Le Yu

Le Yu, Leilei Sun, Bowen Du et al.

We propose DyGFormer, a new Transformer-based architecture for dynamic graph learning. DyGFormer is conceptually simple and only needs to learn from nodes' historical first-hop interactions by: (1) a neighbor co-occurrence encoding scheme that explores the correlations of the source node and destination node based on their historical sequences; (2) a patching technique that divides each sequence into multiple patches and feeds them to Transformer, allowing the model to effectively and efficiently benefit from longer histories. We also introduce DyGLib, a unified library with standard training pipelines, extensible coding interfaces, and comprehensive evaluating protocols to promote reproducible, scalable, and credible dynamic graph learning research. By performing exhaustive experiments on thirteen datasets for dynamic link prediction and dynamic node classification tasks, we find that DyGFormer achieves state-of-the-art performance on most of the datasets, demonstrating its effectiveness in capturing nodes' correlations and long-term temporal dependencies. Moreover, some results of baselines are inconsistent with previous reports, which may be caused by their diverse but less rigorous implementations, showing the importance of DyGLib. All the used resources are publicly available at https://github.com/yule-BUAA/DyGLib.

Tao Zou, Le Yu, Yifei Huang et al.

In many real-world scenarios (e.g., academic networks, social platforms), different types of entities are not only associated with texts but also connected by various relationships, which can be abstracted as Text-Attributed Heterogeneous Graphs (TAHGs). Current pretraining tasks for Language Models (LMs) primarily focus on separately learning the textual information of each entity and overlook the crucial aspect of capturing topological connections among entities in TAHGs. In this paper, we present a new pretraining framework for LMs that explicitly considers the topological and heterogeneous information in TAHGs. Firstly, we define a context graph as neighborhoods of a target node within specific orders and propose a topology-aware pretraining task to predict nodes involved in the context graph by jointly optimizing an LM and an auxiliary heterogeneous graph neural network. Secondly, based on the observation that some nodes are text-rich while others have little text, we devise a text augmentation strategy to enrich textless nodes with their neighbors' texts for handling the imbalance issue. We conduct link prediction and node classification tasks on three datasets from various domains. Experimental results demonstrate the superiority of our approach over existing methods and the rationality of each design. Our code is available at https://github.com/Hope-Rita/THLM.

Haozhen Zhang, Le Yu, Xi Xiao et al.

Encrypted traffic classification is receiving widespread attention from researchers and industrial companies. However, the existing methods only extract flow-level features, failing to handle short flows because of unreliable statistical properties, or treat the header and payload equally, failing to mine the potential correlation between bytes. Therefore, in this paper, we propose a byte-level traffic graph construction approach based on point-wise mutual information (PMI), and a model named Temporal Fusion Encoder using Graph Neural Networks (TFE-GNN) for feature extraction. In particular, we design a dual embedding layer, a GNN-based traffic graph encoder as well as a cross-gated feature fusion mechanism, which can first embed the header and payload bytes separately and then fuses them together to obtain a stronger feature representation. The experimental results on two real datasets demonstrate that TFE-GNN outperforms multiple state-of-the-art methods in fine-grained encrypted traffic classification tasks.

Le Yu, Bowen Yu, Haiyang Yu et al.

In this paper, we unveil that Language Models (LMs) can acquire new capabilities by assimilating parameters from homologous models without retraining or GPUs. We first introduce DARE to set most delta parameters (i.e., the disparity between fine-tuned and pre-trained parameters) to zeros without affecting the abilities of Supervised Fine-Tuning (SFT) LMs, which randomly Drops delta parameters with a ratio $p$ And REscales the remaining ones by $1 / (1 - p)$ to approximate the original embeddings. Then, we use DARE as a versatile plug-in to sparsify delta parameters of multiple SFT homologous models for mitigating parameter interference and merge them into a single model by parameter fusing. We experiment with encoder- and decoder-based LMs, showing that: (1) SFT delta parameter value ranges are typically small (within 0.002) with extreme redundancy, and DARE can effortlessly eliminate 90% or even 99% of them; (2) DARE can merge multiple task-specific LMs into one LM with diverse capabilities. Notably, this phenomenon is more pronounced in large-scale LMs, where the merged LM reveals the potential to surpass the performance of any source LM, providing a new discovery. We also utilize DARE to create a merged LM that ranks first among models with 7 billion parameters on the Open LLM Leaderboard.

Le Yu

In this paper, we conduct an empirical evaluation of Temporal Graph Benchmark (TGB) by extending our Dynamic Graph Library (DyGLib) to TGB. Compared with TGB, we include eleven popular dynamic graph learning methods for more exhaustive comparisons. Through the experiments, we find that (1) different models depict varying performance across various datasets, which is in line with previous observations; (2) the performance of some baselines can be significantly improved over the reported results in TGB when using DyGLib. This work aims to ease the researchers' efforts in evaluating various dynamic graph learning methods on TGB and attempts to offer results that can be directly referenced in the follow-up research. All the used resources in this project are publicly available at https://github.com/yule-BUAA/DyGLib_TGB. This work is in progress, and feedback from the community is welcomed for improvements.

Kaifa Zhao, Le Yu, Shiyao Zhou et al.

Privacy protection raises great attention on both legal levels and user awareness. To protect user privacy, countries enact laws and regulations requiring software privacy policies to regulate their behavior. However, privacy policies are written in natural languages with many legal terms and software jargon that prevent users from understanding and even reading them. It is desirable to use NLP techniques to analyze privacy policies for helping users understand them. Furthermore, existing datasets ignore law requirements and are limited to English. In this paper, we construct the first Chinese privacy policy dataset, namely CA4P-483, to facilitate the sequence labeling tasks and regulation compliance identification between privacy policies and software. Our dataset includes 483 Chinese Android application privacy policies, over 11K sentences, and 52K fine-grained annotations. We evaluate families of robust and representative baseline models on our dataset. Based on baseline performance, we provide findings and potential research directions on our dataset. Finally, we investigate the potential applications of CA4P-483 combing regulation requirements and program analysis.

