Liang Zhao

Liang Zhao, Limin Wang

Detectors trained with massive labeled data often exhibit dramatic performance degradation in some particular scenarios with data distribution gap. To alleviate this problem of domain shift, conventional wisdom typically concentrates solely on reducing the discrepancy between the source and target domains via attached domain classifiers, yet ignoring the difficulty of such transferable features in coping with both classification and localization subtasks in object detection. To address this issue, in this paper, we propose Task-specific Inconsistency Alignment (TIA), by developing a new alignment mechanism in separate task spaces, improving the performance of the detector on both subtasks. Specifically, we add a set of auxiliary predictors for both classification and localization branches, and exploit their behavioral inconsistencies as finer-grained domain-specific measures. Then, we devise task-specific losses to align such cross-domain disagreement of both subtasks. By optimizing them individually, we are able to well approximate the category- and boundary-wise discrepancies in each task space, and therefore narrow them in a decoupled manner. TIA demonstrates superior results on various scenarios to the previous state-of-the-art methods. It is also observed that both the classification and localization capabilities of the detector are sufficiently strengthened, further demonstrating the effectiveness of our TIA method. Code and trained models are publicly available at https://github.com/MCG-NJU/TIA.

Xiaoxiao He, Chaowei Tan, Bo Liu et al.

Federated Learning has gained popularity among medical institutions since it enables collaborative training between clients (e.g., hospitals) without aggregating data. However, due to the high cost associated with creating annotations, especially for large 3D image datasets, clinical institutions do not have enough supervised data for training locally. Thus, the performance of the collaborative model is subpar under limited supervision. On the other hand, large institutions have the resources to compile data repositories with high-resolution images and labels. Therefore, individual clients can utilize the knowledge acquired in the public data repositories to mitigate the shortage of private annotated images. In this paper, we propose a federated few-shot learning method with dual knowledge distillation. This method allows joint training with limited annotations across clients without jeopardizing privacy. The supervised learning of the proposed method extracts features from limited labeled data in each client, while the unsupervised data is used to distill both feature and response-based knowledge from a national data repository to further improve the accuracy of the collaborative model and reduce the communication cost. Extensive evaluations are conducted on 3D magnetic resonance knee images from a private clinical dataset. Our proposed method shows superior performance and less training time than other semi-supervised federated learning methods. Codes and additional visualization results are available at https://github.com/hexiaoxiao-cs/fedml-knee.

Jianshen Zhu, Raveena Rai, Taiyo Sohkawa et al.

A novel two-phase molecule inference framework, mol-infer, has recently been developed to infer chemical graphs with prescribed abstract structures and desired property values through mixed integer linear programming (MILP) under the two-layered model, with guaranteed optimality and exactness relative to the given learned prediction function and structural constraints. In this study, we extend this framework to copolymers by introducing a simple feature representation, called the mixing vector (MV) model. In the proposed model, a copolymer feature vector is represented as a convex combination of MILP-tractable monomer descriptors weighted by the mixing ratio of the constituent monomers. This representation does not require explicit sequence-class information and is therefore naturally compatible with MILP-based inverse design. Under this model, we construct prediction functions for several copolymer property datasets using artificial neural networks, reduced quadratic multiple linear regression, and random forests. The proposed representation achieves practically useful predictive performance across multiple physicochemical property datasets; in particular, the best test R^2 score exceeds 0.7 for nine of the ten datasets and exceeds 0.9 for six datasets. We also formulate a multi-monomer inverse-design problem under the MV representation with a prescribed mixing ratio and show that the resulting MILP instances remain tractable, even for three-monomer settings. Finally, we perform an external consistency check by re-evaluating the inferred candidates and comparing the re-computed property values with those predicted by the learned model. Overall, the proposed framework gives a tractable first step toward model-level exact inverse design of copolymers under the two-layered model.

Runpei Dong, Chunrui Han, Yuang Peng et al. · tsinghua

This paper presents DreamLLM, a learning framework that first achieves versatile Multimodal Large Language Models (MLLMs) empowered with frequently overlooked synergy between multimodal comprehension and creation. DreamLLM operates on two fundamental principles. The first focuses on the generative modeling of both language and image posteriors by direct sampling in the raw multimodal space. This approach circumvents the limitations and information loss inherent to external feature extractors like CLIP, and a more thorough multimodal understanding is obtained. Second, DreamLLM fosters the generation of raw, interleaved documents, modeling both text and image contents, along with unstructured layouts. This allows DreamLLM to learn all conditional, marginal, and joint multimodal distributions effectively. As a result, DreamLLM is the first MLLM capable of generating free-form interleaved content. Comprehensive experiments highlight DreamLLM's superior performance as a zero-shot multimodal generalist, reaping from the enhanced learning synergy. Project page: https://dreamllm.github.io.

Tianwen Wei, Liang Zhao, Lichang Zhang et al.

In this technical report, we present Skywork-13B, a family of large language models (LLMs) trained on a corpus of over 3.2 trillion tokens drawn from both English and Chinese texts. This bilingual foundation model is the most extensively trained and openly published LLMs of comparable size to date. We introduce a two-stage training methodology using a segmented corpus, targeting general purpose training and then domain-specific enhancement training, respectively. We show that our model not only excels on popular benchmarks, but also achieves \emph{state of the art} performance in Chinese language modeling on diverse domains. Furthermore, we propose a novel leakage detection method, demonstrating that test data contamination is a pressing issue warranting further investigation by the LLM community. To spur future research, we release Skywork-13B along with checkpoints obtained during intermediate stages of the training process. We are also releasing part of our SkyPile corpus, a collection of over 150 billion tokens of web text, which is the largest high quality open Chinese pre-training corpus to date. We hope Skywork-13B and our open corpus will serve as a valuable open-source resource to democratize access to high-quality LLMs.

Shipeng Liu, Liang Zhao, Dengfeng Chen et al.

