Shaojie Wang

Shaojie Wang, Liang Zhang

Current plan-based reasoning methods improve large language models (LLMs) by inserting a planning stage before execution, giving rise to the question $\rightarrow$ plan $\rightarrow$ cot paradigm. While effective, a closer examination reveals an inherent paradigm-level gap: both the planning and its execution stages decide how to solve a problem, while the prior question of what to solve; recognizing the problem type, the applicable tools, and the foreseeable pitfalls; remains entirely implicit. To bridge this gap, we propose PPC (Preplan-Plan-CoT), a framework that introduces an explicit problem-understanding stage, the preplan, yielding a new question $\rightarrow$ preplan $\rightarrow$ plan $\rightarrow$ cot paradigm. Realizing this paradigm requires safeguarding the conceptual integrity of preplan at both ends. Specifically, we design a three-stage synthesis pipeline with a spoiler-score detector that filters out leakage and spoiler failures to build clean preplan supervision, and a composite GRPO reward enforces that the generated plan genuinely follows from the preplan. Experiments across four backbones and five mathematical reasoning benchmarks show that PPC achieves the best results on 39 of 40 metrics, improving maj@16 and pass@16 by +2.23 and +3.06 over the strongest baseline without introducing additional inference token overhead.

Shaojie Wang, Jinghui Wang, Yinghan Cui et al.

In agentic LLM scenarios, an agent's interaction process during a single rollout often exhibits branching behaviors. Due to memory retrieval and concurrent tool executions at certain decision points, the token trajectory of one task evolves into a tree-like structure rather than a linear sequence. However, current training pipelines decompose such tree-structured trajectories into separate linear segments, treating each branch as an independent sequence. As a result, shared prefixes across these branches are repeatedly recomputed during both forward and backward passes. To address this inefficiency, we propose Tree Training, a paradigm that computes each shared prefix only once and reuses its intermediate results across related branches during both forward and backward passes, substantially improving computation efficiency in large-scale agentic training. This is achieved via (i) Tree Packing, which efficiently reuses shared computations across trajectories, and (ii) Gradient Restoration, which ensures correct gradient propagation across reused prefixes. Experiments on multiple open-source models demonstrate up to 3.9x reduction in total training time, enabling more efficient agentic LLM SFT and RL training.

Zizheng Zhan, Ken Deng, Huaixi Tang et al.

We present Kwaipilot-AutoThink (KAT), an open-source 40B large language model developed to address the overthinking problem in reasoning-intensive tasks, where an automatic thinking training paradigm is proposed to dynamically switch between reasoning and non-reasoning modes based on task complexity. Specifically, first, we construct the dual-regime dataset based on a novel tagging pipeline and a multi-agent synthesis strategy, and then we apply Multi-Token Prediction (MTP)-enhanced knowledge distillation, enabling efficient and fine-grained reasoning transfer with minimal pretraining cost. Besides, we implement a cold-start initialization strategy that introduces mode-selection priors using majority-vote signals and intent-aware prompting. Finally, we propose Step-SRPO, a reinforcement learning algorithm that incorporates intermediate supervision into the GRPO framework, offering structured guidance over both reasoning-mode selection and response accuracy. Extensive experiments across multiple benchmarks demonstrate that KAT consistently matches or even outperforms current state-of-the-art models, including DeepSeek-R1-0528 and Qwen3-235B-A22B, across a wide range of reasoning-intensive tasks while reducing token usage. Notably, KAT outperforms all open-source models and even surpasses o3-mini on the leakage-controlled LiveCodeBench Pro. Beyond academic evaluation, KAT has been successfully deployed in Kwaipilot (i.e., Kuaishou's internal coding assistant), where it improves real-world development workflows with high accuracy, efficiency, and controllable reasoning behaviors. Moreover, we are actively training a 200B Mixture-of-Experts (MoE) model with 40B active parameters, and early results already show significant gains, further demonstrating the scalability of the AutoThink paradigm.

Zhenhua Huang, Kunhao Li, Shaojie Wang et al.

