Liang Zeng

Lanqing Li, Liang Zeng, Ziqi Gao et al.

The last decade has witnessed a prosperous development of computational methods and dataset curation for AI-aided drug discovery (AIDD). However, real-world pharmaceutical datasets often exhibit highly imbalanced distribution, which is overlooked by the current literature but may severely compromise the fairness and generalization of machine learning applications. Motivated by this observation, we introduce ImDrug, a comprehensive benchmark with an open-source Python library which consists of 4 imbalance settings, 11 AI-ready datasets, 54 learning tasks and 16 baseline algorithms tailored for imbalanced learning. It provides an accessible and customizable testbed for problems and solutions spanning a broad spectrum of the drug discovery pipeline such as molecular modeling, drug-target interaction and retrosynthesis. We conduct extensive empirical studies with novel evaluation metrics, to demonstrate that the existing algorithms fall short of solving medicinal and pharmaceutical challenges in the data imbalance scenario. We believe that ImDrug opens up avenues for future research and development, on real-world challenges at the intersection of AIDD and deep imbalanced learning.

Liang Zeng, Lanqing Li, Ziqi Gao et al.

Graph contrastive learning (GCL) has attracted a surge of attention due to its superior performance for learning node/graph representations without labels. However, in practice, the underlying class distribution of unlabeled nodes for the given graph is usually imbalanced. This highly imbalanced class distribution inevitably deteriorates the quality of learned node representations in GCL. Indeed, we empirically find that most state-of-the-art GCL methods cannot obtain discriminative representations and exhibit poor performance on imbalanced node classification. Motivated by this observation, we propose a principled GCL framework on Imbalanced node classification (ImGCL), which automatically and adaptively balances the representations learned from GCL without labels. Specifically, we first introduce the online clustering based progressively balanced sampling (PBS) method with theoretical rationale, which balances the training sets based on pseudo-labels obtained from learned representations in GCL. We then develop the node centrality based PBS method to better preserve the intrinsic structure of graphs, by upweighting the important nodes of the given graph. Extensive experiments on multiple imbalanced graph datasets and imbalanced settings demonstrate the effectiveness of our proposed framework, which significantly improves the performance of the recent state-of-the-art GCL methods. Further experimental ablations and analyses show that the ImGCL framework consistently improves the representation quality of nodes in under-represented (tail) classes.

Liang Zeng, Attila Lengyel, Nergis Tömen et al.

In this work, we leverage estimated depth to boost self-supervised contrastive learning for segmentation of urban scenes, where unlabeled videos are readily available for training self-supervised depth estimation. We argue that the semantics of a coherent group of pixels in 3D space is self-contained and invariant to the contexts in which they appear. We group coherent, semantically related pixels into coherent depth regions given their estimated depth and use copy-paste to synthetically vary their contexts. In this way, cross-context correspondences are built in contrastive learning and a context-invariant representation is learned. For unsupervised semantic segmentation of urban scenes, our method surpasses the previous state-of-the-art baseline by +7.14% in mIoU on Cityscapes and +6.65% on KITTI. For fine-tuning on Cityscapes and KITTI segmentation, our method is competitive with existing models, yet, we do not need to pre-train on ImageNet or COCO, and we are also more computationally efficient. Our code is available on https://github.com/LeungTsang/CPCDR

Shilong Yang, Chulong Zhang, Qi Zang et al.

Breast cancer is a significant global health issue, and the diagnosis of breast imaging has always been challenging. Mammography images typically have extremely high resolution, with lesions occupying only a very small area. Down-sampling in neural networks can easily lead to the loss of microcalcifications or subtle structures, making it difficult for traditional neural network architectures to address these issues. To tackle these challenges, we propose a Context Clustering Network with triple information fusion. Firstly, compared to CNNs or transformers, we find that Context clustering methods (1) are more computationally efficient and (2) can more easily associate structural or pathological features, making them suitable for the clinical tasks of mammography. Secondly, we propose a triple information fusion mechanism that integrates global information, feature-based local information, and patch-based local information. The proposed approach is rigorously evaluated on two public datasets, Vindr-Mammo and CBIS-DDSM, using five independent splits to ensure statistical robustness. Our method achieves an AUC of 0.828 on Vindr-Mammo and 0.805 on CBIS-DDSM, outperforming the next best method by 3.1% and 2.4%, respectively. These improvements are statistically significant (p<0.05), underscoring the benefits of Context Clustering Network with triple information fusion. Overall, our Context Clustering framework demonstrates strong potential as a scalable and cost-effective solution for large-scale mammography screening, enabling more efficient and accurate breast cancer detection. Access to our method is available at https://github.com/Sohyu1/Mammo_Clustering.

Chris Yuhao Liu, Liang Zeng, Jiacai Liu et al.