Le Yu, Leilei Sun, Bowen Du et al.

Graph Neural Networks (GNNs) have been widely applied in the semi-supervised node classification task, where a key point lies in how to sufficiently leverage the limited but valuable label information. Most of the classical GNNs solely use the known labels for computing the classification loss at the output. In recent years, several methods have been designed to additionally utilize the labels at the input. One part of the methods augment the node features via concatenating or adding them with the one-hot encodings of labels, while other methods optimize the graph structure by assuming neighboring nodes tend to have the same label. To bring into full play the rich information of labels, in this paper, we present a label-enhanced learning framework for GNNs, which first models each label as a virtual center for intra-class nodes and then jointly learns the representations of both nodes and labels. Our approach could not only smooth the representations of nodes belonging to the same class, but also explicitly encode the label semantics into the learning process of GNNs. Moreover, a training node selection technique is provided to eliminate the potential label leakage issue and guarantee the model generalization ability. Finally, an adaptive self-training strategy is proposed to iteratively enlarge the training set with more reliable pseudo labels and distinguish the importance of each pseudo-labeled node during the model training process. Experimental results on both real-world and synthetic datasets demonstrate our approach can not only consistently outperform the state-of-the-arts, but also effectively smooth the representations of intra-class nodes.

Le Yu, Zihang Liu, Leilei Sun et al.

Given a sequence of sets, where each set has a timestamp and contains an arbitrary number of elements, temporal sets prediction aims to predict the elements in the subsequent set. Previous studies for temporal sets prediction mainly focus on the modelling of elements and implicitly represent each user's preference based on his/her interacted elements. However, user preferences are often continuously evolving and the evolutionary trend cannot be fully captured with the indirect learning paradigm of user preferences. To this end, we propose a continuous-time user preference modelling framework for temporal sets prediction, which explicitly models the evolving preference of each user by maintaining a memory bank to store the states of all the users and elements. Specifically, we first construct a universal sequence by arranging all the user-set interactions in a non-descending temporal order, and then chronologically learn from each user-set interaction. For each interaction, we continuously update the memories of the related user and elements based on their currently encoded messages and past memories. Moreover, we present a personalized user behavior learning module to discover user-specific characteristics based on each user's historical sequence, which aggregates the previously interacted elements from dual perspectives according to the user and elements. Finally, we develop a set-batch algorithm to improve the model efficiency, which can create time-consistent batches in advance and achieve 3.5x and 3.0x speedups in the training and evaluation process on average. Experiments on four real-world datasets demonstrate the superiority of our approach over state-of-the-arts under both transductive and inductive settings. The good interpretability of our method is also shown.

Binghai Wang, Yantao Liu, Yuxuan Liu et al.

Generative Reward Models (GenRMs) and LLM-as-a-Judge exhibit deceptive alignment by producing correct judgments for incorrect reasons, as they are trained and evaluated to prioritize Outcome Accuracy, which undermines their ability to generalize during RLHF. We introduce Rationale Consistency, a fine-grained metric that quantifies the alignment between the model's reasoning process and human judgment. Our evaluation of frontier models reveals that rationale consistency effectively discriminates among state-of-the-art models and detects deceptive alignment, while outcome accuracy falls short in both respects. To mitigate this gap, we introduce a hybrid signal that combines rationale consistency with outcome accuracy for GenRM training. Our training method achieves state-of-the-art performance on RM-Bench (87.1%) and JudgeBench (82%), surpassing outcome-only baselines by an average of 5%. Using RM during RLHF, our method effectively improves performance as demonstrated on Arena Hard v2, notably yielding a 7% improvement in creative writing tasks. Further analysis confirms that our method escapes the deceptive alignment trap, effectively reversing the decline in rationale consistency observed in outcome-only training.

Le Yu, Bowen Yu, Haiyang Yu et al.

Merging Large Language Models (LLMs) aims to amalgamate multiple homologous LLMs into one with all the capabilities. Ideally, any LLMs sharing the same backbone should be mergeable, irrespective of whether they are Fine-Tuned (FT) with minor parameter changes or Pre-Trained (PT) with substantial parameter shifts. However, existing methods often manually assign the model importance, rendering them feasible only for LLMs with similar parameter alterations, such as multiple FT LLMs. The diverse parameter changed ranges between FT and PT LLMs pose challenges for current solutions in empirically determining the optimal combination. In this paper, we make a pioneering effort to broaden the applicability of merging techniques from FT to PT LLMs. We initially examine the efficacy of current methods in merging FT and PT LLMs, discovering that they struggle to deal with PT LLMs. Subsequently, we introduce an approach based on WeIght DisENtanglement (WIDEN) to effectively extend the merging scope, which first disentangles model weights into magnitude and direction components, and then performs adaptive fusion by considering their respective contributions. In the experiments, we merge Qwen1.5-Chat (an FT LLM with instruction-following skills) with Sailor (a PT LLM with multilingual abilities) across 7B and 14B model scales. Results reveal that: (1) existing solutions usually fail when merging Sailor, either losing both abilities or only retaining instruction-following skills; (2) WIDEN successfully injects the multilingual abilities of Sailor into Qwen1.5-Chat and make it proficient in Southeast Asian languages, achieving enhancements in the fundamental capabilities. In light of previous research, we also merge multiple 13B FT LLMs and observe that WIDEN achieves a balanced amalgamation of instruction following, mathematical reasoning, and code generation skills.