Recent crack segmentation methods often follow generic semantic segmentation designs, using stronger backbones, hybrid CNN-Transformer-Mamba encoders, and auxiliary enhancement branches. Although effective, this raises whether stronger generic feature mixing is the most suitable direction for crack segmentation. We instead formulate crack segmentation as sparse structural recovery. Cracks have limited category-level semantics but strong morphological regularities, being thin, sparse, anisotropic, locally fragmented, and easily confused with textures or shadows. Thus, the key bottleneck lies in preserving weak structural evidence, recovering directional continuity, and suppressing background coupling. We propose RIFT, a compact family of morphology-aligned crack segmentation models. Rather than compressing a complex generic architecture, RIFT is simple by design, preserving local evidence, aggregating cooperative directional continuity, and restoring crack structures through lightweight multi-scale fusion. Experiments on four public benchmarks show that RIFT achieves the best or tied-best results across the 16 main metrics against reproduced representative baselines. RIFT-B gives the strongest overall accuracy, while RIFT-T provides the best deployment efficiency with only 0.47M parameters and high inference speed. Topology-aware evaluation, ablations, transfer experiments, and visualizations further verify that task-aligned simplicity can match or surpass complex hybrid architectures when its inductive bias fits crack morphology. Code: https://github.com/xauat-liushipeng/RIFT

Haoran Wei, Chenglong Liu, Jinyue Chen et al. · tsinghua

Traditional OCR systems (OCR-1.0) are increasingly unable to meet people's usage due to the growing demand for intelligent processing of man-made optical characters. In this paper, we collectively refer to all artificial optical signals (e.g., plain texts, math/molecular formulas, tables, charts, sheet music, and even geometric shapes) as "characters" and propose the General OCR Theory along with an excellent model, namely GOT, to promote the arrival of OCR-2.0. The GOT, with 580M parameters, is a unified, elegant, and end-to-end model, consisting of a high-compression encoder and a long-contexts decoder. As an OCR-2.0 model, GOT can handle all the above "characters" under various OCR tasks. On the input side, the model supports commonly used scene- and document-style images in slice and whole-page styles. On the output side, GOT can generate plain or formatted results (markdown/tikz/smiles/kern) via an easy prompt. Besides, the model enjoys interactive OCR features, i.e., region-level recognition guided by coordinates or colors. Furthermore, we also adapt dynamic resolution and multi-page OCR technologies to GOT for better practicality. In experiments, we provide sufficient results to prove the superiority of our model.

Junxiang Wang, Junji Jiang, Liang Zhao

Localizing the source of graph diffusion phenomena, such as misinformation propagation, is an important yet extremely challenging task. Existing source localization models typically are heavily dependent on the hand-crafted rules. Unfortunately, a large portion of the graph diffusion process for many applications is still unknown to human beings so it is important to have expressive models for learning such underlying rules automatically. This paper aims to establish a generic framework of invertible graph diffusion models for source localization on graphs, namely Invertible Validity-aware Graph Diffusion (IVGD), to handle major challenges including 1) Difficulty to leverage knowledge in graph diffusion models for modeling their inverse processes in an end-to-end fashion, 2) Difficulty to ensure the validity of the inferred sources, and 3) Efficiency and scalability in source inference. Specifically, first, to inversely infer sources of graph diffusion, we propose a graph residual scenario to make existing graph diffusion models invertible with theoretical guarantees; second, we develop a novel error compensation mechanism that learns to offset the errors of the inferred sources. Finally, to ensure the validity of the inferred sources, a new set of validity-aware layers have been devised to project inferred sources to feasible regions by flexibly encoding constraints with unrolled optimization techniques. A linearization technique is proposed to strengthen the efficiency of our proposed layers. The convergence of the proposed IVGD is proven theoretically. Extensive experiments on nine real-world datasets demonstrate that our proposed IVGD outperforms state-of-the-art comparison methods significantly. We have released our code at https://github.com/xianggebenben/IVGD.

Yangkun Chen, Joseph Suarez, Junjie Zhang et al.

We present the results of the second Neural MMO challenge, hosted at IJCAI 2022, which received 1600+ submissions. This competition targets robustness and generalization in multi-agent systems: participants train teams of agents to complete a multi-task objective against opponents not seen during training. The competition combines relatively complex environment design with large numbers of agents in the environment. The top submissions demonstrate strong success on this task using mostly standard reinforcement learning (RL) methods combined with domain-specific engineering. We summarize the competition design and results and suggest that, as an academic community, competitions may be a powerful approach to solving hard problems and establishing a solid benchmark for algorithms. We will open-source our benchmark including the environment wrapper, baselines, a visualization tool, and selected policies for further research.

Liang Zhao, En Yu, Zheng Ge et al. · tsinghua

Human-AI interactivity is a critical aspect that reflects the usability of multimodal large language models (MLLMs). However, existing end-to-end MLLMs only allow users to interact with them through language instructions, leading to the limitation of the interactive accuracy and efficiency. In this study, we present precise referring instructions that utilize diverse reference representations such as points and boxes as referring prompts to refer to the special region. This enables MLLMs to focus on the region of interest and achieve finer-grained interaction. Based on precise referring instruction, we propose ChatSpot, a unified end-to-end multimodal large language model that supports diverse forms of interactivity including mouse clicks, drag-and-drop, and drawing boxes, which provides a more flexible and seamless interactive experience. We also construct a multi-grained vision-language instruction-following dataset based on existing datasets and GPT-4 generating. Furthermore, we design a series of evaluation tasks to assess the effectiveness of region recognition and interaction. Experimental results showcase ChatSpot's promising performance.

Zheng Zhang, Junxiang Wang, Liang Zhao

Graph Neural Networks (GNNs) have achieved great success in representing data with dependencies by recursively propagating and aggregating messages along the edges. However, edges in real-world graphs often have varying degrees of difficulty, and some edges may even be noisy to the downstream tasks. Therefore, existing GNNs may lead to suboptimal learned representations because they usually treat every edge in the graph equally. On the other hand, Curriculum Learning (CL), which mimics the human learning principle of learning data samples in a meaningful order, has been shown to be effective in improving the generalization ability and robustness of representation learners by gradually proceeding from easy to more difficult samples during training. Unfortunately, existing CL strategies are designed for independent data samples and cannot trivially generalize to handle data dependencies. To address these issues, we propose a novel CL strategy to gradually incorporate more edges into training according to their difficulty from easy to hard, where the degree of difficulty is measured by how well the edges are expected given the model training status. We demonstrate the strength of our proposed method in improving the generalization ability and robustness of learned representations through extensive experiments on nine synthetic datasets and nine real-world datasets. The code for our proposed method is available at https://github.com/rollingstonezz/Curriculum_learning_for_GNNs.

Hu Cao, Jingye Li, Fangfang Su et al.

