Jie Cheng

Jie Cheng, Xiaodong Mei, Ming Liu

This study explores the application of self-supervised learning (SSL) to the task of motion forecasting, an area that has not yet been extensively investigated despite the widespread success of SSL in computer vision and natural language processing. To address this gap, we introduce Forecast-MAE, an extension of the mask autoencoders framework that is specifically designed for self-supervised learning of the motion forecasting task. Our approach includes a novel masking strategy that leverages the strong interconnections between agents' trajectories and road networks, involving complementary masking of agents' future or history trajectories and random masking of lane segments. Our experiments on the challenging Argoverse 2 motion forecasting benchmark show that Forecast-MAE, which utilizes standard Transformer blocks with minimal inductive bias, achieves competitive performance compared to state-of-the-art methods that rely on supervised learning and sophisticated designs. Moreover, it outperforms the previous self-supervised learning method by a significant margin. Code is available at https://github.com/jchengai/forecast-mae.

Saizhe Ding, Jinze Chen, Yang Wang et al.

Event cameras, such as dynamic vision sensors (DVS), are biologically inspired vision sensors that have advanced over conventional cameras in high dynamic range, low latency and low power consumption, showing great application potential in many fields. Event cameras are more sensitive to junction leakage current and photocurrent as they output differential signals, losing the smoothing function of the integral imaging process in the RGB camera. The logarithmic conversion further amplifies noise, especially in low-contrast conditions. Recently, researchers proposed a series of datasets and evaluation metrics but limitations remain: 1) the existing datasets are small in scale and insufficient in noise diversity, which cannot reflect the authentic working environments of event cameras; and 2) the existing denoising evaluation metrics are mostly referenced evaluation metrics, relying on APS information or manual annotation. To address the above issues, we construct a large-scale event denoising dataset (multilevel benchmark for event denoising, E-MLB) for the first time, which consists of 100 scenes, each with four noise levels, that is 12 times larger than the largest existing denoising dataset. We also propose the first nonreference event denoising metric, the event structural ratio (ESR), which measures the structural intensity of given events. ESR is inspired by the contrast metric, but is independent of the number of events and projection direction. Based on the proposed benchmark and ESR, we evaluate the most representative denoising algorithms, including classic and SOTA, and provide denoising baselines under various scenes and noise levels. The corresponding results and codes are available at https://github.com/KugaMaxx/cuke-emlb.

Yushun Tang, Ce Zhang, Heng Xu et al. · cmu

Fully test-time adaptation aims to adapt the network model based on sequential analysis of input samples during the inference stage to address the cross-domain performance degradation problem of deep neural networks. We take inspiration from the biological plausibility learning where the neuron responses are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules. Based on these feed-forward learning rules, we design a soft Hebbian learning process which provides an unsupervised and effective mechanism for online adaptation. We observe that the performance of this feed-forward Hebbian learning for fully test-time adaptation can be significantly improved by incorporating a feedback neuro-modulation layer. It is able to fine-tune the neuron responses based on the external feedback generated by the error back-propagation from the top inference layers. This leads to our proposed neuro-modulated Hebbian learning (NHL) method for fully test-time adaptation. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods.

Jie Cheng, Yingbing Chen, Xiaodong Mei et al.

In recent years, imitation-based driving planners have reported considerable success. However, due to the absence of a standardized benchmark, the effectiveness of various designs remains unclear. The newly released nuPlan addresses this issue by offering a large-scale real-world dataset and a standardized closed-loop benchmark for equitable comparisons. Utilizing this platform, we conduct a comprehensive study on two fundamental yet underexplored aspects of imitation-based planners: the essential features for ego planning and the effective data augmentation techniques to reduce compounding errors. Furthermore, we highlight an imitation gap that has been overlooked by current learning systems. Finally, integrating our findings, we propose a strong baseline model-PlanTF. Our results demonstrate that a well-designed, purely imitation-based planner can achieve highly competitive performance compared to state-of-the-art methods involving hand-crafted rules and exhibit superior generalization capabilities in long-tail cases. Our models and benchmarks are publicly available. Project website https://jchengai.github.io/planTF.