In this report, we introduce a collection of methods to enhance reward modeling for LLMs, focusing specifically on data-centric techniques. We propose effective data selection and filtering strategies for curating high-quality open-source preference datasets, culminating in the Skywork-Reward data collection, which contains only 80K preference pairs -- significantly smaller than existing datasets. Using this curated dataset, we developed the Skywork-Reward model series -- Skywork-Reward-Gemma-27B and Skywork-Reward-Llama-3.1-8B -- with the former currently holding the top position on the RewardBench leaderboard. Notably, our techniques and datasets have directly enhanced the performance of many top-ranked models on RewardBench, highlighting the practical impact of our contributions in real-world preference learning applications.

Liang Zeng, Liangjun Zhong, Liang Zhao et al.

In this paper, we investigate the underlying factors that potentially enhance the mathematical reasoning capabilities of large language models (LLMs). We argue that the data scaling law for math reasoning capabilities in modern LLMs is far from being saturated, highlighting how the model's quality improves with increases in data quantity. To support this claim, we introduce the Skywork-Math model series, supervised fine-tuned (SFT) on common 7B LLMs using our proposed 2.5M-instance Skywork-MathQA dataset. Skywork-Math 7B has achieved impressive accuracies of 51.2% on the competition-level MATH benchmark and 83.9% on the GSM8K benchmark using only SFT data, outperforming an early version of GPT-4 on MATH. The superior performance of Skywork-Math models contributes to our novel two-stage data synthesis and model SFT pipelines, which include three different augmentation methods and a diverse seed problem set, ensuring both the quantity and quality of Skywork-MathQA dataset across varying difficulty levels. Most importantly, we provide several practical takeaways to enhance math reasoning abilities in LLMs for both research and industry applications.

Xiangkun Sun, Lingkai Kong, Aoqi Zhang et al.

Language models can be persuaded to abandon factual knowledge. This vulnerability is central to AI safety, but its internal mechanism remains poorly understood. We uncover a compact causal mechanism for persuasion-induced factual errors. A small set of mid-layer attention heads almost entirely determines the model's answer. These heads write answer options into a low-dimensional polyhedron, with options occupying distinct vertices. Persuasion does not blur belief or merely reduce confidence; it causes a discrete latent jump from the correct-answer vertex to the persuasion-target vertex. We show that decision heads are not reasoning over evidence. Instead, they copy whichever option token their attention selects. Persuasion works by redirecting attention. We isolate a rank-one evidence-routing feature that controls the route. Directly modifying this feature steers the model's choice, and removing it blocks persuasion. We then trace the feature back to a band of shallower attention heads that build it from persuasive keywords in the input. Every step is validated by intervention. This mechanism appears across open-source LLMs and realistic poisoning scenarios such as Generative Engine Optimization, revealing persuasion as a narrow, monitorable circuit.

Jujie He, Jiacai Liu, Chris Yuhao Liu et al.

The success of DeepSeek-R1 underscores the significant role of reinforcement learning (RL) in enhancing the reasoning capabilities of large language models (LLMs). In this work, we present Skywork-OR1, an effective and scalable RL implementation for long Chain-of-Thought (CoT) models. Building on the DeepSeek-R1-Distill model series, our RL approach achieves notable performance gains, increasing average accuracy across AIME24, AIME25, and LiveCodeBench from 57.8% to 72.8% (+15.0%) for the 32B model and from 43.6% to 57.5% (+13.9%) for the 7B model. Our Skywork-OR1-32B model surpasses both DeepSeek-R1 and Qwen3-32B on the AIME24 and AIME25 benchmarks, while achieving comparable results on LiveCodeBench. The Skywork-OR1-7B and Skywork-OR1-Math-7B models demonstrate competitive reasoning capabilities among models of similar size. We perform comprehensive ablation studies on the core components of our training pipeline to validate their effectiveness. Additionally, we thoroughly investigate the phenomenon of entropy collapse, identify key factors affecting entropy dynamics, and demonstrate that mitigating premature entropy collapse is critical for improved test performance. To support community research, we fully open-source our model weights, training code, and training datasets.

Guibin Chen, Dixuan Lin, Jiangping Yang et al.