Juepeng Zheng, Shuai Yuan, Weijia Li et al.

Powered by the advances of optical remote sensing sensors, the production of very high spatial resolution multispectral images provides great potential for achieving cost-efficient and high-accuracy forest inventory and analysis in an automated way. Lots of studies that aim at providing an inventory to the level of each individual tree have generated a variety of methods for Individual Tree Crown Detection and Delineation (ITCD). This review covers ITCD methods for detecting and delineating individual tree crowns, and systematically reviews the past and present of ITCD-related researches applied to the optical remote sensing images. With the goal to provide a clear knowledge map of existing ITCD efforts, we conduct a comprehensive review of recent ITCD papers to build a meta-data analysis, including the algorithm, the study site, the tree species, the sensor type, the evaluation method, etc. We categorize the reviewed methods into three classes: (1) traditional image processing methods (such as local maximum filtering, image segmentation, etc.); (2) traditional machine learning methods (such as random forest, decision tree, etc.); and (3) deep learning based methods. With the deep learning-oriented approaches contributing a majority of the papers, we further discuss the deep learning-based methods as semantic segmentation and object detection methods. In addition, we discuss four ITCD-related issues to further comprehend the ITCD domain using optical remote sensing data, such as comparisons between multi-sensor based data and optical data in ITCD domain, comparisons among different algorithms and different ITCD tasks, etc. Finally, this review proposes some ITCD-related applications and a few exciting prospects and potential hot topics in future ITCD research.

Tao Zou, Le Yu, Junchen Ye et al.

Patent classification aims to assign multiple International Patent Classification (IPC) codes to a given patent. Recent methods for automatically classifying patents mainly focus on analyzing the text descriptions of patents. However, apart from the texts, each patent is also associated with some assignees, and the knowledge of their applied patents is often valuable for classification. Furthermore, the hierarchical taxonomy formulated by the IPC system provides important contextual information and enables models to leverage the correlations between IPC codes for more accurate classification. However, existing methods fail to incorporate the above aspects. In this paper, we propose an integrated framework that comprehensively considers the information on patents for patent classification. To be specific, we first present an IPC codes correlations learning module to derive their semantic representations via adaptively passing and aggregating messages within the same level and across different levels along the hierarchical taxonomy. Moreover, we design a historical application patterns learning component to incorporate the corresponding assignee's previous patents by a dual channel aggregation mechanism. Finally, we combine the contextual information of patent texts that contains the semantics of IPC codes, and assignees' sequential preferences to make predictions. Experiments on real-world datasets demonstrate the superiority of our approach over the existing methods. Besides, we present the model's ability to capture the temporal patterns of assignees and the semantic dependencies among IPC codes.

Tao Zou, Le Yu, Leilei Sun et al.

Accurate prediction of what types of patents that companies will apply for in the next period of time can figure out their development strategies and help them discover potential partners or competitors in advance. Although important, this problem has been rarely studied in previous research due to the challenges in modelling companies' continuously evolving preferences and capturing the semantic correlations of classification codes. To fill in this gap, we propose an event-based dynamic graph learning framework for patent application trend prediction. In particular, our method is founded on the memorable representations of both companies and patent classification codes. When a new patent is observed, the representations of the related companies and classification codes are updated according to the historical memories and the currently encoded messages. Moreover, a hierarchical message passing mechanism is provided to capture the semantic proximities of patent classification codes by updating their representations along the hierarchical taxonomy. Finally, the patent application trend is predicted by aggregating the representations of the target company and classification codes from static, dynamic, and hierarchical perspectives. Experiments on real-world data demonstrate the effectiveness of our approach under various experimental conditions, and also reveal the abilities of our method in learning semantics of classification codes and tracking technology developing trajectories of companies.

An Yang, Bowen Yu, Chengyuan Li et al.

We introduce Qwen2.5-1M, a series of models that extend the context length to 1 million tokens. Compared to the previous 128K version, the Qwen2.5-1M series have significantly enhanced long-context capabilities through long-context pre-training and post-training. Key techniques such as long data synthesis, progressive pre-training, and multi-stage supervised fine-tuning are employed to effectively enhance long-context performance while reducing training costs. To promote the use of long-context models among a broader user base, we present and open-source our inference framework. This framework includes a length extrapolation method that can expand the model context lengths by at least four times, or even more, without additional training. To reduce inference costs, we implement a sparse attention method along with chunked prefill optimization for deployment scenarios and a sparsity refinement method to improve precision. Additionally, we detail our optimizations in the inference engine, including kernel optimization, pipeline parallelism, and scheduling optimization, which significantly enhance overall inference performance. By leveraging our inference framework, the Qwen2.5-1M models achieve a remarkable 3x to 7x prefill speedup in scenarios with 1 million tokens of context. This framework provides an efficient and powerful solution for developing applications that require long-context processing using open-source models. The Qwen2.5-1M series currently includes the open-source models Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, as well as the API-accessed model Qwen2.5-Turbo. Evaluations show that Qwen2.5-1M models have been greatly improved in long-context tasks without compromising performance in short-context scenarios. Specifically, the Qwen2.5-14B-Instruct-1M model significantly outperforms GPT-4o-mini in long-context tasks and supports contexts eight times longer.