Event extraction (EE) is an essential task of information extraction, which aims to extract structured event information from unstructured text. Most prior work focuses on extracting flat events while neglecting overlapped or nested ones. A few models for overlapped and nested EE includes several successive stages to extract event triggers and arguments,which suffer from error propagation. Therefore, we design a simple yet effective tagging scheme and model to formulate EE as word-word relation recognition, called OneEE. The relations between trigger or argument words are simultaneously recognized in one stage with parallel grid tagging, thus yielding a very fast event extraction speed. The model is equipped with an adaptive event fusion module to generate event-aware representations and a distance-aware predictor to integrate relative distance information for word-word relation recognition, which are empirically demonstrated to be effective mechanisms. Experiments on 3 overlapped and nested EE benchmarks, namely FewFC, Genia11, and Genia13, show that OneEE achieves the state-of-the-art (SOTA) results. Moreover, the inference speed of OneEE is faster than those of baselines in the same condition, and can be further substantially improved since it supports parallel inference.

Xiaoyuan Zhang, Liang Zhao, Yingying Yu et al.

Multiobjective optimization problems (MOPs) are prevalent in machine learning, with applications in multi-task learning, learning under fairness or robustness constraints, etc. Instead of reducing multiple objective functions into a scalar objective, MOPs aim to optimize for the so-called Pareto optimality or Pareto set learning, which involves optimizing more than one objective function simultaneously, over models with thousands / millions of parameters. Existing benchmark libraries for MOPs mainly focus on evolutionary algorithms, most of which are zeroth-order / meta-heuristic methods that do not effectively utilize higher-order information from objectives and cannot scale to large-scale models with thousands / millions of parameters. In light of the above gap, this paper introduces LibMOON, the first multiobjective optimization library that supports state-of-the-art gradient-based methods, provides a fair benchmark, and is open-sourced for the community.

Yiwen Shi, Ping Ren, Jing Wang et al.

Food effect summarization from New Drug Application (NDA) is an essential component of product-specific guidance (PSG) development and assessment. However, manual summarization of food effect from extensive drug application review documents is time-consuming, which arouses a need to develop automated methods. Recent advances in large language models (LLMs) such as ChatGPT and GPT-4, have demonstrated great potential in improving the effectiveness of automated text summarization, but its ability regarding the accuracy in summarizing food effect for PSG assessment remains unclear. In this study, we introduce a simple yet effective approach, iterative prompting, which allows one to interact with ChatGPT or GPT-4 more effectively and efficiently through multi-turn interaction. Specifically, we propose a three-turn iterative prompting approach to food effect summarization in which the keyword-focused and length-controlled prompts are respectively provided in consecutive turns to refine the quality of the generated summary. We conduct a series of extensive evaluations, ranging from automated metrics to FDA professionals and even evaluation by GPT-4, on 100 NDA review documents selected over the past five years. We observe that the summary quality is progressively improved throughout the process. Moreover, we find that GPT-4 performs better than ChatGPT, as evaluated by FDA professionals (43% vs. 12%) and GPT-4 (64% vs. 35%). Importantly, all the FDA professionals unanimously rated that 85% of the summaries generated by GPT-4 are factually consistent with the golden reference summary, a finding further supported by GPT-4 rating of 72% consistency. These results strongly suggest a great potential for GPT-4 to draft food effect summaries that could be reviewed by FDA professionals, thereby improving the efficiency of PSG assessment cycle and promoting the generic drug product development.

Ailin Huang, Chengyuan Yao, Chunrui Han et al.

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

Chen Ling, Xujiang Zhao, Xuchao Zhang et al.

Open Information Extraction (OIE) task aims at extracting structured facts from unstructured text, typically in the form of (subject, relation, object) triples. Despite the potential of large language models (LLMs) like ChatGPT as a general task solver, they lag behind state-of-the-art (supervised) methods in OIE tasks due to two key issues. First, LLMs struggle to distinguish irrelevant context from relevant relations and generate structured output due to the restrictions on fine-tuning the model. Second, LLMs generates responses autoregressively based on probability, which makes the predicted relations lack confidence. In this paper, we assess the capabilities of LLMs in improving the OIE task. Particularly, we propose various in-context learning strategies to enhance LLM's instruction-following ability and a demonstration uncertainty quantification module to enhance the confidence of the generated relations. Our experiments on three OIE benchmark datasets show that our approach holds its own against established supervised methods, both quantitatively and qualitatively.

Xiaomi LLM-Core Team, Dong Zhang, Gang Wang et al.

Existing audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversion, style transfer, and speech editing. MiMo-Audio-7B-Base also demonstrates powerful speech continuation capabilities, capable of generating highly realistic talk shows, recitations, livestreaming and debates. At the post-training stage, we curate a diverse instruction-tuning corpus and introduce thinking mechanisms into both audio understanding and generation. MiMo-Audio-7B-Instruct achieves open-source SOTA on audio understanding benchmarks (MMSU, MMAU, MMAR, MMAU-Pro), spoken dialogue benchmarks (Big Bench Audio, MultiChallenge Audio) and instruct-TTS evaluations, approaching or surpassing closed-source models. Model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-Audio.

Muniba Batool, Naveed Ahmed Azam, Jianshen Zhu et al.

Aqueous solubility (AS) is a key physiochemical property that plays a crucial role in drug discovery and material design. We report a novel unified approach to predict and infer chemical compounds with the desired AS based on simple deterministic graph-theoretic descriptors, multiple linear regression (MLR) and mixed integer linear programming (MILP). Selected descriptors based on a forward stepwise procedure enabled the simplest regression model, MLR, to achieve significantly good prediction accuracy compared to the existing approaches, achieving the accuracy in the range [0.7191, 0.9377] for 29 diverse datasets. By simulating these descriptors and learning models as MILPs, we inferred mathematically exact and optimal compounds with the desired AS, prescribed structures, and up to 50 non-hydrogen atoms in a reasonable time range [6, 1204] seconds. These findings indicate a strong correlation between the simple graph-theoretic descriptors and the AS of compounds, potentially leading to a deeper understanding of their AS without relying on widely used complicated chemical descriptors and complex machine learning models that are computationally expensive, and therefore difficult to use for inference. An implementation of the proposed approach is available at https://github.com/ku-dml/mol-infer/tree/master/AqSol.

Hejie Cui, Jiaying Lu, Ran Xu et al.