SkyReels V4 is a unified multi modal video foundation model for joint video audio generation, inpainting, and editing. The model adopts a dual stream Multimodal Diffusion Transformer (MMDiT) architecture, where one branch synthesizes video and the other generates temporally aligned audio, while sharing a powerful text encoder based on the Multimodal Large Language Models (MMLM). SkyReels V4 accepts rich multi modal instructions, including text, images, video clips, masks, and audio references. By combining the MMLMs multi modal instruction following capability with in context learning in the video branch MMDiT, the model can inject fine grained visual guidance under complex conditioning, while the audio branch MMDiT simultaneously leverages audio references to guide sound generation. On the video side, we adopt a channel concatenation formulation that unifies a wide range of inpainting style tasks, such as image to video, video extension, and video editing under a single interface, and naturally extends to vision referenced inpainting and editing via multi modal prompts. SkyReels V4 supports up to 1080p resolution, 32 FPS, and 15 second duration, enabling high fidelity, multi shot, cinema level video generation with synchronized audio. To make such high resolution, long-duration generation computationally feasible, we introduce an efficiency strategy: Joint generation of low resolution full sequences and high-resolution keyframes, followed by dedicated super-resolution and frame interpolation models. To our knowledge, SkyReels V4 is the first video foundation model that simultaneously supports multi-modal input, joint video audio generation, and a unified treatment of generation, inpainting, and editing, while maintaining strong efficiency and quality at cinematic resolutions and durations.

Valerio Marsocci, Yuru Jia, Georges Le Bellier et al.

Geospatial Foundation Models (GFMs) have emerged as powerful tools for extracting representations from Earth observation data, but their evaluation remains inconsistent and narrow. Existing works often evaluate on suboptimal downstream datasets and tasks, that are often too easy or too narrow, limiting the usefulness of the evaluations to assess the real-world applicability of GFMs. Additionally, there is a distinct lack of diversity in current evaluation protocols, which fail to account for the multiplicity of image resolutions, sensor types, and temporalities, which further complicates the assessment of GFM performance. In particular, most existing benchmarks are geographically biased towards North America and Europe, questioning the global applicability of GFMs. To overcome these challenges, we introduce PANGAEA, a standardized evaluation protocol that covers a diverse set of datasets, tasks, resolutions, sensor modalities, and temporalities. It establishes a robust and widely applicable benchmark for GFMs. We evaluate the most popular GFMs openly available on this benchmark and analyze their performance across several domains. In particular, we compare these models to supervised baselines (e.g. UNet and vanilla ViT), and assess their effectiveness when faced with limited labeled data. Our findings highlight the limitations of GFMs, under different scenarios, showing that they do not consistently outperform supervised models. PANGAEA is designed to be highly extensible, allowing for the seamless inclusion of new datasets, models, and tasks in future research. By releasing the evaluation code and benchmark, we aim to enable other researchers to replicate our experiments and build upon our work, fostering a more principled evaluation protocol for large pre-trained geospatial models. The code is available at https://github.com/VMarsocci/pangaea-bench.

Liang Zeng, Valerio Marsocci, Wufan Zhao et al.

Masked Image Modeling has been one of the most popular self-supervised learning paradigms to learn representations from large-scale, unlabeled Earth Observation images. While incorporating multi-modal and multi-temporal Earth Observation data into Masked Image Modeling has been widely explored, the spatial dependencies between images captured from neighboring areas remains largely overlooked. Since the Earth's surface is continuous, neighboring images are highly related and offer rich contextual information for self-supervised learning. To close this gap, we propose NeighborMAE, which learns spatial dependencies by joint reconstruction of neighboring Earth Observation images. To ensure that the reconstruction remains challenging, we leverage a heuristic strategy to dynamically adjust the mask ratio and the pixel-level loss weight. Experimental results across various pretraining datasets and downstream tasks show that NeighborMAE significantly outperforms existing baselines, underscoring the value of neighboring images in Masked Image Modeling for Earth Observation and the efficacy of our designs.

Chaojie Wang, Yanchen Deng, Zhiyi Lyu et al.

Large Language Models (LLMs) have demonstrated impressive capability in many natural language tasks. However, the auto-regressive generation process makes LLMs prone to produce errors, hallucinations and inconsistent statements when performing multi-step reasoning. In this paper, by casting multi-step reasoning of LLMs as a heuristic search problem, we aim to alleviate the pathology by introducing Q*, a general, versatile and agile framework for guiding LLMs decoding process with deliberative planning. By learning a plug-and-play Q-value model as heuristic function for estimating expected future rewards, our Q* can effectively guide LLMs to select the most promising next reasoning step without fine-tuning LLMs for the current task, which avoids the significant computational overhead and potential risk of performance degeneration on other tasks. Extensive experiments on GSM8K, MATH and MBPP demonstrate the superiority of our method, contributing to improving the reasoning performance of existing open-source LLMs.

Tonghan Wang, Liang Zeng, Weijun Dong et al.

Learning sparse coordination graphs adaptive to the coordination dynamics among agents is a long-standing problem in cooperative multi-agent learning. This paper studies this problem and proposes a novel method using the variance of payoff functions to construct context-aware sparse coordination topologies. We theoretically consolidate our method by proving that the smaller the variance of payoff functions is, the less likely action selection will change after removing the corresponding edge. Moreover, we propose to learn action representations to effectively reduce the influence of payoff functions' estimation errors on graph construction. To empirically evaluate our method, we present the Multi-Agent COordination (MACO) benchmark by collecting classic coordination problems in the literature, increasing their difficulty, and classifying them into different types. We carry out a case study and experiments on the MACO and StarCraft II micromanagement benchmark to demonstrate the dynamics of sparse graph learning, the influence of graph sparseness, and the learning performance of our method. (The MACO benchmark and codes are publicly available at https://github.com/TonghanWang/CASEC-MACO-benchmark.)