Jin Xu, Zhifang Guo, Hangrui Hu et al. · pku

We present Qwen3-Omni, a single multimodal model that, for the first time, maintains state-of-the-art performance across text, image, audio, and video without any degradation relative to single-modal counterparts. Qwen3-Omni matches the performance of same-sized single-modal models within the Qwen series and excels particularly on audio tasks. Across 36 audio and audio-visual benchmarks, Qwen3-Omni achieves open-source SOTA on 32 benchmarks and overall SOTA on 22, outperforming strong closed-source models such as Gemini-2.5-Pro, Seed-ASR, and GPT-4o-Transcribe. Qwen3-Omni adopts a Thinker-Talker MoE architecture that unifies perception and generation across text, images, audio, and video, yielding fluent text and natural real-time speech. It supports text interaction in 119 languages, speech understanding in 19 languages, and speech generation in 10 languages. To reduce first-packet latency in streaming synthesis, Talker autoregressively predicts discrete speech codecs using a multi-codebook scheme. Leveraging the representational capacity of these codebooks, we replace computationally intensive block-wise diffusion with a lightweight causal ConvNet, enabling streaming from the first codec frame. In cold-start settings, Qwen3-Omni achieves a theoretical end-to-end first-packet latency of 234 ms. To further strengthen multimodal reasoning, we introduce a Thinking model that explicitly reasons over inputs from any modality. Since the research community currently lacks a general-purpose audio captioning model, we fine-tuned Qwen3-Omni-30B-A3B to obtain Qwen3-Omni-30B-A3B-Captioner, which produces detailed, low-hallucination captions for arbitrary audio inputs. Qwen3-Omni-30B-A3B, Qwen3-Omni-30B-A3B-Thinking, and Qwen3-Omni-30B-A3B-Captioner are publicly released under the Apache 2.0 license.

Zihang Liu, Le Yu, Tongyu Zhu et al.

Spatial-temporal data modeling aims to mine the underlying spatial relationships and temporal dependencies of objects in a system. However, most existing methods focus on the modeling of spatial-temporal data in a single mode, lacking the understanding of multiple modes. Though very few methods have been presented to learn the multi-mode relationships recently, they are built on complicated components with higher model complexities. In this paper, we propose a simple framework for multi-mode spatial-temporal data modeling to bring both effectiveness and efficiency together. Specifically, we design a general cross-mode spatial relationships learning component to adaptively establish connections between multiple modes and propagate information along the learned connections. Moreover, we employ multi-layer perceptrons to capture the temporal dependencies and channel correlations, which are conceptually and technically succinct. Experiments on three real-world datasets show that our model can consistently outperform the baselines with lower space and time complexity, opening up a promising direction for modeling spatial-temporal data. The generalizability of the cross-mode spatial relationships learning module is also validated.

Zihan Qiu, Zekun Wang, Bo Zheng et al.

Gating mechanisms have been widely utilized, from early models like LSTMs and Highway Networks to recent state space models, linear attention, and also softmax attention. Yet, existing literature rarely examines the specific effects of gating. In this work, we conduct comprehensive experiments to systematically investigate gating-augmented softmax attention variants. Specifically, we perform a comprehensive comparison over 30 variants of 15B Mixture-of-Experts (MoE) models and 1.7B dense models trained on a 3.5 trillion token dataset. Our central finding is that a simple modification-applying a head-specific sigmoid gate after the Scaled Dot-Product Attention (SDPA)-consistently improves performance. This modification also enhances training stability, tolerates larger learning rates, and improves scaling properties. By comparing various gating positions and computational variants, we attribute this effectiveness to two key factors: (1) introducing non-linearity upon the low-rank mapping in the softmax attention, and (2) applying query-dependent sparse gating scores to modulate the SDPA output. Notably, we find this sparse gating mechanism mitigates 'attention sink' and enhances long-context extrapolation performance, and we also release related $\href{https://github.com/qiuzh20/gated_attention}{codes}$ and $\href{https://huggingface.co/QwQZh/gated_attention}{models}$ to facilitate future research.

An Yang, Anfeng Li, Baosong Yang et al. · tsinghua

In this work, we present Qwen3, the latest version of the Qwen model family. Qwen3 comprises a series of large language models (LLMs) designed to advance performance, efficiency, and multilingual capabilities. The Qwen3 series includes models of both dense and Mixture-of-Expert (MoE) architectures, with parameter scales ranging from 0.6 to 235 billion. A key innovation in Qwen3 is the integration of thinking mode (for complex, multi-step reasoning) and non-thinking mode (for rapid, context-driven responses) into a unified framework. This eliminates the need to switch between different models--such as chat-optimized models (e.g., GPT-4o) and dedicated reasoning models (e.g., QwQ-32B)--and enables dynamic mode switching based on user queries or chat templates. Meanwhile, Qwen3 introduces a thinking budget mechanism, allowing users to allocate computational resources adaptively during inference, thereby balancing latency and performance based on task complexity. Moreover, by leveraging the knowledge from the flagship models, we significantly reduce the computational resources required to build smaller-scale models, while ensuring their highly competitive performance. Empirical evaluations demonstrate that Qwen3 achieves state-of-the-art results across diverse benchmarks, including tasks in code generation, mathematical reasoning, agent tasks, etc., competitive against larger MoE models and proprietary models. Compared to its predecessor Qwen2.5, Qwen3 expands multilingual support from 29 to 119 languages and dialects, enhancing global accessibility through improved cross-lingual understanding and generation capabilities. To facilitate reproducibility and community-driven research and development, all Qwen3 models are publicly accessible under Apache 2.0.

Qwen, An Yang, Baosong Yang et al.