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.

Jingcheng Hu, Yinmin Zhang, Shijie Shang et al.

We introduce Parallel Coordinated Reasoning (PaCoRe), a training-and-inference framework designed to overcome a central limitation of contemporary language models: their inability to scale test-time compute (TTC) far beyond sequential reasoning under a fixed context window. PaCoRe departs from the traditional sequential paradigm by driving TTC through massive parallel exploration coordinated via a message-passing architecture in multiple rounds. Each round launches many parallel reasoning trajectories, compacts their findings into context-bounded messages, and synthesizes these messages to guide the next round and ultimately produce the final answer. Trained end-to-end with large-scale, outcome-based reinforcement learning, the model masters the synthesis abilities required by PaCoRe and scales to multi-million-token effective TTC without exceeding context limits. The approach yields strong improvements across diverse domains, and notably pushes reasoning beyond frontier systems in mathematics: an 8B model reaches 94.5% on HMMT 2025, surpassing GPT-5's 93.2% by scaling effective TTC to roughly two million tokens. We open-source model checkpoints, training data, and the full inference pipeline to accelerate follow-up work.

Pengfei Zhao, Qigong Sun, Xiaolin Tian et al.

There has been growing interest in facial video-based remote photoplethysmography (rPPG) measurement recently, with a focus on assessing various vital signs such as heart rate and heart rate variability. Despite previous efforts on static datasets, their approaches have been hindered by inaccurate region of interest (ROI) localization and motion issues, and have shown limited generalization in real-world scenarios. To address these challenges, we propose a novel masked attention regularization (MAR-rPPG) framework that mitigates the impact of ROI localization and complex motion artifacts. Specifically, our approach first integrates a masked attention regularization mechanism into the rPPG field to capture the visual semantic consistency of facial clips, while it also employs a masking technique to prevent the model from overfitting on inaccurate ROIs and subsequently degrading its performance. Furthermore, we propose an enhanced rPPG expert aggregation (EREA) network as the backbone to obtain rPPG signals and attention maps simultaneously. Our EREA network is capable of discriminating divergent attentions from different facial areas and retaining the consistency of spatiotemporal attention maps. For motion robustness, a simple open source detector MediaPipe for data preprocessing is sufficient for our framework due to its superior capability of rPPG signal extraction and attention regularization. Exhaustive experiments on three benchmark datasets (UBFC-rPPG, PURE, and MMPD) substantiate the superiority of our proposed method, outperforming recent state-of-the-art works by a considerable margin.

Guiqin Wang, Peng Zhao, Cong Zhao et al.

Weakly-supervised action localization aims to recognize and localize action instancese in untrimmed videos with only video-level labels. Most existing models rely on multiple instance learning(MIL), where the predictions of unlabeled instances are supervised by classifying labeled bags. The MIL-based methods are relatively well studied with cogent performance achieved on classification but not on localization. Generally, they locate temporal regions by the video-level classification but overlook the temporal variations of feature semantics. To address this problem, we propose a novel attention-based hierarchically-structured latent model to learn the temporal variations of feature semantics. Specifically, our model entails two components, the first is an unsupervised change-points detection module that detects change-points by learning the latent representations of video features in a temporal hierarchy based on their rates of change, and the second is an attention-based classification model that selects the change-points of the foreground as the boundaries. To evaluate the effectiveness of our model, we conduct extensive experiments on two benchmark datasets, THUMOS-14 and ActivityNet-v1.3. The experiments show that our method outperforms current state-of-the-art methods, and even achieves comparable performance with fully-supervised methods.

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.

Rui Zhang, Luziwei Leng, Kaiwei Che et al.