This comprehensive review aims to provide an overview of the current state of Healthcare Knowledge Graphs (HKGs), including their construction, utilization models, and applications across various healthcare and biomedical research domains. We thoroughly analyzed existing literature on HKGs, covering their construction methodologies, utilization techniques, and applications in basic science research, pharmaceutical research and development, clinical decision support, and public health. The review encompasses both model-free and model-based utilization approaches and the integration of HKGs with large language models (LLMs). We searched Google Scholar for relevant papers on HKGs and classified them into the following topics: HKG construction, HKG utilization, and their downstream applications in various domains. We also discussed their special challenges and the promise for future work. The review highlights the potential of HKGs to significantly impact biomedical research and clinical practice by integrating vast amounts of biomedical knowledge from multiple domains. The synergy between HKGs and LLMs offers promising opportunities for constructing more comprehensive knowledge graphs and improving the accuracy of healthcare applications. HKGs have emerged as a powerful tool for structuring medical knowledge, with broad applications across biomedical research, clinical decision-making, and public health. This survey serves as a roadmap for future research and development in the field of HKGs, highlighting the potential of combining knowledge graphs with advanced machine learning models for healthcare transformation.

Victor Y. Pan, Liang Zhao

Low-rank approximation of a matrix by means of random sampling has been consistently efficient in its empirical studies by many scientists who applied it with various sparse and structured multipliers, but adequate formal support for this empirical phenomenon has been missing so far. Our novel insight into the subject leads to such an elusive formal support and promises significant acceleration of the known algorithms for some fundamental problems of matrix computations and data mining and analysis. Our formal results and our numerical tests are in good accordance with each other. We also outline extensions of low-rank approximation algorithms and of our progress to the Least Squares Regression, the Fast Multipole Method, and the Conjugate Gradient algorithms.

Zheng Chai, Guangji Bai, Liang Zhao et al.

Despite the recent success of Graph Neural Networks, it remains challenging to train a GNN on large graphs with millions of nodes and billions of edges, which are prevalent in many graph-based applications. Traditional sampling-based methods accelerate GNN training by dropping edges and nodes, which impairs the graph integrity and model performance. Differently, distributed GNN algorithms accelerate GNN training by utilizing multiple computing devices and can be classified into two types: "partition-based" methods enjoy low communication costs but suffer from information loss due to dropped edges, while "propagation-based" methods avoid information loss but suffer from prohibitive communication overhead caused by the neighbor explosion. To jointly address these problems, this paper proposes DIGEST (DIstributed Graph reprEsentation SynchronizaTion), a novel distributed GNN training framework that synergizes the complementary strength of both categories of existing methods. We propose to allow each device to utilize the stale representations of its neighbors in other subgraphs during subgraph parallel training. This way, our method preserves global graph information from neighbors to avoid information loss and reduce communication costs. Our convergence analysis demonstrates that DIGEST enjoys a state-of-the-art convergence rate. Extensive experimental evaluation on large, real-world graph datasets shows that DIGEST achieves up to 21.82 speedups without compromising performance compared to state-of-the-art distributed GNN training frameworks.

Jiang Liu, Donghong Ji, Jingye Li et al.

So far, discontinuous named entity recognition (NER) has received increasing research attention and many related methods have surged such as hypergraph-based methods, span-based methods, and sequence-to-sequence (Seq2Seq) methods, etc. However, these methods more or less suffer from some problems such as decoding ambiguity and efficiency, which limit their performance. Recently, grid-tagging methods, which benefit from the flexible design of tagging systems and model architectures, have shown superiority to adapt for various information extraction tasks. In this paper, we follow the line of such methods and propose a competitive grid-tagging model for discontinuous NER. We call our model TOE because we incorporate two kinds of Tag-Oriented Enhancement mechanisms into a state-of-the-art (SOTA) grid-tagging model that casts the NER problem into word-word relationship prediction. First, we design a Tag Representation Embedding Module (TREM) to force our model to consider not only word-word relationships but also word-tag and tag-tag relationships. Concretely, we construct tag representations and embed them into TREM, so that TREM can treat tag and word representations as queries/keys/values and utilize self-attention to model their relationships. On the other hand, motivated by the Next-Neighboring-Word (NNW) and Tail-Head-Word (THW) tags in the SOTA model, we add two new symmetric tags, namely Previous-Neighboring-Word (PNW) and Head-Tail-Word (HTW), to model more fine-grained word-word relationships and alleviate error propagation from tag prediction. In the experiments of three benchmark datasets, namely CADEC, ShARe13 and ShARe14, our TOE model pushes the SOTA results by about 0.83%, 0.05% and 0.66% in F1, demonstrating its effectiveness.

Han Yi Chiu, Liang Zhao, Anqi Wu

The decoding of brain neural networks has been an intriguing topic in neuroscience for a well-rounded understanding of different types of brain disorders and cognitive stimuli. Integrating different types of connectivity, e.g., Functional Connectivity (FC) and Structural Connectivity (SC), from multi-modal imaging techniques can take their complementary information into account and therefore have the potential to get better decoding capability. However, traditional approaches for integrating FC and SC overlook the dynamical variations, which stand a great chance to over-generalize the brain neural network. In this paper, we propose a Joint kernel Graph Attention Network (JGAT), which is a new multi-modal temporal graph attention network framework. It integrates the data from functional Magnetic Resonance Images (fMRI) and Diffusion Weighted Imaging (DWI) while preserving the dynamic information at the same time. We conduct brain-decoding tasks with our JGAT on four independent datasets: three of 7T fMRI datasets from the Human Connectome Project (HCP) and one from animal neural recordings. Furthermore, with Attention Scores (AS) and Frame Scores (FS) computed and learned from the model, we can locate several informative temporal segments and build meaningful dynamical pathways along the temporal domain for the HCP datasets. The URL to the code of JGAT model: https://github.com/BRAINML-GT/JGAT.

En Yu, Liang Zhao, Yana Wei et al.

Humans possess the remarkable ability to foresee the future to a certain extent based on present observations, a skill we term as foresight minds. However, this capability remains largely under explored within existing Multimodal Large Language Models (MLLMs), hindering their capacity to learn the fundamental principles of how things operate and the intentions behind the observed subjects. To address this issue, we introduce the integration of future modeling into the existing learning frameworks of MLLMs. By utilizing the subject trajectory, a highly structured representation of a consecutive frame sequence, as a learning objective, we aim to bridge the gap between the past and the future. We propose two innovative methods to empower MLLMs with foresight minds, Foresight Pre-Training (FPT) and Foresight Instruction-Tuning (FIT), which are inspired by the modern learning paradigm of LLMs. Specifically, FPT jointly training various tasks centered on trajectories, enabling MLLMs to learn how to attend and predict entire trajectories from a given initial observation. Then, FIT requires MLLMs to first predict trajectories of related objects and then reason about potential future events based on them. Aided by FPT and FIT, we build a novel and unified MLLM named Merlin that supports multi-images input and analysis about potential actions of multiple objects for the future reasoning. Experimental results show Merlin powerful foresight minds with impressive performance on both future reasoning and visual comprehension tasks.