Wentao Zhu, Yufang Huang, Liang Zeng et al.

Methods: Our deep learning model, called AnatomyNet, segments OARs from head and neck CT images in an end-to-end fashion, receiving whole-volume HaN CT images as input and generating masks of all OARs of interest in one shot. AnatomyNet is built upon the popular 3D U-net architecture, but extends it in three important ways: 1) a new encoding scheme to allow auto-segmentation on whole-volume CT images instead of local patches or subsets of slices, 2) incorporating 3D squeeze-and-excitation residual blocks in encoding layers for better feature representation, and 3) a new loss function combining Dice scores and focal loss to facilitate the training of the neural model. These features are designed to address two main challenges in deep-learning-based HaN segmentation: a) segmenting small anatomies (i.e., optic chiasm and optic nerves) occupying only a few slices, and b) training with inconsistent data annotations with missing ground truth for some anatomical structures. Results: We collected 261 HaN CT images to train AnatomyNet, and used MICCAI Head and Neck Auto Segmentation Challenge 2015 as a benchmark dataset to evaluate the performance of AnatomyNet. The objective is to segment nine anatomies: brain stem, chiasm, mandible, optic nerve left, optic nerve right, parotid gland left, parotid gland right, submandibular gland left, and submandibular gland right. Compared to previous state-of-the-art results from the MICCAI 2015 competition, AnatomyNet increases Dice similarity coefficient by 3.3% on average. AnatomyNet takes about 0.12 seconds to fully segment a head and neck CT image of dimension 178 x 302 x 225, significantly faster than previous methods. In addition, the model is able to process whole-volume CT images and delineate all OARs in one pass, requiring little pre- or post-processing. https://github.com/wentaozhu/AnatomyNet-for-anatomical-segmentation.git.

Peiran Yun, Wenxin Xu, Jiayuan Liu et al.

As Large Language Models (LLMs) transition into conversational agents, generative advertising emerges as a crucial monetization strategy. However, embedding advertisements within unstructured LLM outputs introduces a critical trilemma: balancing advertiser payoffs, platform revenue, and user experience. Existing methods, such as prompt injection or rigid position slots, disrupt semantic coherence and lack a parametric framework for independent control, rendering rigorous mechanism design intractable. To bridge this gap, we introduce Neuron Auctions, a novel paradigm that shifts the auction object from the surface text space to the LLM's internal representations. Leveraging mechanistic interpretability, we identify brand-specific feed-forward network (FFN) neurons and demonstrate that competing brands activate within approximately orthogonal subspaces. This near-perfect independence allows us to define continuous, disentangled intervention budgets (specifically, neuron counts and amplification factors) as auctionable commodities. Building on this computational carrier, we design a continuous menu-based auction mechanism that naturally guarantees strategy-proofness and optimizes revenue for the platform. By explicitly incorporating a user utility penalty into the platform's optimization objective, our framework dynamically prices out overly aggressive interventions. Extensive experiments demonstrate that Neuron Auctions effectively preserve natural discourse quality while achieving an optimal alignment between commercial incentives and user satisfaction.

Chris Yuhao Liu, Liang Zeng, Yuzhen Xiao et al.

Despite the critical role of reward models (RMs) in reinforcement learning from human feedback (RLHF), current state-of-the-art open RMs perform poorly on most existing evaluation benchmarks, failing to capture the spectrum of nuanced and sophisticated human preferences. Even approaches that incorporate advanced training techniques have not yielded meaningful performance improvements. We hypothesize that this brittleness stems primarily from limitations in preference datasets, which are often narrowly scoped, synthetically labeled, or lack rigorous quality control. To address these challenges, we present a large-scale preference dataset comprising 40 million preference pairs, named SynPref-40M. To enable data curation at scale, we design a human-AI synergistic two-stage pipeline that leverages the complementary strengths of human annotation quality and AI scalability. In this pipeline, humans provide verified annotations, while large language models perform automatic curation based on human guidance. Training on this preference mixture, we introduce Skywork-Reward-V2, a suite of eight reward models ranging from 0.6B to 8B parameters, trained on a carefully curated subset of 26 million preference pairs from SynPref-40M. We demonstrate that Skywork-Reward-V2 is versatile across a wide range of capabilities, including alignment with human preferences, objective correctness, safety, resistance to stylistic biases, and best-of-N scaling, achieving state-of-the-art performance across seven major reward model benchmarks. Ablation studies confirm that the effectiveness of our approach stems not only from data scale but also from high-quality curation. The Skywork-Reward-V2 series represents substantial progress in open reward models, highlighting the untapped potential of existing preference datasets and demonstrating how human-AI curation synergy can unlock significantly higher data quality.