Spiking neural networks (SNNs), known for their low-power, event-driven computation and intrinsic temporal dynamics, are emerging as promising solutions for processing dynamic, asynchronous signals from event-based sensors. Despite their potential, SNNs face challenges in training and architectural design, resulting in limited performance in challenging event-based dense prediction tasks compared to artificial neural networks (ANNs). In this work, we develop an efficient spiking encoder-decoder network (SpikingEDN) for large-scale event-based semantic segmentation tasks. To enhance the learning efficiency from dynamic event streams, we harness the adaptive threshold which improves network accuracy, sparsity and robustness in streaming inference. Moreover, we develop a dual-path Spiking Spatially-Adaptive Modulation module, which is specifically tailored to enhance the representation of sparse events and multi-modal inputs, thereby considerably improving network performance. Our SpikingEDN attains a mean intersection over union (MIoU) of 72.57\% on the DDD17 dataset and 58.32\% on the larger DSEC-Semantic dataset, showing competitive results to the state-of-the-art ANNs while requiring substantially fewer computational resources. Our results shed light on the untapped potential of SNNs in event-based vision applications. The source code will be made publicly available.

Jie Cheng, Qishuai Liu, Qing Hui et al.

Electricity, water, and gas systems are critical infrastructures that are sustaining our daily lives. This paper studies the joint operation of these systems through a proposed optimization model and explores the advantage of considering the system of systems. Individual and joint optimizations are studied and compared. The numerical results show that the total electricity cost for these three systems can be reduced by 9% via joint optimization. Because the water system and gas system intrinsically include the storages in their systems, the power system can use these storages as the regulation capacity to shift load from peak hours to off-peak hours. Since the saving on the power generation cost surpasses the incremental cost in the operation and maintenance (O&M), the overall economic performance is improved by the joint optimization.

Jie Cheng, Gang Xiong, Xingyuan Dai et al.

Preference-based Reinforcement Learning (PbRL) circumvents the need for reward engineering by harnessing human preferences as the reward signal. However, current PbRL methods excessively depend on high-quality feedback from domain experts, which results in a lack of robustness. In this paper, we present RIME, a robust PbRL algorithm for effective reward learning from noisy preferences. Our method utilizes a sample selection-based discriminator to dynamically filter out noise and ensure robust training. To counteract the cumulative error stemming from incorrect selection, we suggest a warm start for the reward model, which additionally bridges the performance gap during the transition from pre-training to online training in PbRL. Our experiments on robotic manipulation and locomotion tasks demonstrate that RIME significantly enhances the robustness of the state-of-the-art PbRL method. Code is available at https://github.com/CJReinforce/RIME_ICML2024.

Jie Cheng, Gang Xiong, Ruixi Qiao et al.

Process reward models (PRMs) have proven effective for test-time scaling of Large Language Models (LLMs) on challenging reasoning tasks. However, reward hacking issues with PRMs limit their successful application in reinforcement fine-tuning. In this paper, we identify the main cause of PRM-induced reward hacking: the canonical summation-form credit assignment in reinforcement learning (RL), which defines the value as cumulative gamma-decayed future rewards, easily induces LLMs to hack steps with high rewards. To address this, we propose PURE: Process sUpervised Reinforcement lEarning. The key innovation of PURE is a min-form credit assignment that formulates the value function as the minimum of future rewards. This method significantly alleviates reward hacking by limiting the value function range and distributing advantages more reasonably. Through extensive experiments on 3 base models, we show that PRM-based approaches enabling min-form credit assignment achieve comparable reasoning performance to verifiable reward-based methods within only 30% steps. In contrast, the canonical sum-form credit assignment collapses training even at the beginning! Additionally, when we supplement PRM-based fine-tuning with just 10% verifiable rewards, we further alleviate reward hacking and produce the best fine-tuned model based on Qwen2.5-Math-7B in our experiments, achieving 82.5% accuracy on AMC23 and 53.3% average accuracy across 5 benchmarks. Moreover, we summarize the observed reward hacking cases and analyze the causes of training collapse. We release our code and model weights at https://github.com/CJReinforce/PURE.