In this report, we introduce Qwen2.5, a comprehensive series of large language models (LLMs) designed to meet diverse needs. Compared to previous iterations, Qwen 2.5 has been significantly improved during both the pre-training and post-training stages. In terms of pre-training, we have scaled the high-quality pre-training datasets from the previous 7 trillion tokens to 18 trillion tokens. This provides a strong foundation for common sense, expert knowledge, and reasoning capabilities. In terms of post-training, we implement intricate supervised finetuning with over 1 million samples, as well as multistage reinforcement learning. Post-training techniques enhance human preference, and notably improve long text generation, structural data analysis, and instruction following. To handle diverse and varied use cases effectively, we present Qwen2.5 LLM series in rich sizes. Open-weight offerings include base and instruction-tuned models, with quantized versions available. In addition, for hosted solutions, the proprietary models currently include two mixture-of-experts (MoE) variants: Qwen2.5-Turbo and Qwen2.5-Plus, both available from Alibaba Cloud Model Studio. Qwen2.5 has demonstrated top-tier performance on a wide range of benchmarks evaluating language understanding, reasoning, mathematics, coding, human preference alignment, etc. Specifically, the open-weight flagship Qwen2.5-72B-Instruct outperforms a number of open and proprietary models and demonstrates competitive performance to the state-of-the-art open-weight model, Llama-3-405B-Instruct, which is around 5 times larger. Qwen2.5-Turbo and Qwen2.5-Plus offer superior cost-effectiveness while performing competitively against GPT-4o-mini and GPT-4o respectively. Additionally, as the foundation, Qwen2.5 models have been instrumental in training specialized models such as Qwen2.5-Math, Qwen2.5-Coder, QwQ, and multimodal models.

Haozhen Zhang, Xi Xiao, Le Yu et al.

As network security receives widespread attention, encrypted traffic classification has become the current research focus. However, existing methods conduct traffic classification without sufficiently considering the common characteristics between data samples, leading to suboptimal performance. Moreover, they train the packet-level and flow-level classification tasks independently, which is redundant because the packet representations learned in the packet-level task can be exploited by the flow-level task. Therefore, in this paper, we propose an effective model named a Contrastive Learning Enhanced Temporal Fusion Encoder (CLE-TFE). In particular, we utilize supervised contrastive learning to enhance the packet-level and flow-level representations and perform graph data augmentation on the byte-level traffic graph so that the fine-grained semantic-invariant characteristics between bytes can be captured through contrastive learning. We also propose cross-level multi-task learning, which simultaneously accomplishes the packet-level and flow-level classification tasks in the same model with one training. Further experiments show that CLE-TFE achieves the best overall performance on the two tasks, while its computational overhead (i.e., floating point operations, FLOPs) is only about 1/14 of the pre-trained model (e.g., ET-BERT). We release the code at https://github.com/ViktorAxelsen/CLE-TFE

Qiaoyu Tang, Hao Xiang, Le Yu et al.

While recent advances in reasoning models have demonstrated cognitive behaviors through reinforcement learning, existing approaches struggle to invoke deep reasoning capabilities in multi-turn agents with long-horizon interactions. We propose DeepMiner, a novel framework that elicits such abilities by introducing high-difficulty training tasks and dynamic context window. DeepMiner presents a reverse construction method to generate complex but verifiable question-answer pairs from authentic web sources, which ensures the challenge and reliability of training data while injecting cognitive capabilities into multi-turn reasoning scenarios. We further design an elegant yet effective dynamic context management strategy for both training and inference, utilizing sliding window mechanisms while eliminating the dependency on external summarization models, thereby efficiently empowering the model to handle continuously expanding long-horizon contexts. Through reinforcement learning on Qwen3-32B, we develop DeepMiner-32B, which achieves substantial performance improvements across multiple search agent benchmarks. DeepMiner attains 33.5% accuracy on BrowseComp-en, surpassing the previous best open-source agent by almost 20 percentage points, and demonstrates consistent improvements on BrowseComp-zh, XBench-DeepSearch, and GAIA. Notably, our dynamic context management enables sustained interactions of nearly 100 turns within standard 32k context length, effectively addressing the context limitations that constrain existing multi-turn interaction systems.

Shenzhi Wang, Le Yu, Chang Gao et al. · tsinghua

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs), while its mechanisms are not yet well understood. In this work, we undertake a pioneering exploration of RLVR through the novel perspective of token entropy patterns, comprehensively analyzing how different tokens influence reasoning performance. By examining token entropy patterns in Chain-of-Thought (CoT) reasoning, we observe that only a small fraction of tokens exhibit high entropy, and these tokens act as critical forks that steer the model toward diverse reasoning pathways. Furthermore, studying how entropy patterns evolve during RLVR training reveals that RLVR largely adheres to the base model's entropy patterns, primarily adjusting the entropy of high-entropy tokens. These findings highlight the significance of high-entropy tokens (i.e., forking tokens) to RLVR. We ultimately improve RLVR by restricting policy gradient updates to forking tokens and uncover a finding even beyond the 80/20 rule: utilizing only 20% of the tokens while maintaining performance comparable to full-gradient updates on the Qwen3-8B base model and significantly surpassing full-gradient updates on the Qwen3-32B (+11.04 on AIME'25 and +7.71 on AIME'24) and Qwen3-14B (+4.79 on AIME'25 and +5.21 on AIME'24) base models, highlighting a strong scaling trend. In contrast, training exclusively on the 80% lowest-entropy tokens leads to a marked decline in performance. These findings indicate that the efficacy of RLVR primarily arises from optimizing the high-entropy tokens that decide reasoning directions. Collectively, our results highlight the potential to understand RLVR through a token-entropy perspective and optimize RLVR by leveraging high-entropy minority tokens to further improve LLM reasoning.

Xiaoqi Li, Ting Chen, Xiapu Luo et al.

More than eight million smart contracts have been deployed into Ethereum, which is the most popular blockchain that supports smart contract. However, less than 1% of deployed smart contracts are open-source, and it is difficult for users to understand the functionality and internal mechanism of those closed-source contracts. Although a few decompilers for smart contracts have been recently proposed, it is still not easy for users to grasp the semantic information of the contract, not to mention the potential misleading due to decompilation errors. In this paper, we propose the first system named STAN to generate descriptions for the bytecodes of smart contracts to help users comprehend them. In particular, for each interface in a smart contract, STAN can generate four categories of descriptions, including functionality description, usage description, behavior description, and payment description, by leveraging symbolic execution and NLP (Natural Language Processing) techniques. Extensive experiments show that STAN can generate adequate, accurate, and readable descriptions for contract's bytecodes, which have practical value for users.