Despite the Graph Neural Networks' (GNNs) proficiency in analyzing graph data, achieving high-accuracy and interpretable predictions remains challenging. Existing GNN interpreters typically provide post-hoc explanations disjointed from GNNs' predictions, resulting in misrepresentations. Self-explainable GNNs offer built-in explanations during the training process. However, they cannot exploit the explanatory outcomes to augment prediction performance, and they fail to provide high-quality explanations of node features and require additional processes to generate explainable subgraphs, which is costly. To address the aforementioned limitations, we propose a self-explained and self-supervised graph neural network (SES) to bridge the gap between explainability and prediction. SES comprises two processes: explainable training and enhanced predictive learning. During explainable training, SES employs a global mask generator co-trained with a graph encoder and directly produces crucial structure and feature masks, reducing time consumption and providing node feature and subgraph explanations. In the enhanced predictive learning phase, mask-based positive-negative pairs are constructed utilizing the explanations to compute a triplet loss and enhance the node representations by contrastive learning.

Zizheng Zhan, Ken Deng, Jinghui Wang et al.

Recent advances in large language models (LLMs) have enabled progress in agentic coding, where models autonomously reason, plan, and act within interactive software development workflows. However, bridging the gap between static text-based training and dynamic real-world agentic execution remains a core challenge. In this technical report, we present KAT-Coder, a large-scale agentic code model trained through a multi-stage curriculum encompassing Mid-Term Training, Supervised Fine-Tuning (SFT), Reinforcement Fine-Tuning (RFT), and Reinforcement-to-Deployment Adaptation. The Mid-Term stage enhances reasoning, planning, and reflection capabilities through a corpus of real software engineering data and synthetic agentic interactions. The SFT stage constructs a million-sample dataset balancing twenty programming languages, ten development contexts, and ten task archetypes. The RFT stage introduces a novel multi-ground-truth reward formulation for stable and sample-efficient policy optimization. Finally, the Reinforcement-to-Deployment phase adapts the model to production-grade IDE environments using Error-Masked SFT and Tree-Structured Trajectory Training. In summary, these stages enable KAT-Coder to achieve robust tool-use reliability, instruction alignment, and long-context reasoning, forming a deployable foundation for real-world intelligent coding agents. Our KAT series 32B model, KAT-Dev, has been open-sourced on https://huggingface.co/Kwaipilot/KAT-Dev.

Tianyou Chen, Jin Xiao, Xiaoguang Hu et al.

Camouflaged objects are typically assimilated into their backgrounds and exhibit fuzzy boundaries. The complex environmental conditions and the high intrinsic similarity between camouflaged targets and their surroundings pose significant challenges in accurately locating and segmenting these objects in their entirety. While existing methods have demonstrated remarkable performance in various real-world scenarios, they still face limitations when confronted with difficult cases, such as small targets, thin structures, and indistinct boundaries. Drawing inspiration from human visual perception when observing images containing camouflaged objects, we propose a three-stage model that enables coarse-to-fine segmentation in a single iteration. Specifically, our model employs three decoders to sequentially process subsampled features, cropped features, and high-resolution original features. This proposed approach not only reduces computational overhead but also mitigates interference caused by background noise. Furthermore, considering the significance of multi-scale information, we have designed a multi-scale feature enhancement module that enlarges the receptive field while preserving detailed structural cues. Additionally, a boundary enhancement module has been developed to enhance performance by leveraging boundary information. Subsequently, a mask-guided fusion module is proposed to generate fine-grained results by integrating coarse prediction maps with high-resolution feature maps. Our network surpasses state-of-the-art CNN-based counterparts without unnecessary complexities. Upon acceptance of the paper, the source code will be made publicly available at https://github.com/clelouch/BTSNet.

Zhenhua Huang, Kunhao Li, Shaojie Wang et al.

Graph neural networks (GNNs) are widely applied in graph data modeling. However, existing GNNs are often trained in a task-driven manner that fails to fully capture the intrinsic nature of the graph structure, resulting in sub-optimal node and graph representations. To address this limitation, we propose a novel Graph structure Prompt Learning method (GPL) to enhance the training of GNNs, which is inspired by prompt mechanisms in natural language processing. GPL employs task-independent graph structure losses to encourage GNNs to learn intrinsic graph characteristics while simultaneously solving downstream tasks, producing higher-quality node and graph representations. In extensive experiments on eleven real-world datasets, after being trained by GPL, GNNs significantly outperform their original performance on node classification, graph classification, and edge prediction tasks (up to 10.28%, 16.5%, and 24.15%, respectively). By allowing GNNs to capture the inherent structural prompts of graphs in GPL, they can alleviate the issue of over-smooth and achieve new state-of-the-art performances, which introduces a novel and effective direction for GNN research with potential applications in various domains.