Tongtian Yue, Jie Cheng, Longteng Guo et al.

Recent trends in Large Vision Language Models (LVLMs) research have been increasingly focusing on advancing beyond general image understanding towards more nuanced, object-level referential comprehension. In this paper, we present and delve into the self-consistency capability of LVLMs, a crucial aspect that reflects the models' ability to both generate informative captions for specific objects and subsequently utilize these captions to accurately re-identify the objects in a closed-loop process. This capability significantly mirrors the precision and reliability of fine-grained visual-language understanding. Our findings reveal that the self-consistency level of existing LVLMs falls short of expectations, posing limitations on their practical applicability and potential. To address this gap, we introduce a novel fine-tuning paradigm named Self-Consistency Tuning (SC-Tune). It features the synergistic learning of a cyclic describer-locator system. This paradigm is not only data-efficient but also exhibits generalizability across multiple LVLMs. Through extensive experiments, we demonstrate that SC-Tune significantly elevates performance across a spectrum of object-level vision-language benchmarks and maintains competitive or improved performance on image-level vision-language benchmarks. Both our model and code will be publicly available at https://github.com/ivattyue/SC-Tune.

Peng Tao, Xiaohu Hao, Jie Cheng et al.

Making an accurate prediction of an unknown system only from a short-term time series is difficult due to the lack of sufficient information, especially in a multi-step-ahead manner. However, a high-dimensional short-term time series contains rich dynamical information, and also becomes increasingly available in many fields. In this work, by exploiting spatiotemporal information (STI) transformation scheme that transforms such high-dimensional/spatial information to temporal information, we developed a new method called MT-GPRMachine to achieve accurate prediction from a short-term time series. Specifically, we first construct a specific multi-task GPR which is multiple linked STI mappings to transform high dimensional/spatial information into temporal/dynamical information of any given target variable, and then makes multi step-ahead prediction of the target variable by solving those STI mappings. The multi-step-ahead prediction results on various synthetic and real-world datasets clearly validated that MT-GPRMachine outperformed other existing approaches.

StepFun, Bin Wang, Bojun Wang et al.

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

Qiuli Wang, Xinhuan Sun, Fengxi Chen et al.

Gadoxetate disodium-enhanced MRI is essential for the detection and characterization of hepatocellular carcinoma. However, acquisition of the hepatobiliary phase (HBP) requires a prolonged post-contrast delay, which reduces workflow efficiency and increases the risk of motion artifacts. In this study, we propose a Triple-Phase Sequential Fusion Network (TriPF-Net) to synthesize HBP images by leveraging the sequential information from pre-HBP sequences: while T1-weighted imaging serves as the indispensable baseline, the model adaptively integrates arterial-phase (AP) and venous-phase (VP) features when available. By modeling the tissue-specific contrast uptake and excretion dynamics across these three phases, TriPF-Net ensures robust HBP synthesis even under the stochastic absence of one or both dynamic contrast-enhanced sequences. The framework comprises an Enhanced Region-Guided Encoder and a Dynamic Feature Unification Module, optimized with a Region-Guided Sequential Fusion Loss to maintain physiological consistency. In addition, clinical variables, including age, sex, total bilirubin, and albumin, are incorporated to enhance physiological consistency. Compared with conventional methods, TriPF-Net achieved superior performance on datasets from two centers. On the internal dataset, the model achieved an MAE of 10.65, a PSNR of 23.27, and an SSIM of 0.76. On the external validation dataset, the corresponding values were 12.41, 23.11, and 0.78, respectively. This flexible solution enhances clinical workflow and lesion depiction, potentially eliminating the need for delayed HBP acquisition in HCC imaging.

Tongtian Yue, Longteng Guo, Jie Cheng et al.