Jiacai Liu, Chaojie Wang, Chris Yuhao Liu et al.

The role of reinforcement learning (RL) in enhancing the reasoning of large language models (LLMs) is becoming increasingly significant. Despite the success of RL in many scenarios, there are still many challenges in improving the reasoning of LLMs. One challenge is the sparse reward, which makes optimization difficult for RL and necessitates a large amount of data samples. Another challenge stems from the inherent instability of RL, particularly when using Actor-Critic (AC) methods to derive optimal policies, which often leads to unstable training processes. To address these issues, we introduce Direct Advantage Policy Optimization (DAPO), an novel step-level offline RL algorithm. Unlike standard alignment that rely solely outcome rewards to optimize policies (such as DPO), DAPO employs a critic function to predict the reasoning accuracy at each step, thereby generating dense signals to refine the generation strategy. Additionally, the Actor and Critic components in DAPO are trained independently, avoiding the co-training instability observed in standard AC algorithms like PPO. We train DAPO on mathematical and code query datasets and then evaluate its performance on multiple benchmarks. Our results show that DAPO can effectively enhance the mathematical and code capabilities on both SFT models and RL models, demonstrating the effectiveness of DAPO.

Baoyun Zhao, He Wang, Liang Zeng

While Large Language Models (LLMs) have recently shown promise in Automated Heuristic Design (AHD), existing approaches typically formulate AHD around constructive priority rules or parameterized local search guidance, thereby restricting the search space to fixed heuristic forms. Such designs offer limited capacity for structural exploration, making it difficult to escape deep local optima in complex Combinatorial Optimization Problems (COPs). In this work, we propose G-LNS, a generative evolutionary framework that extends LLM-based AHD to the automated design of Large Neighborhood Search (LNS) operators. Unlike prior methods that evolve heuristics in isolation, G-LNS leverages LLMs to co-evolve tightly coupled pairs of destroy and repair operators. A cooperative evaluation mechanism explicitly captures their interaction, enabling the discovery of complementary operator logic that jointly performs effective structural disruption and reconstruction. Extensive experiments on challenging COP benchmarks, such as Traveling Salesman Problems (TSP) and Capacitated Vehicle Routing Problems (CVRP), demonstrate that G-LNS significantly outperforms LLM-based AHD methods as well as strong classical solvers. The discovered heuristics not only achieve near-optimal solutions with reduced computational budgets but also exhibit robust generalization across diverse and unseen instance distributions.

Liang Zeng, Yongcong Li, Yuzhen Xiao et al.

Software engineering (SWE) has recently emerged as a crucial testbed for next-generation LLM agents, demanding inherent capabilities in two critical dimensions: sustained iterative problem-solving (e.g., >50 interaction rounds) and long-context dependency resolution (e.g., >32k tokens). However, the data curation process in SWE remains notoriously time-consuming, as it heavily relies on manual annotation for code file filtering and the setup of dedicated runtime environments to execute and validate unit tests. Consequently, most existing datasets are limited to only a few thousand GitHub-sourced instances. To this end, we propose an incremental, automated data-curation pipeline that systematically scales both the volume and diversity of SWE datasets. Our dataset comprises 10,169 real-world Python task instances from 2,531 distinct GitHub repositories, each accompanied by a task specified in natural language and a dedicated runtime-environment image for automated unit-test validation. We have carefully curated over 8,000 successfully runtime-validated training trajectories from our proposed SWE dataset. When fine-tuning the Skywork-SWE model on these trajectories, we uncover a striking data scaling phenomenon: the trained model's performance for software engineering capabilities in LLMs continues to improve as the data size increases, showing no signs of saturation. Notably, our Skywork-SWE model achieves 38.0% pass@1 accuracy on the SWE-bench Verified benchmark without using verifiers or multiple rollouts, establishing a new state-of-the-art (SOTA) among the Qwen2.5-Coder-32B-based LLMs built on the OpenHands agent framework. Furthermore, with the incorporation of test-time scaling techniques, the performance further improves to 47.0% accuracy, surpassing the previous SOTA results for sub-32B parameter models. We release the Skywork-SWE-32B model checkpoint to accelerate future research.

Xinyi Yang, Liang Zeng, Heng Dong et al.