V. Y. Pan, F. Soleymani, Liang Zhao

We propose a hyperpower iteration for numerical computation of the outer generalized inverse of a matrix which achieves the 18th order of convergence by using only seven matrix multiplication per iteration loop. This is the record high efficiency for that computational task. The algorithm has a relatively mild numerical instability, and we stabilize it at the price of adding one extra matrix multiplication per iteration loop. This imlplies an efficiency index that significantly exceeds the known record for numerically stable iterations for this task. Our numerical tests cover a variety of examples such as Drazin case, rectangular case, and preconditioning of linear systems. The test results are in good accordance with our formal study and indicate that our algorithms can be of interest for the user.

Yuyang Gao, Siyi Gu, Junji Jiang et al.

As the societal impact of Deep Neural Networks (DNNs) grows, the goals for advancing DNNs become more complex and diverse, ranging from improving a conventional model accuracy metric to infusing advanced human virtues such as fairness, accountability, transparency (FaccT), and unbiasedness. Recently, techniques in Explainable Artificial Intelligence (XAI) are attracting considerable attention, and have tremendously helped Machine Learning (ML) engineers in understanding AI models. However, at the same time, we started to witness the emerging need beyond XAI among AI communities; based on the insights learned from XAI, how can we better empower ML engineers in steering their DNNs so that the model's reasonableness and performance can be improved as intended? This article provides a timely and extensive literature overview of the field Explanation-Guided Learning (EGL), a domain of techniques that steer the DNNs' reasoning process by adding regularization, supervision, or intervention on model explanations. In doing so, we first provide a formal definition of EGL and its general learning paradigm. Secondly, an overview of the key factors for EGL evaluation, as well as summarization and categorization of existing evaluation procedures and metrics for EGL are provided. Finally, the current and potential future application areas and directions of EGL are discussed, and an extensive experimental study is presented aiming at providing comprehensive comparative studies among existing EGL models in various popular application domains, such as Computer Vision (CV) and Natural Language Processing (NLP) domains.

Shiyu Wang, Xiaojie Guo, Xuanyang Lin et al.

Developing deep generative models has been an emerging field due to the ability to model and generate complex data for various purposes, such as image synthesis and molecular design. However, the advancement of deep generative models is limited by challenges to generate objects that possess multiple desired properties: 1) the existence of complex correlation among real-world properties is common but hard to identify; 2) controlling individual property enforces an implicit partially control of its correlated properties, which is difficult to model; 3) controlling multiple properties under various manners simultaneously is hard and under-explored. We address these challenges by proposing a novel deep generative framework that recovers semantics and the correlation of properties through disentangled latent vectors. The correlation is handled via an explainable mask pooling layer, and properties are precisely retained by generated objects via the mutual dependence between latent vectors and properties. Our generative model preserves properties of interest while handling correlation and conflicts of properties under a multi-objective optimization framework. The experiments demonstrate our model's superior performance in generating data with desired properties.

Victor Y. Pan, Qi Luan, John Svadlenka et al.

We study superfast algorithms that computes low rank approximation of a matrix (hereafter referred to as LRA) that use much fewer memory cells and arithmetic operations than the input matrix has entries. We first specify a family of 2mn matrices of size m*n such that for almost 50% of them any superfast LRA algorithm fails to improve the poor trivial approximation by the matrix filled with zeros, but then we prove that the class of all such hard inputs is narrow - the cross-approximation (hereafter {C-A}) superfast iterations as well as some more primitive superfast algorithms compute reasonably accurate LRAs in their transparent CUR form (i) to any matrix allowing close LRA except for small norm perturbations of matrices of an algebraic variety of a smaller dimension, (ii) to the average matrix allowing close LRA, (iii) to the average sparse matrix allowing close LRA and (iv) with a high probability to any matrix allowing close LRA if it is pre-processed fast with a random Gaussian, SRHT or SRFT multiplier. Moreover empirically the output LRAs remain accurate when we perform the computations superfast by replacing such a multiplier with one of our sparse and structured multipliers. Our techniques, auxiliary results and extensions may be of some independent interest. We analyze C-A and other superfast algorithms twice -- based on two well-known sufficient criteria for obtaining accurate LRAs. We provide a distinct proof in the case of superfast variant of randomized algorithms of [DMM08], improve a decade-old estimate for the norm of the inverse of a Gaussian matrix, prove such an estimate also in the case of a sparse Gaussian matrix, present some novel advanced pre-processing techniques for fast and superfast computation of LRA, and extend our results to dramatic acceleration of the Fast Multipole Method (FMM) and the Conjugate Gradient algorithms.

Jiang Liu, Hao Fei, Fei Li et al.

Few-shot named entity recognition (NER) exploits limited annotated instances to identify named mentions. Effectively transferring the internal or external resources thus becomes the key to few-shot NER. While the existing prompt tuning methods have shown remarkable few-shot performances, they still fail to make full use of knowledge. In this work, we investigate the integration of rich knowledge to prompt tuning for stronger few-shot NER. We propose incorporating the deep prompt tuning framework with threefold knowledge (namely TKDP), including the internal 1) context knowledge and the external 2) label knowledge & 3) sememe knowledge. TKDP encodes the three feature sources and incorporates them into the soft prompt embeddings, which are further injected into an existing pre-trained language model to facilitate predictions. On five benchmark datasets, our knowledge-enriched model boosts by at most 11.53% F1 over the raw deep prompt method, and significantly outperforms 8 strong-performing baseline systems in 5-/10-/20-shot settings, showing great potential in few-shot NER. Our TKDP can be broadly adapted to other few-shot tasks without effort.

Shiyu Wang, Yuanqi Du, Xiaojie Guo et al.

Designing and generating new data under targeted properties has been attracting various critical applications such as molecule design, image editing and speech synthesis. Traditional hand-crafted approaches heavily rely on expertise experience and intensive human efforts, yet still suffer from the insufficiency of scientific knowledge and low throughput to support effective and efficient data generation. Recently, the advancement of deep learning has created the opportunity for expressive methods to learn the underlying representation and properties of data. Such capability provides new ways of determining the mutual relationship between the structural patterns and functional properties of the data and leveraging such relationships to generate structural data, given the desired properties. This article is a systematic review that explains this promising research area, commonly known as controllable deep data generation. First, the article raises the potential challenges and provides preliminaries. Then the article formally defines controllable deep data generation, proposes a taxonomy on various techniques and summarizes the evaluation metrics in this specific domain. After that, the article introduces exciting applications of controllable deep data generation, experimentally analyzes and compares existing works. Finally, this article highlights the promising future directions of controllable deep data generation and identifies five potential challenges.