Xiaojiang Zhang, Jinghui Wang, Zifei Cheng et al.

Recent advances of reasoning models, exemplified by OpenAI's o1 and DeepSeek's R1, highlight the significant potential of Reinforcement Learning (RL) to enhance the reasoning capabilities of Large Language Models (LLMs). However, replicating these advancements across diverse domains remains challenging due to limited methodological transparency. In this work, we present two-Staged history-Resampling Policy Optimization (SRPO), which surpasses the performance of DeepSeek-R1-Zero-32B on the AIME24 and LiveCodeBench benchmarks. SRPO achieves this using the same base model as DeepSeek (i.e. Qwen2.5-32B), using only about 1/10 of the training steps required by DeepSeek-R1-Zero-32B, demonstrating superior efficiency. Building upon Group Relative Policy Optimization (GRPO), we introduce two key methodological innovations: (1) a two-stage cross-domain training paradigm designed to balance the development of mathematical reasoning and coding proficiency, and (2) History Resampling (HR), a technique to address ineffective samples. Our comprehensive experiments validate the effectiveness of our approach, offering valuable insights into scaling LLM reasoning capabilities across diverse tasks.

Shaojie Wang, Sirui Ding, Na Zou

Recent advances in large language models (LLMs) have revolutionized natural language processing, yet evaluating their intrinsic linguistic understanding remains challenging. Moving beyond specialized evaluation tasks, we propose an information-theoretic framework grounded in mutual information (MI) to achieve this. We formalize the understanding as MI between an input sentence and its latent representation (sentence-level MI), measuring how effectively input information is preserved in latent representation. Given that LLMs learn embeddings for individual tokens, we decompose sentence-level MI into token-level MI between tokens and sentence embeddings, establishing theoretical bounds connecting these measures. Based on this foundation, we theoretically derive a computable lower bound for token-level MI using Fano's inequality, which directly relates to token-level recoverability-the ability to predict original tokens from sentence embedding. We implement this recoverability task to comparatively measure MI across different LLMs, revealing that encoder-only models consistently maintain higher information fidelity than their decoder-only counterparts, with the latter exhibiting a distinctive late-layer "forgetting" pattern where mutual information is first enhanced and then discarded. Moreover, fine-tuning to maximize token-level recoverability consistently improves understanding ability of LLMs on tasks without task-specific supervision, demonstrating that mutual information can serve as a foundation for understanding and improving language model capabilities.

Shaojie Wang, Liang Zhang

Recent studies reveal that large language models (LLMs) exhibit limited logical reasoning abilities in mathematical problem-solving, instead often relying on pattern-matching and memorization. We systematically analyze this limitation, focusing on logical relationship understanding, which is a core capability underlying genuine logical reasoning, and reveal that errors related to this capability account for over 90\% of incorrect predictions, with Chain-of-Thought Supervised Fine-Tuning (CoT-SFT) failing to substantially reduce these errors. To address this bottleneck, we propose First-Step Logical Reasoning (FSLR), a lightweight training framework targeting logical relationship understanding. Our key insight is that the first planning step-identifying which variables to use and which operation to apply-encourages the model to derive logical relationships directly from the problem statement. By training models on this isolated step, FSLR provides explicit supervision for logical relationship understanding, unlike CoT-SFT which implicitly embeds such relationships within complete solution trajectories. Extensive experiments across multiple models and datasets demonstrate that FSLR consistently outperforms CoT-SFT under both in-distribution and out-of-distribution settings, with average improvements of 3.2\% and 4.6\%, respectively. Moreover, FSLR achieves 4-6x faster training and reduces training token consumption by over 80\%.