As humans increasingly share environments with diverse agents powered by RL, LLMs, and beyond, the ability to explain their policies in natural language will be vital for reliable coexistence. In this paper, we build a model-agnostic explanation generator based on an LLM. The technical novelty is that the rewards for training this LLM are generated by a generative flow matching model. This model has a specially designed structure with a hidden layer merged with an LLM to harness the linguistic cues of explanations into generating appropriate rewards. Experiments on both RL and LLM tasks demonstrate that our method can generate dense and effective rewards while saving on expensive human feedback; it thus enables effective explanations and even improves the accuracy of the decisions in original tasks.

Wentao Zhang, Liang Zeng, Yuzhen Xiao et al.

Recent advances in LLMs-based agent systems have demonstrated remarkable capabilities in solving complex tasks. Nevertheless, current protocols (e.g., A2A and MCP) suffer from insufficient capabilities in context management, limited adaptability to diverse environments, and the absence of dynamic agent architectures. To address these limitations, we propose the Tool-Environment-Agent (TEA) Protocol, which establishes a principled basis for integrating environments, agents, and tools into an unified system. The TEA protocol treats environments and agents as first-class resources, enabling comprehensive context management and adaptive environment integration. Based on this protocol, we introduce AgentOrchestra, a hierarchical multi-agent framework with a central planning agent that decomposes complex objectives and coordinates specialized agents. Each sub-agent is dedicated to specific functions, providing capabilities for data analysis, file operations, web navigation, and interactive reasoning. Notably, AgentOrchestra introduces a tool manager agent that supports intelligent evolution through dynamic tool creation, retrieval, and reuse mechanisms. Experiments on three widely used benchmarks show that AgentOrchestra consistently outperforms existing baselines, achieving state-of-the-art performance of 83.39% on GAIA and ranking among the top general-purpose LLM-based agents. These results highlight the effectiveness of the TEA Protocol and hierarchical organization in building general-purpose multi-agent systems.

He Wang, Liang Zeng

Automated algorithm discovery in scientific computing faces fundamental challenges: vast design spaces with expensive evaluations, domain-specific physical constraints requiring expert knowledge, and the necessity for interpretable solutions that scientists can validate and understand. We present the Evo-MCTS (Evolutionary Monte Carlo Tree Search) framework, integrating large language models (LLMs) with tree-structured evolutionary search for interpretable algorithm discovery. Evo-MCTS combines reflective code synthesis leveraging LLM domain knowledge, multi-scale evolutionary operations on structured code representations, and interpretable algorithmic pathways emerging from tree-guided exploration. When applied to gravitational wave detection-a challenging domain with continuous parameter spaces and strict physical constraints-Evo-MCTS achieves 20.2% improvement over domain-specific methods and 59.1% over LLM-based optimization frameworks. This improvement arises from its ability to consistently converge toward interpretable algorithmic structures that integrate multiple functional components. Our domain-agnostic architecture establishes a generalizable methodology for automated algorithm discovery in scientific computing, where algorithmic transparency and physical validity are as essential as performance optimization.

Lu Zhang, Liang Zeng

The rapid proliferation of AI-generated images necessitates effective watermarking techniques to protect intellectual property and detect fraudulent content. While existing training-based watermarking methods show promise, they often struggle with generalizing across diverse prompts and tend to introduce visible artifacts. To this end, we propose a novel, provably generalizable image watermarking approach for Latent Diffusion Models, termed Self-Augmented Training (SAT-LDM). Our method aligns the training and testing phases through a free generation distribution, thereby enhancing the watermarking module's generalization capabilities. We theoretically consolidate SAT-LDM by proving that the free generation distribution contributes to its tight generalization bound, without the need for additional data collection. Extensive experiments show that SAT-LDM not only achieves robust watermarking but also significantly improves the quality of watermarked images across a wide range of prompts. Moreover, our experimental analyses confirm the strong generalization abilities of SAT-LDM. We hope that our method provides a practical and efficient solution for securing high-fidelity AI-generated content.

Tianwen Wei, Bo Zhu, Liang Zhao et al.

In this technical report, we introduce the training methodologies implemented in the development of Skywork-MoE, a high-performance mixture-of-experts (MoE) large language model (LLM) with 146 billion parameters and 16 experts. It is initialized from the pre-existing dense checkpoints of our Skywork-13B model. We explore the comparative effectiveness of upcycling versus training from scratch initializations. Our findings suggest that the choice between these two approaches should consider both the performance of the existing dense checkpoints and the MoE training budget. We highlight two innovative techniques: gating logit normalization, which improves expert diversification, and adaptive auxiliary loss coefficients, allowing for layer-specific adjustment of auxiliary loss coefficients. Our experimental results validate the effectiveness of these methods. Leveraging these techniques and insights, we trained our upcycled Skywork-MoE on a condensed subset of our SkyPile corpus. The evaluation results demonstrate that our model delivers strong performance across a wide range of benchmarks.

Liang Zhao, Tianwen Wei, Liang Zeng et al.

