology is demonstrated for three synthetic grids with sizes ranging from 118 to 2848 buses. Our results demonstrate that GNNs are capable of providing fast and accurate prediction of QoIs and can be good proxies for computationally expensive MILP algorithms. The excellent accuracy of GNN-based reliability and risk assessment suggests that GNN models can substantially improve situational awareness by quickly providing rigorous reliability and risk estimates.

Haiduo Huang, Jiangcheng Song, Yadong Zhang et al.

Knowledge distillation (KD) is a standard route to compress Large Language Models (LLMs) into compact students, yet most pipelines uniformly apply token-wise loss regardless of teacher confidence. This indiscriminate supervision amplifies noisy, high-entropy signals and is especially harmful under large teacher-student capacity gaps. We introduce SelecTKD, a plug-and-play Selective Token-Weighted distillation framework that shifts the focus from "how to measure divergence" to "where to apply learning". At each step, the student proposes tokens that are verified by the teacher through a robust propose-and-verify procedure with two variants: greedy Top-k and non-greedy Spec-k. Accepted tokens receive full loss, while rejected tokens are masked or down-weighted. This objective-agnostic design works with on- and off-policy data, induces an implicit curriculum quantified by Token Acceptance Rate (TAR), and stabilizes optimization. Across instruction following, mathematical reasoning, code generation, and a VLM setting, SelecTKD consistently improves strong baselines and achieves state-of-the-art results for small models without architectural changes or extra reference models.

Haiduo Huang, Yadong Zhang, Yinghui Xu et al.

Dynamic convolution enhances model capacity by adaptively combining multiple kernels, yet faces critical trade-offs: prior works either (1) incur significant parameter overhead by scaling kernel numbers linearly, (2) compromise inference speed through complex kernel interactions, or (3) struggle to jointly optimize dynamic attention and static kernels. We observe that pre-trained Convolutional Neural Networks (CNNs) exhibit inter-layer redundancy akin to that in Large Language Models (LLMs). Specifically, dense convolutional layers can be efficiently replaced by derived "child" layers generated from a shared "parent" convolutional kernel through an adapter. To address these limitations and implement the weight-sharing mechanism, we propose a lightweight convolution kernel plug-in, named KernelDNA. It decouples kernel adaptation into input-dependent dynamic routing and pre-trained static modulation, ensuring both parameter efficiency and hardware-friendly inference. Unlike existing dynamic convolutions that expand parameters via multi-kernel ensembles, our method leverages cross-layer weight sharing and adapter-based modulation, enabling dynamic kernel specialization without altering the standard convolution structure. This design preserves the native computational efficiency of standard convolutions while enhancing representation power through input-adaptive kernel adjustments. Experiments on image classification and dense prediction tasks demonstrate that KernelDNA achieves a state-of-the-art accuracy-efficiency balance among dynamic convolution variants.

Jinghao Hu, Yuhe Zhang, GuoHua Geng et al.

Traditionally, style has been primarily considered in terms of artistic elements such as colors, brushstrokes, and lighting. However, identical semantic subjects, like people, boats, and houses, can vary significantly across different artistic traditions, indicating that style also encompasses the underlying semantics. Therefore, in this study, we propose a zero-shot scheme for image variation with coordinated semantics. Specifically, our scheme transforms the image-to-image problem into an image-to-text-to-image problem. The image-to-text operation employs vision-language models e.g., BLIP) to generate text describing the content of the input image, including the objects and their positions. Subsequently, the input style keyword is elaborated into a detailed description of this style and then merged with the content text using the reasoning capabilities of ChatGPT. Finally, the text-to-image operation utilizes a Diffusion model to generate images based on the text prompt. To enable the Diffusion model to accommodate more styles, we propose a fine-tuning strategy that injects text and style constraints into cross-attention. This ensures that the output image exhibits similar semantics in the desired style. To validate the performance of the proposed scheme, we constructed a benchmark comprising images of various styles and scenes and introduced two novel metrics. Despite its simplicity, our scheme yields highly plausible results in a zero-shot manner, particularly for generating stylized images with high-fidelity semantics.

Yuwen Qin, Ningbo Cai, Chen Gao et al.