Shaojie Wang, Liang Zhang

Current LLM post-training methods optimize complete reasoning trajectories through Supervised Fine-Tuning (SFT) followed by outcome-based Reinforcement Learning (RL). While effective, a closer examination reveals a fundamental gap: this approach does not align with how humans actually solve problems. Human cognition naturally decomposes problem-solving into two distinct stages: first acquiring abstract strategies (i.e., meta-knowledge) that generalize across problems, then adapting them to specific instances. In contrast, by treating complete trajectories as basic units, current methods are inherently problem-centric, entangling abstract strategies with problem-specific execution. To address this misalignment, we propose a cognitively-inspired framework that explicitly mirrors the two-stage human cognitive process. Specifically, Chain-of-Meta-Thought (CoMT) focuses supervised learning on abstract reasoning patterns without specific executions, enabling acquisition of generalizable strategies. Confidence-Calibrated Reinforcement Learning (CCRL) then optimizes task adaptation via confidence-aware rewards on intermediate steps, preventing overconfident errors from cascading and improving execution reliability. Experiments across four models and eight benchmarks show 2.19\% and 4.63\% improvements in-distribution and out-of-distribution respectively over standard methods, while reducing training time by 65-70% and token consumption by 50%, demonstrating that aligning post-training with human cognitive principles yields not only superior generalization but also enhanced training efficiency.

Jinghui Wang, Shaojie Wang, Yinghan Cui et al.

We introduce SeamlessFlow, a server based reinforcement learning (RL) framework that addresses two core challenges in industrial scale RL: (1) decoupling RL training from the complex execution flow of agents; (2) maximizing GPU utilization with minimal idle time while preserving the stability and scalability required for large-scale deployments. First, SeamlessFlow introduces a data plane that decouples the RL trainer from diverse, complex agent implementations while sustaining high throughput. A central trajectory manager maintains complete interaction histories and supports partial rollout, allowing rollout to pause for weight updates and resume seamlessly, keeping agents unaware of service interruptions. Second, we propose a tag driven scheduling paradigm that abstracts hardware into capability tagged resources, unifying colocated and disaggregated architectures. Based on this, SeamlessFlow introduces a spatiotemporal multiplexing pipeline that dynamically reassigns idle training nodes to rollout in a train rollout separated setup, eliminating pipeline bubbles and fully exploiting heterogeneous cluster resources. By combining these innovations, SeamlessFlow delivers both stability and high performance, making it well suited for multi agent, long horizon, and other complex RL tasks.

Mingyang Xie, Manav Kulshrestha, Shaojie Wang et al.

Modern AI tools, such as generative adversarial networks, have transformed our ability to create and modify visual data with photorealistic results. However, one of the deleterious side-effects of these advances is the emergence of nefarious uses in manipulating information in visual data, such as through the use of deep fakes. We propose a novel architecture for preserving the provenance of semantic information in images to make them less susceptible to deep fake attacks. Our architecture includes semantic signing and verification steps. We apply this architecture to verifying two types of semantic information: individual identities (faces) and whether the photo was taken indoors or outdoors. Verification accounts for a collection of common image transformation, such as translation, scaling, cropping, and small rotations, and rejects adversarial transformations, such as adversarially perturbed or, in the case of face verification, swapped faces. Experiments demonstrate that in the case of provenance of faces in an image, our approach is robust to black-box adversarial transformations (which are rejected) as well as benign transformations (which are accepted), with few false negatives and false positives. Background verification, on the other hand, is susceptible to black-box adversarial examples, but becomes significantly more robust after adversarial training.

Shaojie Wang, Tong Wu, Ayan Chakrabarti et al.

We experimentally study the robustness of deep camera-LiDAR fusion architectures for 2D object detection in autonomous driving. First, we find that the fusion model is usually both more accurate, and more robust against single-source attacks than single-sensor deep neural networks. Furthermore, we show that without adversarial training, early fusion is more robust than late fusion, whereas the two perform similarly after adversarial training. However, we note that single-channel adversarial training of deep fusion is often detrimental even to robustness. Moreover, we observe cross-channel externalities, where single-channel adversarial training reduces robustness to attacks on the other channel. Additionally, we observe that the choice of adversarial model in adversarial training is critical: using attacks restricted to cars' bounding boxes is more effective in adversarial training and exhibits less significant cross-channel externalities. Finally, we find that joint-channel adversarial training helps mitigate many of the issues above, but does not significantly boost adversarial robustness.

Ziyi Kou, Guofeng Cui, Shaojie Wang et al.