We introduce LongSkywork, a long-context Large Language Model (LLM) capable of processing up to 200,000 tokens. We provide a training recipe for efficiently extending context length of LLMs. We identify that the critical element in enhancing long-context processing capability is to incorporate a long-context SFT stage following the standard SFT stage. A mere 200 iterations can convert the standard SFT model into a long-context model. To reduce the effort in collecting and annotating data for long-context language modeling, we develop two novel methods for creating synthetic data. These methods are applied during the continual pretraining phase as well as the Supervised Fine-Tuning (SFT) phase, greatly enhancing the training efficiency of our long-context LLMs. Our findings suggest that synthetic long-context SFT data can surpass the performance of data curated by humans to some extent. LongSkywork achieves outstanding performance on a variety of long-context benchmarks. In the Needle test, a benchmark for long-context information retrieval, our models achieved perfect accuracy across multiple context spans. Moreover, in realistic application scenarios, LongSkywork-13B demonstrates performance on par with Claude2.1, the leading long-context model, underscoring the effectiveness of our proposed methods.

Bo Li, Yuyan Chen, Liang Zeng

Multi-Label Text Classification (MLTC) is a fundamental task in the field of Natural Language Processing (NLP) that involves the assignment of multiple labels to a given text. MLTC has gained significant importance and has been widely applied in various domains such as topic recognition, recommendation systems, sentiment analysis, and information retrieval. However, traditional machine learning and Deep neural network have not yet addressed certain issues, such as the fact that some documents are brief but have a large number of labels and how to establish relationships between the labels. It is imperative to additionally acknowledge that the significance of knowledge is substantiated in the realm of MLTC. To address this issue, we provide a novel approach known as Knowledge-enhanced Doc-Label Attention Network (KeNet). Specifically, we design an Attention Network that incorporates external knowledge, label embedding, and a comprehensive attention mechanism. In contrast to conventional methods, we use comprehensive representation of documents, knowledge and labels to predict all labels for each single text. Our approach has been validated by comprehensive research conducted on three multi-label datasets. Experimental results demonstrate that our method outperforms state-of-the-art MLTC method. Additionally, a case study is undertaken to illustrate the practical implementation of KeNet.

Liang Zeng, Lanqing Li, Jian Li

How to effectively represent molecules is a long-standing challenge for molecular property prediction and drug discovery. This paper studies this problem and proposes to incorporate chemical domain knowledge, specifically related to chemical reactions, for learning effective molecular representations. However, the inherent cross-modality property between chemical reactions and molecules presents a significant challenge to address. To this end, we introduce a novel method, namely MolKD, which Distills cross-modal Knowledge in chemical reactions to assist Molecular property prediction. Specifically, the reaction-to-molecule distillation model within MolKD transfers cross-modal knowledge from a pre-trained teacher network learning with one modality (i.e., reactions) into a student network learning with another modality (i.e., molecules). Moreover, MolKD learns effective molecular representations by incorporating reaction yields to measure transformation efficiency of the reactant-product pair when pre-training on reactions. Extensive experiments demonstrate that MolKD significantly outperforms various competitive baseline models, e.g., 2.1% absolute AUC-ROC gain on Tox21. Further investigations demonstrate that pre-trained molecular representations in MolKD can distinguish chemically reasonable molecular similarities, which enables molecular property prediction with high robustness and interpretability.

Yuchen Fang, Yanjun Qin, Haiyong Luo et al.

Traffic forecasting is crucial for public safety and resource optimization, yet is very challenging due to three aspects: i) current existing works mostly exploit intricate temporal patterns (e.g., the short-term thunderstorm and long-term daily trends) within a single method, which fail to accurately capture spatio-temporal dependencies under different schemas; ii) the under-exploration of the graph positional encoding limit the extraction of spatial information in the commonly used full graph attention network; iii) the quadratic complexity of the full graph attention introduces heavy computational needs. To achieve the effective traffic flow forecasting, we propose an efficient spectral graph attention network with disentangled traffic sequences. Specifically, the discrete wavelet transform is leveraged to obtain the low- and high-frequency components of traffic sequences, and a dual-channel encoder is elaborately designed to accurately capture the spatio-temporal dependencies under long- and short-term schemas of the low- and high-frequency components. Moreover, a novel wavelet-based graph positional encoding and a query sampling strategy are introduced in our spectral graph attention to effectively guide message passing and efficiently calculate the attention. Extensive experiments on four real-world datasets show the superiority of our model, i.e., the higher traffic forecasting precision with lower computational cost.

Yuchen Fang, Yanjun Qin, Haiyong Luo et al.