Victor Y. Pan, Liang Zhao

1. A standard Gaussian random matrix has full rank with probability 1 and is well-conditioned with a probability quite close to 1 and converging to 1 fast as the matrix deviates from square shape and becomes more rectangular. 2. If we append sufficiently many standard Gaussian random rows or columns to any normalized matrix A, then the augmented matrix has full rank with probability 1 and is well-conditioned with a probability close to 1, even if the matrix A is rank deficient or ill-conditioned. 3. We specify and prove these properties of augmentation and extend them to additive preprocessing, that is, to adding a product of two rectangular Gaussian matrices. 4. By applying our randomization techniques to a matrix that has numerical rank r, we accelerate the known algorithms for the approximation of its leading and trailing singular spaces associated with its r largest and with all its remaining singular values, respectively. 5. Our algorithms use much fewer random parameters and run much faster when various random sparse and structured preprocessors replace Gaussian. Empirically the outputs of the resulting algorithms is as accurate as the outputs under Gaussian preprocessing. 6. Our novel duality techniques provides formal support, so far missing, for these empirical observations and opens door to derandomization of our preprocessing and to further acceleration and simplification of our algorithms by using more efficient sparse and structured preprocessors. 7. Our techniques and our progress can be applied to various other fundamental matrix computations such as the celebrated low-rank approximation of a matrix by means of random sampling.

DeepSeek-AI, Daya Guo, Dejian Yang et al. · stanford, tsinghua

We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrates remarkable reasoning capabilities. Through RL, DeepSeek-R1-Zero naturally emerges with numerous powerful and intriguing reasoning behaviors. However, it encounters challenges such as poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks. To support the research community, we open-source DeepSeek-R1-Zero, DeepSeek-R1, and six dense models (1.5B, 7B, 8B, 14B, 32B, 70B) distilled from DeepSeek-R1 based on Qwen and Llama.

Yuyang Gao, Tong Steven Sun, Guangji Bai et al.

Despite the fast progress of explanation techniques in modern Deep Neural Networks (DNNs) where the main focus is handling "how to generate the explanations", advanced research questions that examine the quality of the explanation itself (e.g., "whether the explanations are accurate") and improve the explanation quality (e.g., "how to adjust the model to generate more accurate explanations when explanations are inaccurate") are still relatively under-explored. To guide the model toward better explanations, techniques in explanation supervision - which add supervision signals on the model explanation - have started to show promising effects on improving both the generalizability as and intrinsic interpretability of Deep Neural Networks. However, the research on supervising explanations, especially in vision-based applications represented through saliency maps, is in its early stage due to several inherent challenges: 1) inaccuracy of the human explanation annotation boundary, 2) incompleteness of the human explanation annotation region, and 3) inconsistency of the data distribution between human annotation and model explanation maps. To address the challenges, we propose a generic RES framework for guiding visual explanation by developing a novel objective that handles inaccurate boundary, incomplete region, and inconsistent distribution of human annotations, with a theoretical justification on model generalizability. Extensive experiments on two real-world image datasets demonstrate the effectiveness of the proposed framework on enhancing both the reasonability of the explanation and the performance of the backbone DNNs model.

Tong Steven Sun, Yuyang Gao, Shubham Khaladkar et al.

The local explanation provides heatmaps on images to explain how Convolutional Neural Networks (CNNs) derive their output. Due to its visual straightforwardness, the method has been one of the most popular explainable AI (XAI) methods for diagnosing CNNs. Through our formative study (S1), however, we captured ML engineers' ambivalent perspective about the local explanation as a valuable and indispensable envision in building CNNs versus the process that exhausts them due to the heuristic nature of detecting vulnerability. Moreover, steering the CNNs based on the vulnerability learned from the diagnosis seemed highly challenging. To mitigate the gap, we designed DeepFuse, the first interactive design that realizes the direct feedback loop between a user and CNNs in diagnosing and revising CNN's vulnerability using local explanations. DeepFuse helps CNN engineers to systemically search "unreasonable" local explanations and annotate the new boundaries for those identified as unreasonable in a labor-efficient manner. Next, it steers the model based on the given annotation such that the model doesn't introduce similar mistakes. We conducted a two-day study (S2) with 12 experienced CNN engineers. Using DeepFuse, participants made a more accurate and "reasonable" model than the current state-of-the-art. Also, participants found the way DeepFuse guides case-based reasoning can practically improve their current practice. We provide implications for design that explain how future HCI-driven design can move our practice forward to make XAI-driven insights more actionable.

Jianshen Zhu, Naveed Ahmed Azam, Shengjuan Cao et al.

A novel framework has recently been proposed for designing the molecular structure of chemical compounds with a desired chemical property, where design of novel drugs is an important topic in bioinformatics and chemo-informatics. The framework infers a desired chemical graph by solving a mixed integer linear program (MILP) that simulates the computation process of a feature function defined by a two-layered model on chemical graphs and a prediction function constructed by a machine learning method. A set of graph theoretical descriptors in the feature function plays a key role to derive a compact formulation of such an MILP. To improve the learning performance of prediction functions in the framework maintaining the compactness of the MILP, this paper utilizes the product of two of those descriptors as a new descriptor and then designs a method of reducing the number of descriptors. The results of our computational experiments suggest that the proposed method improved the learning performance for many chemical properties and can infer a chemical structure with up to 50 non-hydrogen atoms.

DeepSeek-AI, Aixin Liu, Bei Feng et al. · pku

We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times. We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential. Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.

Taha ValizadehAslani, Yiwen Shi, Jing Wang et al.