In the era of Large Language Models (LLMs), Mixture-of-Experts (MoE) architectures offer a promising approach to managing computational costs while scaling up model parameters. Conventional MoE-based LLMs typically employ static Top-K routing, which activates a fixed and equal number of experts for each token regardless of their significance within the context. In this paper, we propose a novel Ada-K routing strategy that dynamically adjusts the number of activated experts for each token, thereby improving the balance between computational efficiency and model performance. Specifically, our strategy incorporates learnable and lightweight allocator modules that decide customized expert resource allocation tailored to the contextual needs for each token. These allocators are designed to be fully pluggable, making it broadly applicable across all mainstream MoE-based LLMs. We leverage the Proximal Policy Optimization (PPO) algorithm to facilitate an end-to-end learning process for this non-differentiable decision-making framework. Extensive evaluations on four popular baseline models demonstrate that our Ada-K routing method significantly outperforms conventional Top-K routing. Compared to Top-K, our method achieves over 25% reduction in FLOPs and more than 20% inference speedup while still improving performance across various benchmarks. Moreover, the training of Ada-K is highly efficient. Even for Mixtral-8x22B, a MoE-based LLM with more than 140B parameters, the training time is limited to 8 hours. Detailed analysis shows that harder tasks, middle layers, and content words tend to activate more experts, providing valuable insights for future adaptive MoE system designs. Both the training code and model checkpoints will be publicly available.

Sheng Wang, Yao Tian, Xiaodong Mei et al.

Decision-making and planning in autonomous driving critically reflect the safety of the system, making effective planning imperative. Current imitation learning-based planning algorithms often merge historical trajectories with present observations to predict future candidate paths. However, these algorithms typically assess the current and historical plans independently, leading to discontinuities in driving intentions and an accumulation of errors with each step in a discontinuous plan. To tackle this challenge, this paper introduces LHPF, an imitation learning planner that integrates historical planning information. Our approach employs a historical intention aggregation module that pools historical planning intentions, which are then combined with a spatial query vector to decode the final planning trajectory. Furthermore, we incorporate a comfort auxiliary task to enhance the human-like quality of the driving behavior. Extensive experiments using both real-world and synthetic data demonstrate that LHPF not only surpasses existing advanced learning-based planners in planning performance but also marks the first instance of a purely learning-based planner outperforming the expert. Additionally, the application of the historical intention aggregation module across various backbones highlights the considerable potential of the proposed method. The code will be made publicly available.

Xiaodong Mei, Sheng Wang, Jie Cheng et al.

Motion forecasting represents a critical challenge in autonomous driving systems, requiring accurate prediction of surrounding agents' future trajectories. While existing approaches predict future motion states with the extracted scene context feature from historical agent trajectories and road layouts, they suffer from the information degradation during the scene feature encoding. To address the limitation, we propose HAMF, a novel motion forecasting framework that learns future motion representations with the scene context encoding jointly, to coherently combine the scene understanding and future motion state prediction. We first embed the observed agent states and map information into 1D token sequences, together with the target multi-modal future motion features as a set of learnable tokens. Then we design a unified Attention-based encoder, which synergistically combines self-attention and cross-attention mechanisms to model the scene context information and aggregate future motion features jointly. Complementing the encoder, we implement the Mamba module in the decoding stage to further preserve the consistency and correlations among the learned future motion representations, to generate the accurate and diverse final trajectories. Extensive experiments on Argoverse 2 benchmark demonstrate that our hybrid Attention-Mamba model achieves state-of-the-art motion forecasting performance with the simple and lightweight architecture.