Sicong Cao, Jinxuan Xu, Le Yu et al.

Accurate vulnerability-inducing commit identification serves as a foundation for a series of software security tasks, such as vulnerability detection and affected version analysis. A straightforward solution is the SZZ algorithm, which traces back through the code history to identify the earliest commit that modify the vulnerable code. Unfortunately, neither the customized V-SZZ nor state-of-the-art LLM4SZZ perform satisfactorily due to the incorrect anchor selection and inadequate backtracking capability, making them far beyond a reliable usage in practice. To overcome these challenges, we propose a multi-agentic SZZ algorithm, named MAS-SZZ, that facilitates the identification of vulnerability-inducing commits through collaboration among agents. Specifically, given a CVE description and its corresponding fixing commit, MAS-SZZ summarizes the root cause of the vulnerability and employs a structured step-forward prompting strategy to localize vulnerability-related statements based on the change intent of each patch hunk. These vulnerable statements serve as anchors from which MAS-SZZ autonomously traces backward through the repository's history to find the commit that first introduced the vulnerability. Extensive experiments show that MAS-SZZ outperforms the state-of-the-art baselines across datasets and programming languages, achieving F1-score gains of up to 65.22% over the best-performing SZZ algorithm.

Binghai Wang, Runji Lin, Keming Lu et al. · tsinghua

Motivated by scaling laws in language modeling that demonstrate how test loss scales as a power law with model and dataset sizes, we find that similar laws exist in preference modeling. We propose World Preference Modeling$ (WorldPM) to emphasize this scaling potential, where World Preference embodies a unified representation of human preferences. In this paper, we collect preference data from public forums covering diverse user communities, and conduct extensive training using 15M-scale data across models ranging from 1.5B to 72B parameters. We observe distinct patterns across different evaluation metrics: (1) Adversarial metrics (ability to identify deceptive features) consistently scale up with increased training data and base model size; (2) Objective metrics (objective knowledge with well-defined answers) show emergent behavior in larger language models, highlighting WorldPM's scalability potential; (3) Subjective metrics (subjective preferences from a limited number of humans or AI) do not demonstrate scaling trends. Further experiments validate the effectiveness of WorldPM as a foundation for preference fine-tuning. Through evaluations on 7 benchmarks with 20 subtasks, we find that WorldPM broadly improves the generalization performance across human preference datasets of varying sizes (7K, 100K and 800K samples), with performance gains exceeding 5% on many key subtasks. Integrating WorldPM into our internal RLHF pipeline, we observe significant improvements on both in-house and public evaluation sets, with notable gains of 4% to 8% in our in-house evaluations.

Haozhen Zhang, Haodong Yue, Xi Xiao et al.

With the growing significance of network security, the classification of encrypted traffic has emerged as an urgent challenge. Traditional byte-based traffic analysis methods are constrained by the rigid granularity of information and fail to fully exploit the diverse correlations between bytes. To address these limitations, this paper introduces MH-Net, a novel approach for classifying network traffic that leverages multi-view heterogeneous traffic graphs to model the intricate relationships between traffic bytes. The essence of MH-Net lies in aggregating varying numbers of traffic bits into multiple types of traffic units, thereby constructing multi-view traffic graphs with diverse information granularities. By accounting for different types of byte correlations, such as header-payload relationships, MH-Net further endows the traffic graph with heterogeneity, significantly enhancing model performance. Notably, we employ contrastive learning in a multi-task manner to strengthen the robustness of the learned traffic unit representations. Experiments conducted on the ISCX and CIC-IoT datasets for both the packet-level and flow-level traffic classification tasks demonstrate that MH-Net achieves the best overall performance compared to dozens of SOTA methods.

Qiaoyu Tang, Hao Xiang, Le Yu et al.

With the rapid advancement of Large Language Models (LLMs), developing effective critic modules for precise guidance has become crucial yet challenging. In this paper, we initially demonstrate that supervised fine-tuning for building critic modules (which is widely adopted in current solutions) fails to genuinely enhance models' critique abilities, producing superficial critiques with insufficient reflections and verifications. To unlock the unprecedented critique capabilities, we propose RefCritic, a long-chain-of-thought critic module based on reinforcement learning with dual rule-based rewards: (1) instance-level correctness of solution judgments and (2) refinement accuracies of the policy model based on critiques, aiming to generate high-quality evaluations with actionable feedback that effectively guides model refinement. We evaluate RefCritic on Qwen2.5-14B-Instruct and DeepSeek-R1-Distill-Qwen-14B across five benchmarks. On critique and refinement settings, RefCritic demonstrates consistent advantages across all benchmarks, e.g., 6.8\% and 7.2\% gains on AIME25 for the respective base models. Notably, under majority voting, policy models filtered by RefCritic show superior scaling with increased voting numbers. Moreover, despite training on solution-level supervision, RefCritic outperforms step-level supervised approaches on ProcessBench, a benchmark to identify erroneous steps in mathematical reasoning.

Qiaoyu Tang, Le Yu, Bowen Yu et al.