Classification on long-tailed distributed data is a challenging problem, which suffers from serious class-imbalance and hence poor performance on tail classes with only a few samples. Owing to this paucity of samples, learning on the tail classes is especially challenging for the fine-tuning when transferring a pretrained model to a downstream task. In this work, we present a simple modification of standard fine-tuning to cope with these challenges. Specifically, we propose a two-stage fine-tuning: we first fine-tune the final layer of the pretrained model with class-balanced reweighting loss, and then we perform the standard fine-tuning. Our modification has several benefits: (1) it leverages pretrained representations by only fine-tuning a small portion of the model parameters while keeping the rest untouched; (2) it allows the model to learn an initial representation of the specific task; and importantly (3) it protects the learning of tail classes from being at a disadvantage during the model updating. We conduct extensive experiments on synthetic datasets of both two-class and multi-class tasks of text classification as well as a real-world application to ADME (i.e., absorption, distribution, metabolism, and excretion) semantic labeling. The experimental results show that the proposed two-stage fine-tuning outperforms both fine-tuning with conventional loss and fine-tuning with a reweighting loss on the above datasets.

Chen Ling, Junji Jiang, Junxiang Wang et al.

Graph diffusion problems such as the propagation of rumors, computer viruses, or smart grid failures are ubiquitous and societal. Hence it is usually crucial to identify diffusion sources according to the current graph diffusion observations. Despite its tremendous necessity and significance in practice, source localization, as the inverse problem of graph diffusion, is extremely challenging as it is ill-posed: different sources may lead to the same graph diffusion patterns. Different from most traditional source localization methods, this paper focuses on a probabilistic manner to account for the uncertainty of different candidate sources. Such endeavors require overcoming challenges including 1) the uncertainty in graph diffusion source localization is hard to be quantified; 2) the complex patterns of the graph diffusion sources are difficult to be probabilistically characterized; 3) the generalization under any underlying diffusion patterns is hard to be imposed. To solve the above challenges, this paper presents a generic framework: Source Localization Variational AutoEncoder (SL-VAE) for locating the diffusion sources under arbitrary diffusion patterns. Particularly, we propose a probabilistic model that leverages the forward diffusion estimation model along with deep generative models to approximate the diffusion source distribution for quantifying the uncertainty. SL-VAE further utilizes prior knowledge of the source-observation pairs to characterize the complex patterns of diffusion sources by a learned generative prior. Lastly, a unified objective that integrates the forward diffusion estimation model is derived to enforce the model to generalize under arbitrary diffusion patterns. Extensive experiments are conducted on 7 real-world datasets to demonstrate the superiority of SL-VAE in reconstructing the diffusion sources by excelling other methods on average 20% in AUC score.

Chen Ling, Tanmoy Chowdhury, Junji Jiang et al.

Analogical reasoning is the process of discovering and mapping correspondences from a target subject to a base subject. As the most well-known computational method of analogical reasoning, Structure-Mapping Theory (SMT) abstracts both target and base subjects into relational graphs and forms the cognitive process of analogical reasoning by finding a corresponding subgraph (i.e., correspondence) in the target graph that is aligned with the base graph. However, incorporating deep learning for SMT is still under-explored due to several obstacles: 1) the combinatorial complexity of searching for the correspondence in the target graph; 2) the correspondence mining is restricted by various cognitive theory-driven constraints. To address both challenges, we propose a novel framework for Analogical Reasoning (DeepGAR) that identifies the correspondence between source and target domains by assuring cognitive theory-driven constraints. Specifically, we design a geometric constraint embedding space to induce subgraph relation from node embeddings for efficient subgraph search. Furthermore, we develop novel learning and optimization strategies that could end-to-end identify correspondences that are strictly consistent with constraints driven by the cognitive theory. Extensive experiments are conducted on synthetic and real-world datasets to demonstrate the effectiveness of the proposed DeepGAR over existing methods.

Liang Zhao, Kunming Shao, Zhipeng Liao et al.

FP8 low-precision formats have gained significant adoption in Transformer inference and training. However, existing digital compute-in-memory (DCIM) architectures face challenges in supporting variable FP8 aligned-mantissa bitwidths, as unified alignment strategies and fixed-precision multiply-accumulate (MAC) units struggle to handle input data with diverse distributions. This work presents a flexible FP8 DCIM accelerator with three innovations: (1) a dynamic shift-aware bitwidth prediction (DSBP) with on-the-fly input prediction that adaptively adjusts weight (2/4/6/8b) and input (2$\sim$12b) aligned-mantissa precision; (2) a FIFO-based input alignment unit (FIAU) replacing complex barrel shifters with pointer-based control; and (3) a precision-scalable INT MAC array achieving flexible weight precision with minimal overhead. Implemented in 28nm CMOS with a 64$\times$96 CIM array, the design achieves 20.4 TFLOPS/W for fixed E5M7, demonstrating 2.8$\times$ higher FP8 efficiency than previous work while supporting all FP8 formats. Results on Llama-7b show that the DSBP achieves higher efficiency than fixed bitwidth mode at the same accuracy level on both BoolQ and Winogrande datasets, with configurable parameters enabling flexible accuracy-efficiency trade-offs.

Jianshen Zhu, Naveed Ahmed Azam, Kazuya Haraguchi et al.

A novel framework for designing the molecular structure of chemical compounds with a desired chemical property has recently been proposed. The framework infers a desired chemical graph by solving a mixed integer linear program (MILP) that simulates the computation process of a feature function defined by a two-layered model on chemical graphs and a prediction function constructed by a machine learning method. To improve the learning performance of prediction functions in the framework, we design a method that splits a given data set $\mathcal{C}$ into two subsets $\mathcal{C}^{(i)},i=1,2$ by a hyperplane in a chemical space so that most compounds in the first (resp., second) subset have observed values lower (resp., higher) than a threshold $θ$. We construct a prediction function $ψ$ to the data set $\mathcal{C}$ by combining prediction functions $ψ_i,i=1,2$ each of which is constructed on $\mathcal{C}^{(i)}$ independently. The results of our computational experiments suggest that the proposed method improved the learning performance for several chemical properties to which a good prediction function has been difficult to construct.

Ailin Huang, Ang Li, Aobo Kong et al.

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

Zhiyu Wu, Xiaokang Chen, Zizheng Pan et al.

We present DeepSeek-VL2, an advanced series of large Mixture-of-Experts (MoE) Vision-Language Models that significantly improves upon its predecessor, DeepSeek-VL, through two key major upgrades. For the vision component, we incorporate a dynamic tiling vision encoding strategy designed for processing high-resolution images with different aspect ratios. For the language component, we leverage DeepSeekMoE models with the Multi-head Latent Attention mechanism, which compresses Key-Value cache into latent vectors, to enable efficient inference and high throughput. Trained on an improved vision-language dataset, DeepSeek-VL2 demonstrates superior capabilities across various tasks, including but not limited to visual question answering, optical character recognition, document/table/chart understanding, and visual grounding. Our model series is composed of three variants: DeepSeek-VL2-Tiny, DeepSeek-VL2-Small and DeepSeek-VL2, with 1.0B, 2.8B and 4.5B activated parameters respectively. DeepSeek-VL2 achieves competitive or state-of-the-art performance with similar or fewer activated parameters compared to existing open-source dense and MoE-based models. Codes and pre-trained models are publicly accessible at https://github.com/deepseek-ai/DeepSeek-VL2.