Eric Cheng, Jie Cheng

Decision trees are a commonly used class of machine learning models valued for their interpretability and versatility, capable of both classification and regression. We propose ZTree, a novel decision tree learning framework that replaces CART's traditional purity based splitting with statistically principled subgroup identification. At each node, ZTree applies hypothesis testing (e.g., z-tests, t-tests, Mann-Whitney U, log-rank) to assess whether a candidate subgroup differs meaningfully from the complement. To adjust for the complication of multiple testing, we employ a cross-validation-based approach to determine if further node splitting is needed. This robust stopping criterion eliminates the need for post-pruning and makes the test threshold (z-threshold) the only parameter for controlling tree complexity. Because of the simplicity of the tree growing procedure, once a detailed tree is learned using the most lenient z-threshold, all simpler trees can be derived by simply removing nodes that do not meet the larger z-thresholds. This makes parameter tuning intuitive and efficient. Furthermore, this z-threshold is essentially a p-value, allowing users to easily plug in appropriate statistical tests into our framework without adjusting the range of parameter search. Empirical evaluation on five large-scale UCI datasets demonstrates that ZTree consistently delivers strong performance, especially at low data regimes. Compared to CART, ZTree also tends to grow simpler trees without sacrificing performance. ZTree introduces a statistically grounded alternative to traditional decision tree splitting by leveraging hypothesis testing and a cross-validation approach to multiple testing correction, resulting in an efficient and flexible framework.

Haimei Pan, Jiyun Zhang, Qinxi Wei et al.

Fire is a highly destructive disaster, but effective prevention can significantly reduce its likelihood of occurrence. When it happens, deploying emergency robots in fire-risk scenarios can help minimize the danger to human responders. However, current research on pre-disaster warnings and disaster-time rescue still faces significant challenges due to incomplete perception, inadequate fire situational awareness, and delayed response. To enhance intelligent perception and response planning for robots in fire scenarios, we first construct a knowledge graph (KG) by leveraging large language models (LLMs) to integrate fire domain knowledge derived from fire prevention guidelines and fire rescue task information from robotic emergency response documents. We then propose a new framework called Insights-on-Graph (IOG), which integrates the structured fire information of KG and Large Multimodal Models (LMMs). The framework generates perception-driven risk graphs from real-time scene imagery to enable early fire risk detection and provide interpretable emergency responses for task module and robot component configuration based on the evolving risk situation. Extensive simulations and real-world experiments show that IOG has good applicability and practical application value in fire risk detection and rescue decision-making.

Shanghai AI Lab, Yicheng Bao, Guanxu Chen et al.

We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54\%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.

Jie Cheng, Hao Zheng, Meiguang Zheng et al.

Source-Free Domain Adaptation (SFDA) aims to train a target model without source data, and the key is to generate pseudo-labels using a pre-trained source model. However, we observe that the source model often produces highly uncertain pseudo-labels for hard samples, particularly those heavily affected by domain shifts, leading to these noisy pseudo-labels being introduced even before adaptation and further reinforced through parameter updates. Additionally, they continuously influence neighbor samples through propagation in the feature space.To eliminate the issue of noise accumulation, we propose a novel Progressive Curriculum Labeling (ElimPCL) method, which iteratively filters trustworthy pseudo-labeled samples based on prototype consistency to exclude high-noise samples from training. Furthermore, a Dual MixUP technique is designed in the feature space to enhance the separability of hard samples, thereby mitigating the interference of noisy samples on their neighbors.Extensive experiments validate the effectiveness of ElimPCL, achieving up to a 3.4% improvement on challenging tasks compared to state-of-the-art methods.

RuiXi Qiao, Jie Cheng, Xingyuan Dai et al.

Skills have been introduced to offline reinforcement learning (RL) as temporal abstractions to tackle complex, long-horizon tasks, promoting consistent behavior and enabling meaningful exploration. While skills in offline RL are predominantly modeled within a continuous latent space, the potential of discrete skill spaces remains largely underexplored. In this paper, we propose a compact discrete skill space for offline RL tasks supported by state-of-the-art transformer-based encoder and diffusion-based decoder. Coupled with a high-level policy trained via offline RL techniques, our method establishes a hierarchical RL framework where the trained diffusion decoder plays a pivotal role. Empirical evaluations show that the proposed algorithm, Discrete Diffusion Skill (DDS), is a powerful offline RL method. DDS performs competitively on Locomotion and Kitchen tasks and excels on long-horizon tasks, achieving at least a 12 percent improvement on AntMaze-v2 benchmarks compared to existing offline RL approaches. Furthermore, DDS offers improved interpretability, training stability, and online exploration compared to previous skill-based methods.