The precise estimate of remaining useful life (RUL) is vital for the prognostic analysis and predictive maintenance that can significantly reduce failure rate and maintenance costs. The degradation-related features extracted from the sensor streaming data with neural networks can dramatically improve the accuracy of the RUL prediction. The Temporal deep degradation network (TDDN) model is proposed to make the RUL prediction with the degradation-related features given by the one-dimensional convolutional neural network (1D CNN) feature extraction and attention mechanism. 1D CNN is used to extract the temporal features from the streaming sensor data. Temporal features have monotonic degradation trends from the fluctuating raw sensor streaming data. Attention mechanism can improve the RUL prediction performance by capturing the fault characteristics and the degradation development with the attention weights. The performance of the TDDN model is evaluated on the public C-MAPSS dataset and compared with the existing methods. The results show that the TDDN model can achieve the best RUL prediction accuracy in complex conditions compared to current machine learning models. The degradation-related features extracted from the high-dimension sensor streaming data demonstrate the clear degradation trajectories and degradation stages that enable TDDN to predict the turbofan-engine RUL accurately and efficiently.

Yadong Zhang, Chenye Zou, Xin Chen

Battery discharge capacity forecasting is critically essential for the applications of lithium-ion batteries. The capacity degeneration can be treated as the memory of the initial battery state of charge from the data point of view. The streaming sensor data collected by battery management systems (BMS) reflect the usable battery capacity degradation rates under various operational working conditions. The battery capacity in different cycles can be measured with the temporal patterns extracted from the streaming sensor data based on the attention mechanism. The attention-based similarity regarding the first cycle can describe the battery capacity degradation in the following cycles. The deep degradation network (DDN) is developed with the attention mechanism to measure similarity and predict battery capacity. The DDN model can extract the degeneration-related temporal patterns from the streaming sensor data and perform the battery capacity prediction efficiently online in real-time. Based on the MIT-Stanford open-access battery aging dataset, the root-mean-square error of the capacity estimation is 1.3 mAh. The mean absolute percentage error of the proposed DDN model is 0.06{\%}. The DDN model also performance well in the Oxford Battery Degradation Dataset with dynamic load profiles. Therefore, the high accuracy and strong robustness of the proposed algorithm are verified.

Yadong Zhang, Xin Chen

In the time-series analysis, the time series motifs and the order patterns in time series can reveal general temporal patterns and dynamic features. Triadic Motif Field (TMF) is a simple and effective time-series image encoding method based on triadic time series motifs. Electrocardiography (ECG) signals are time-series data widely used to diagnose various cardiac anomalies. The TMF images contain the features characterizing the normal and Atrial Fibrillation (AF) ECG signals. Considering the quasi-periodic characteristics of ECG signals, the dynamic features can be extracted from the TMF images with the transfer learning pre-trained convolutional neural network (CNN) models. With the extracted features, the simple classifiers, such as the Multi-Layer Perceptron (MLP), the logistic regression, and the random forest, can be applied for accurate anomaly detection. With the test dataset of the PhysioNet Challenge 2017 database, the TMF classification model with the VGG16 transfer learning model and MLP classifier demonstrates the best performance with the 95.50% ROC-AUC and 88.43% F1 score in the AF classification. Besides, the TMF classification model can identify AF patients in the test dataset with high precision. The feature vectors extracted from the TMF images show clear patient-wise clustering with the t-distributed Stochastic Neighbor Embedding technique. Above all, the TMF classification model has very good clinical interpretability. The patterns revealed by symmetrized Gradient-weighted Class Activation Mapping have a clear clinical interpretation at the beat and rhythm levels.

Yadong Zhang, Xin Chen

Inspired by the well-known permutation entropy (PE), an effective image encoding scheme for chaotic time series, Triad State Space Construction (TSSC), is proposed. The TSSC image can recognize higher-order temporal patterns and identify new forbidden regions in time series motifs beyond the Bandt-Pompe probabilities. The Convolutional Neural Network (ConvNet) is widely used in image classification. The ConvNet classifier based on TSSC images (TSSC-ConvNet) are highly accurate and very robust in the chaotic signal classification.

Yadong Zhang, Xin Chen

Time series motifs play an important role in the time series analysis. The motif-based time series clustering is used for the discovery of higher-order patterns or structures in time series data. Inspired by the convolutional neural network (CNN) classifier based on the image representations of time series, motif difference field (MDF) is proposed. Compared to other image representations of time series, MDF is simple and easy to construct. With the Fully Convolution Network (FCN) as the classifier, MDF demonstrates the superior performance on the UCR time series dataset in benchmark with other time series classification methods. It is interesting to find that the triadic time series motifs give the best result in the test. Due to the motif clustering reflected in MDF, the significant motifs are detected with the help of the Gradient-weighted Class Activation Mapping (Grad-CAM). The areas in MDF with high weight in Grad-CAM have a high contribution from the significant motifs with the desired ordinal patterns associated with the signature patterns in time series. However, the signature patterns cannot be identified with the neural network classifiers directly based on the time series.