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.

Zheng Zhang, Hossein Amiri, Zhenke Liu et al.

Identifying anomalous human spatial trajectory patterns can indicate dynamic changes in mobility behavior with applications in domains like infectious disease monitoring and elderly care. Recent advancements in large language models (LLMs) have demonstrated their ability to reason in a manner akin to humans. This presents significant potential for analyzing temporal patterns in human mobility. In this paper, we conduct empirical studies to assess the capabilities of leading LLMs like GPT-4 and Claude-2 in detecting anomalous behaviors from mobility data, by comparing to specialized methods. Our key findings demonstrate that LLMs can attain reasonable anomaly detection performance even without any specific cues. In addition, providing contextual clues about potential irregularities could further enhances their prediction efficacy. Moreover, LLMs can provide reasonable explanations for their judgments, thereby improving transparency. Our work provides insights on the strengths and limitations of LLMs for human spatial trajectory analysis.

Guangji Bai, Chen Ling, Liang Zhao

Temporal domain generalization is a promising yet extremely challenging area where the goal is to learn models under temporally changing data distributions and generalize to unseen data distributions following the trends of the change. The advancement of this area is challenged by: 1) characterizing data distribution drift and its impacts on models, 2) expressiveness in tracking the model dynamics, and 3) theoretical guarantee on the performance. To address them, we propose a Temporal Domain Generalization with Drift-Aware Dynamic Neural Network (DRAIN) framework. Specifically, we formulate the problem into a Bayesian framework that jointly models the relation between data and model dynamics. We then build a recurrent graph generation scenario to characterize the dynamic graph-structured neural networks learned across different time points. It captures the temporal drift of model parameters and data distributions and can predict models in the future without the presence of future data. In addition, we explore theoretical guarantees of the model performance under the challenging temporal DG setting and provide theoretical analysis, including uncertainty and generalization error. Finally, extensive experiments on several real-world benchmarks with temporal drift demonstrate the effectiveness and efficiency of the proposed method.

Zheng Zhang, Chen Zheng, Da Tang et al.

This paper introduces a multifaceted methodology for fine-tuning and evaluating large language models (LLMs) for specialized monetization tasks. The goal is to balance general language proficiency with domain-specific skills. The methodology has three main components: 1) Carefully blending in-domain and general-purpose data during fine-tuning to achieve an optimal balance between general and specialized capabilities; 2) Designing a comprehensive evaluation framework with 45 questions tailored to assess performance on functionally relevant dimensions like reliability, consistency, and business impact; 3) Analyzing how model size and continual training influence metrics to guide efficient resource allocation during fine-tuning. The paper details the design, data collection, analytical techniques, and results validating the proposed frameworks. It aims to provide businesses and researchers with actionable insights on effectively adapting LLMs for specialized contexts. We also intend to make public the comprehensive evaluation framework, which includes the 45 tailored questions and their respective scoring guidelines, to foster transparency and collaboration in adapting LLMs for specialized tasks.

Muhammad Hamza, Ammar Hawbani, Sami Ul Rehman et al.

Underwater image processing and analysis have been a hotspot of study in recent years, as more emphasis has been focused to underwater monitoring and usage of marine resources. Compared with the open environment, underwater image encountered with more complicated conditions such as light abortion, scattering, turbulence, nonuniform illumination and color diffusion. Although considerable advances and enhancement techniques achieved in resolving these issues, they treat low-frequency information equally across the entire channel, which results in limiting the network's representativeness. We propose a deep learning and feature-attention-based end-to-end network (FA-Net) to solve this problem. In particular, we propose a Residual Feature Attention Block (RFAB), containing the channel attention, pixel attention, and residual learning mechanism with long and short skip connections. RFAB allows the network to focus on learning high-frequency information while skipping low-frequency information on multi-hop connections. The channel and pixel attention mechanism considers each channel's different features and the uneven distribution of haze over different pixels in the image. The experimental results shows that the FA-Net propose by us provides higher accuracy, quantitatively and qualitatively and superiority to previous state-of-the-art methods.

Yuntong Hu, Zheng Zhang, Liang Zhao

Large language models (LLMs) have achieved impressive performance on many natural language processing tasks. However, their capabilities on graph-structured data remain relatively unexplored. In this paper, we conduct a series of experiments benchmarking leading LLMs on diverse graph prediction tasks spanning node, edge, and graph levels. We aim to assess whether LLMs can effectively process graph data and leverage topological structures to enhance performance, compared to specialized graph neural networks. Through varied prompt formatting and task/dataset selection, we analyze how well LLMs can interpret and utilize graph structures. By comparing LLMs' performance with specialized graph models, we offer insights into the strengths and limitations of employing LLMs for graph analytics. Our findings provide insights into LLMs' capabilities and suggest avenues for further exploration in applying them to graph analytics.

Guangji Bai, Ziyang Yu, Zheng Chai et al.

Despite the recent success of Graph Neural Networks (GNNs), it remains challenging to train GNNs on large-scale graphs due to neighbor explosions. As a remedy, distributed computing becomes a promising solution by leveraging abundant computing resources (e.g., GPU). However, the node dependency of graph data increases the difficulty of achieving high concurrency in distributed GNN training, which suffers from the massive communication overhead. To address it, Historical value approximation is deemed a promising class of distributed training techniques. It utilizes an offline memory to cache historical information (e.g., node embedding) as an affordable approximation of the exact value and achieves high concurrency. However, such benefits come at the cost of involving dated training information, leading to staleness, imprecision, and convergence issues. To overcome these challenges, this paper proposes SAT (Staleness-Alleviated Training), a novel and scalable distributed GNN training framework that reduces the embedding staleness adaptively. The key idea of SAT is to model the GNN's embedding evolution as a temporal graph and build a model upon it to predict future embedding, which effectively alleviates the staleness of the cached historical embedding. We propose an online algorithm to train the embedding predictor and the distributed GNN alternatively and further provide a convergence analysis. Empirically, we demonstrate that SAT can effectively reduce embedding staleness and thus achieve better performance and convergence speed on multiple large-scale graph datasets.