Traffic forecasting is important in intelligent transportation systems of webs and beneficial to traffic safety, yet is very challenging because of the complex and dynamic spatio-temporal dependencies in real-world traffic systems. Prior methods use the pre-defined or learnable static graph to extract spatial correlations. However, the static graph-based methods fail to mine the evolution of the traffic network. Researchers subsequently generate the dynamic graph for each time slice to reflect the changes of spatial correlations, but they follow the paradigm of independently modeling spatio-temporal dependencies, ignoring the cross-time spatial influence. In this paper, we propose a novel cross-time dynamic graph-based deep learning model, named CDGNet, for traffic forecasting. The model is able to effectively capture the cross-time spatial dependence between each time slice and its historical time slices by utilizing the cross-time dynamic graph. Meanwhile, we design a gating mechanism to sparse the cross-time dynamic graph, which conforms to the sparse spatial correlations in the real world. Besides, we propose a novel encoder-decoder architecture to incorporate the cross-time dynamic graph-based GCN for multi-step traffic forecasting. Experimental results on three real-world public traffic datasets demonstrate that CDGNet outperforms the state-of-the-art baselines. We additionally provide a qualitative study to analyze the effectiveness of our architecture.

Wei Shen, Chuheng Zhang, Yun Tian et al.

Recently, the graph neural network (GNN) has shown great power in matrix completion by formulating a rating matrix as a bipartite graph and then predicting the link between the corresponding user and item nodes. The majority of GNN-based matrix completion methods are based on Graph Autoencoder (GAE), which considers the one-hot index as input, maps a user (or item) index to a learnable embedding, applies a GNN to learn the node-specific representations based on these learnable embeddings and finally aggregates the representations of the target users and its corresponding item nodes to predict missing links. However, without node content (i.e., side information) for training, the user (or item) specific representation can not be learned in the inductive setting, that is, a model trained on one group of users (or items) cannot adapt to new users (or items). To this end, we propose an inductive matrix completion method using GAE (IMC-GAE), which utilizes the GAE to learn both the user-specific (or item-specific) representation for personalized recommendation and local graph patterns for inductive matrix completion. Specifically, we design two informative node features and employ a layer-wise node dropout scheme in GAE to learn local graph patterns which can be generalized to unseen data. The main contribution of our paper is the capability to efficiently learn local graph patterns in GAE, with good scalability and superior expressiveness compared to previous GNN-based matrix completion methods. Furthermore, extensive experiments demonstrate that our model achieves state-of-the-art performance on several matrix completion benchmarks. Our official code is publicly available.

Liang Zeng, Lei Wang, Hui Niu et al.

Price movement forecasting, aimed at predicting financial asset trends based on current market information, has achieved promising advancements through machine learning (ML) methods. Most existing ML methods, however, struggle with the extremely low signal-to-noise ratio and stochastic nature of financial data, often mistaking noises for real trading signals without careful selection of potentially profitable samples. To address this issue, we propose LARA, a novel price movement forecasting framework with two main components: Locality-Aware Attention (LA-Attention) and Iterative Refinement Labeling (RA-Labeling). (1) LA-Attention, enhanced by metric learning techniques, automatically extracts the potentially profitable samples through masked attention scheme and task-specific distance metrics. (2) RA-Labeling further iteratively refines the noisy labels of potentially profitable samples, and combines the learned predictors robust to the unseen and noisy samples. In a set of experiments on three real-world financial markets: stocks, cryptocurrencies, and ETFs, LARA significantly outperforms several machine learning based methods on the Qlib quantitative investment platform. Extensive ablation studies confirm LARA's superior ability in capturing more reliable trading opportunities.

Liang Zeng, Jin Xu, Zijun Yao et al.

Graph Neural Networks (GNNs) have already been widely used in various graph mining tasks. However, recent works reveal that the learned weights (channels) in well-trained GNNs are highly redundant, which inevitably limits the performance of GNNs. Instead of removing these redundant channels for efficiency consideration, we aim to reactivate them to enlarge the representation capacity of GNNs for effective graph learning. In this paper, we propose to substitute these redundant channels with other informative channels to achieve this goal. We introduce a novel GNN learning framework named AKE-GNN, which performs the Adaptive Knowledge Exchange strategy among multiple graph views generated by graph augmentations. AKE-GNN first trains multiple GNNs each corresponding to one graph view to obtain informative channels. Then, AKE-GNN iteratively exchanges redundant channels in the weight parameter matrix of one GNN with informative channels of another GNN in a layer-wise manner. Additionally, existing GNNs can be seamlessly incorporated into our framework. AKE-GNN achieves superior performance compared with various baselines across a suite of experiments on node classification, link prediction, and graph classification. In particular, we conduct a series of experiments on 15 public benchmark datasets, 8 popular GNN models, and 3 graph tasks and show that AKE-GNN consistently outperforms existing popular GNN models and even their ensembles. Extensive ablation studies and analyses on knowledge exchange methods validate the effectiveness of AKE-GNN.