Post-training has emerged as a crucial paradigm for adapting large-scale pre-trained models to various tasks, whose effects are fully reflected by delta parameters (i.e., the disparity between post-trained and pre-trained parameters). While numerous studies have explored delta parameter properties via operations like pruning, quantization, low-rank approximation, and extrapolation, a unified framework for systematically examining these characteristics has been lacking. In this paper, we propose a novel perspective based on Riemann sum approximation of the loss function to elucidate delta parameter editing operations. Our analysis categorizes existing methods into three classes based on their post-editing performance: competitive, decreased, and improved, explaining how they are expressed by the Riemann sum approximation term and how they alter the model performance. Extensive experiments on both visual and language models, including ViT, LLaMA 3, Qwen 2, and Mistral, corroborate our theoretical findings. Furthermore, we introduce extensions to existing techniques like DARE and BitDelta, highlighting their limitations in leveraging the properties of delta parameters and reorganizing them into general expressions to enhance the applicability and effectiveness of delta parameter editing in post-trained models.

Liping Han, Tingting Nie, Le Yu et al.

Flight control software is typically designed with numerous configurable parameters governing multiple functionalities, enabling flexible adaptation to mission diversity and environmental uncertainty. Although developers and manufacturers usually provide recommendations for these parameters to ensure safe and stable operations, certain combinations of parameters with recommended values may still lead to unstable flight behaviors, thereby degrading the drone's robustness. To this end, we propose a Large Language Model (LLM) based approach for real-time repair of risk-prone configurations (named RisConFix) that degrade drone robustness. RisConFix continuously monitors the drone's operational state and automatically triggers a repair mechanism once abnormal flight behaviors are detected. The repair mechanism leverages an LLM to analyze relationships between configuration parameters and flight states, and then generates corrective parameter updates to restore flight stability. To ensure the validity of the updated configuration, RisConFix operates as an iterative process; it continuously monitors the drone's flight state and, if an anomaly persists after applying an update, automatically triggers the next repair cycle. We evaluated RisConFix through a case study of ArduPilot (with 1,421 groups of misconfigurations). Experimental results show that RisConFix achieved a best repair success rate of 97% and an optimal average number of repairs of 1.17, demonstrating its capability to effectively and efficiently repair risk-prone configurations in real time.

Le Yu, Zhengyue Zhao, Yawen Zheng et al.

Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces. In this work, we find that the safety alignment of RVLMs can be easily break through a novel attack method termed \textbf{Stealth Fine-Tuning}. Our method elicits harmful reasoning traces through \textbf{segment-level interference} and reuses the self-generated outputs as supervised fine-tuning data. Through a \textbf{turn-based weighted} loss design, yielding a lightweight, distribution-consistent finetuning method. In our experiment, with only 499 samples and under 3 hours on a single A100 (QLoRA), Stealth Fine-Tuning outperforms IDEATOR by 38.52\% ASR while preserving general reasoning ability, as the tuned model retains the original representation distribution. Experiments on AdvBench and several general benchmarks demonstrate that Stealth Fine-Tuning is a low-cost and highly effective way to bypass alignment defenses. \textcolor{red}{\textbf{Disclaimer: This paper contains content that may be disturbing or offensive.}}

Le Yu, Kaishen Wang, Jianlong Xiong et al.

Though Large Vision-Language Models (LVLMs) have achieved remarkable performance across various tasks, they are still prone to hallucinations-generating outputs that are textually plausible but visually ungrounded. While prior approaches generally address this issue through data-centric fine-tuning or innovative decoding strategies, these methods often require substantial resources or task-specific configurations. In this work, we introduce an architecture-level solution, HalluRNN, which enhances model stability through recurrent cross-layer reasoning. Specifically, we propose a novel Dual-Gated Depth Propagation Unit (DG-DPU) module, which is shared across layers and recurrently refines hidden states. This allows for the adaptive propagation of information throughout the model, enforces consistency across layers, and mitigates hallucinations caused by representational drift. By fine-tuning only the DG-DPU module, HalluRNN achieves strong and robust performance across multiple benchmarks.

Le Yu, Leilei Sun, Bowen Du et al.

Representation learning on heterogeneous graphs aims to obtain meaningful node representations to facilitate various downstream tasks, such as node classification and link prediction. Existing heterogeneous graph learning methods are primarily developed by following the propagation mechanism of node representations. There are few efforts on studying the role of relations for improving the learning of more fine-grained node representations. Indeed, it is important to collaboratively learn the semantic representations of relations and discern node representations with respect to different relation types. To this end, in this paper, we propose a novel Relation-aware Heterogeneous Graph Neural Network, namely R-HGNN, to learn node representations on heterogeneous graphs at a fine-grained level by considering relation-aware characteristics. Specifically, a dedicated graph convolution component is first designed to learn unique node representations from each relation-specific graph separately. Then, a cross-relation message passing module is developed to improve the interactions of node representations across different relations. Also, the relation representations are learned in a layer-wise manner to capture relation semantics, which are used to guide the node representation learning process. Moreover, a semantic fusing module is presented to aggregate relation-aware node representations into a compact representation with the learned relation representations. Finally, we conduct extensive experiments on a variety of graph learning tasks, and experimental results demonstrate that our approach consistently outperforms existing methods among all the tasks.

Le Yu, Leilei Sun, Bowen Du et al.

Heterogeneous graphs are pervasive in practical scenarios, where each graph consists of multiple types of nodes and edges. Representation learning on heterogeneous graphs aims to obtain low-dimensional node representations that could preserve both node attributes and relation information. However, most of the existing graph convolution approaches were designed for homogeneous graphs, and therefore cannot handle heterogeneous graphs. Some recent methods designed for heterogeneous graphs are also faced with several issues, including the insufficient utilization of heterogeneous properties, structural information loss, and lack of interpretability. In this paper, we propose HGConv, a novel Heterogeneous Graph Convolution approach, to learn comprehensive node representations on heterogeneous graphs with a hybrid micro/macro level convolutional operation. Different from existing methods, HGConv could perform convolutions on the intrinsic structure of heterogeneous graphs directly at both micro and macro levels: A micro-level convolution to learn the importance of nodes within the same relation, and a macro-level convolution to distinguish the subtle difference across different relations. The hybrid strategy enables HGConv to fully leverage heterogeneous information with proper interpretability. Moreover, a weighted residual connection is designed to aggregate both inherent attributes and neighbor information of the focal node adaptively. Extensive experiments on various tasks demonstrate not only the superiority of HGConv over existing methods, but also the intuitive interpretability of our approach for graph analysis.