Weakly Supervised Object Localization (WSOL) methods generate both classification and localization results by learning from only image category labels. Previous methods usually utilize class activation map (CAM) to obtain target object regions. However, most of them only focus on improving foreground object parts in CAM, but ignore the important effect of its background contents. In this paper, we propose a confidence segmentation (ConfSeg) module that builds confidence score for each pixel in CAM without introducing additional hyper-parameters. The generated sample-specific confidence mask is able to indicate the extent of determination for each pixel in CAM, and further supervises additional CAM extended from internal feature maps. Besides, we introduce Co-supervised Augmentation (CoAug) module to capture feature-level representation for foreground and background parts in CAM separately. Then a metric loss is applied at batch sample level to augment distinguish ability of our model, which helps a lot to localize more related object parts. Our final model, CSoA, combines the two modules and achieves superior performance, e.g. $37.69\%$ and $48.81\%$ Top-1 localization error on CUB-200 and ILSVRC datasets, respectively, which outperforms all previous methods and becomes the new state-of-the-art.

Ziyi Kou, Wentian Zhao, Guofeng Cui et al.

Weakly Supervised Object Localization (WSOL) methodsusually rely on fully convolutional networks in order to ob-tain class activation maps(CAMs) of targeted labels. How-ever, these networks always highlight the most discriminativeparts to perform the task, the located areas are much smallerthan entire targeted objects. In this work, we propose a novelend-to-end model to enlarge CAMs generated from classifi-cation models, which can localize targeted objects more pre-cisely. In detail, we add an additional module in traditionalclassification networks to extract foreground object propos-als from images without classifying them into specific cate-gories. Then we set these normalized regions as unrestrictedpixel-level mask supervision for the following classificationtask. We collect a set of images defined as Background ImageSet from the Internet. The number of them is much smallerthan the targeted dataset but surprisingly well supports themethod to extract foreground regions from different pictures.The region extracted is independent from classification task,where the extracted region in each image covers almost en-tire object rather than just a significant part. Therefore, theseregions can serve as masks to supervise the response mapgenerated from classification models to become larger andmore precise. The method achieves state-of-the-art results onCUB-200-2011 in terms of Top-1 and Top-5 localization er-ror while has a competitive result on ILSVRC2016 comparedwith other approaches.

Shaojie Wang, Wentian Zhao, Ziyi Kou et al.

Understanding web instructional videos is an essential branch of video understanding in two aspects. First, most existing video methods focus on short-term actions for a-few-second-long video clips; these methods are not directly applicable to long videos. Second, unlike unconstrained long videos, e.g., movies, instructional videos are more structured in that they have step-by-step procedure constraining the understanding task. In this paper, we study reasoning on instructional videos via question-answering (QA). Surprisingly, it has not been an emphasis in the video community despite its rich applications. We thereby introduce YouQuek, an annotated QA dataset for instructional videos based on the recent YouCook2. The questions in YouQuek are not limited to cues on one frame but related to logical reasoning in the temporal dimension. Observing the lack of effective representations for modeling long videos, we propose a set of carefully designed models including a novel Recurrent Graph Convolutional Network (RGCN) that captures both temporal order and relation information. Furthermore, we study multiple modalities including description and transcripts for the purpose of boosting video understanding. Extensive experiments on YouQuek suggest that RGCN performs the best in terms of QA accuracy and a better performance is gained by introducing human annotated description.

Wentian Zhao, Shaojie Wang, Zhihuai Xie et al.

Expectation maximization (EM) algorithm is to find maximum likelihood solution for models having latent variables. A typical example is Gaussian Mixture Model (GMM) which requires Gaussian assumption, however, natural images are highly non-Gaussian so that GMM cannot be applied to perform clustering task on pixel space. To overcome such limitation, we propose a GAN based EM learning framework that can maximize the likelihood of images and estimate the latent variables with only the constraint of L-Lipschitz continuity. We call this model GAN-EM, which is a framework for image clustering, semi-supervised classification and dimensionality reduction. In M-step, we design a novel loss function for discriminator of GAN to perform maximum likelihood estimation (MLE) on data with soft class label assignments. Specifically, a conditional generator captures data distribution for $K$ classes, and a discriminator tells whether a sample is real or fake for each class. Since our model is unsupervised, the class label of real data is regarded as latent variable, which is estimated by an additional network (E-net) in E-step. The proposed GAN-EM achieves state-of-the-art clustering and semi-supervised classification results on MNIST, SVHN and CelebA, as well as comparable quality of generated images to other recently developed generative models.