Juepeng Zheng, Haohuan Fu, Weijia Li et al.

Providing an accurate evaluation of palm tree plantation in a large region can bring meaningful impacts in both economic and ecological aspects. However, the enormous spatial scale and the variety of geological features across regions has made it a grand challenge with limited solutions based on manual human monitoring efforts. Although deep learning based algorithms have demonstrated potential in forming an automated approach in recent years, the labelling efforts needed for covering different features in different regions largely constrain its effectiveness in large-scale problems. In this paper, we propose a novel domain adaptive oil palm tree detection method, i.e., a Multi-level Attention Domain Adaptation Network (MADAN) to reap cross-regional oil palm tree counting and detection. MADAN consists of 4 procedures: First, we adopted a batch-instance normalization network (BIN) based feature extractor for improving the generalization ability of the model, integrating batch normalization and instance normalization. Second, we embedded a multi-level attention mechanism (MLA) into our architecture for enhancing the transferability, including a feature level attention and an entropy level attention. Then we designed a minimum entropy regularization (MER) to increase the confidence of the classifier predictions through assigning the entropy level attention value to the entropy penalty. Finally, we employed a sliding window-based prediction and an IOU based post-processing approach to attain the final detection results. We conducted comprehensive ablation experiments using three different satellite images of large-scale oil palm plantation area with six transfer tasks. MADAN improves the detection accuracy by 14.98% in terms of average F1-score compared with the Baseline method (without DA), and performs 3.55%-14.49% better than existing domain adaptation methods.

Ming Fan, Le Yu, Sen Chen et al.

The purpose of the General Data Protection Regulation (GDPR) is to provide improved privacy protection. If an app controls personal data from users, it needs to be compliant with GDPR. However, GDPR lists general rules rather than exact step-by-step guidelines about how to develop an app that fulfills the requirements. Therefore, there may exist GDPR compliance violations in existing apps, which would pose severe privacy threats to app users. In this paper, we take mobile health applications (mHealth apps) as a peephole to examine the status quo of GDPR compliance in Android apps. We first propose an automated system, named \mytool, to bridge the semantic gap between the general rules of GDPR and the app implementations by identifying the data practices declared in the app privacy policy and the data relevant behaviors in the app code. Then, based on \mytool, we detect three kinds of GDPR compliance violations, including the incompleteness of privacy policy, the inconsistency of data collections, and the insecurity of data transmission. We perform an empirical evaluation of 796 mHealth apps. The results reveal that 189 (23.7\%) of them do not provide complete privacy policies. Moreover, 59 apps collect sensitive data through different measures, but 46 (77.9\%) of them contain at least one inconsistent collection behavior. Even worse, among the 59 apps, only 8 apps try to ensure the transmission security of collected data. However, all of them contain at least one encryption or SSL misuse. Our work exposes severe privacy issues to raise awareness of privacy protection for app users and developers.

Le Yu, Leilei Sun, Bowen Du et al.

Given a sequence of sets, where each set contains an arbitrary number of elements, the problem of temporal sets prediction aims to predict the elements in the subsequent set. In practice, temporal sets prediction is much more complex than predictive modelling of temporal events and time series, and is still an open problem. Many possible existing methods, if adapted for the problem of temporal sets prediction, usually follow a two-step strategy by first projecting temporal sets into latent representations and then learning a predictive model with the latent representations. The two-step approach often leads to information loss and unsatisfactory prediction performance. In this paper, we propose an integrated solution based on the deep neural networks for temporal sets prediction. A unique perspective of our approach is to learn element relationship by constructing set-level co-occurrence graph and then perform graph convolutions on the dynamic relationship graphs. Moreover, we design an attention-based module to adaptively learn the temporal dependency of elements and sets. Finally, we provide a gated updating mechanism to find the hidden shared patterns in different sequences and fuse both static and dynamic information to improve the prediction performance. Experiments on real-world data sets demonstrate that our approach can achieve competitive performances even with a portion of the training data and can outperform existing methods with a significant margin.

Muyuan Li, Haojin Zhu, Zhaoyu Gao et al.

Many popular location-based social networks (LBSNs) support built-in location-based social discovery with hundreds of millions of users around the world. While user (near) realtime geographical information is essential to enable location-based social discovery in LBSNs, the importance of user location privacy has also been recognized by leading real-world LBSNs. To protect user's exact geographical location from being exposed, a number of location protection approaches have been adopted by the industry so that only relative location information are publicly disclosed. These techniques are assumed to be secure and are exercised on the daily base. In this paper, we question the safety of these location-obfuscation techniques used by existing LBSNs. We show, for the first time, through real world attacks that they can all be easily destroyed by an attacker with the capability of no more than a regular LBSN user. In particular, by manipulating location information fed to LBSN client app, an ill-intended regular user can easily deduce the exact location information by running LBSN apps as location oracle and performing a series of attacking strategies. We develop an automated user location tracking system and test it on the most popular LBSNs including Wechat, Skout and Momo. We demonstrate its effectiveness and efficiency via a 3 week real-world experiment with 30 volunteers. Our evaluation results show that we could geo-locate a target with high accuracy and can readily recover users' Top 5 locations. We also propose to use grid reference system and location classification to mitigate the attacks. Our work shows that the current industrial best practices on user location privacy protection are completely broken, and it is critical to address this immediate threat.