Yikang Zhang, Chuang-Wei Liu, Jiahang Li et al.

Road inspection is essential for ensuring road maintenance and traffic safety, as road defects gradually emerge and compromise road functionality. Traditional methods, which rely on manual evaluations, are labor-intensive, costly, and time-consuming. Although data-driven approaches are gaining traction, the scarcity and spatial sparsity of road defects in the real world pose significant challenges in acquiring high-quality datasets. Existing simulators designed to generate detailed synthetic driving scenes, however, lack models for road defects. Furthermore, advanced driving tasks involving interactions with road surfaces, such as planning and control in defective areas, remain underexplored. To address these limitations, we propose a system based on Urban Digital Twin (UDT) technology for intelligent road inspection. First, hierarchical road models are constructed from real-world driving data, creating highly detailed representations of road defect structures and surface elevations. Next, digital road twins are generated to create simulation environments for comprehensive analysis and evaluation. These scenarios are subsequently imported into a simulator to enable both data acquisition and physical simulation. Experimental results demonstrate that driving tasks, including perception and decision-making, can be significantly improved using the high-fidelity road defect scenes generated by our system.

Kai Zhu, Yang Cao, Wei Zhai et al.

Few-shot class-incremental learning is to recognize the new classes given few samples and not forget the old classes. It is a challenging task since representation optimization and prototype reorganization can only be achieved under little supervision. To address this problem, we propose a novel incremental prototype learning scheme. Our scheme consists of a random episode selection strategy that adapts the feature representation to various generated incremental episodes to enhance the corresponding extensibility, and a self-promoted prototype refinement mechanism which strengthens the expression ability of the new classes by explicitly considering the dependencies among different classes. Particularly, a dynamic relation projection module is proposed to calculate the relation matrix in a shared embedding space and leverage it as the factor for bootstrapping the update of prototypes. Extensive experiments on three benchmark datasets demonstrate the above-par incremental performance, outperforming state-of-the-art methods by a margin of 13%, 17% and 11%, respectively.

Dong He, Jie Cheng, Jong-Hwan Kim

Ground segmentation of point clouds remains challenging because of the sparse and unordered data structure. This paper proposes the GSECnet - Ground Segmentation network for Edge Computing, an efficient ground segmentation framework of point clouds specifically designed to be deployable on a low-power edge computing unit. First, raw point clouds are converted into a discretization representation by pillarization. Afterward, features of points within pillars are fed into PointNet to get the corresponding pillars feature map. Then, a depthwise-separable U-Net with the attention module learns the classification from the pillars feature map with an enormously diminished model parameter size. Our proposed framework is evaluated on SemanticKITTI against both point-based and discretization-based state-of-the-art learning approaches, and achieves an excellent balance between high accuracy and low computing complexity. Remarkably, our framework achieves the inference runtime of 135.2 Hz on a desktop platform. Moreover, experiments verify that it is deployable on a low-power edge computing unit powered 10 watts only.

Tianyu Liu, Qinghai Liao, Lu Gan et al.

Since early 2020, the coronavirus disease 2019 (COVID-19) has spread rapidly across the world. As at the date of writing this article, the disease has been globally reported in 223 countries and regions, infected over 108 million people and caused over 2.4 million deaths (https://covid19.who.int/, accessed on Feb. 17, 2021). Avoiding person-to-person transmission is an effective approach to control and prevent the pandemic. However, many daily activities, such as transporting goods in our daily life, inevitably involve person-to-person contact. Using an autonomous logistic vehicle to achieve contact-less goods transportation could alleviate this issue. For example, it can reduce the risk of virus transmission between the driver and customers. Moreover, many countries have imposed tough lockdown measures to reduce the virus transmission (e.g., retail, catering) during the pandemic, which causes inconveniences for human daily life. Autonomous vehicle can deliver the goods bought by humans, so that humans can get the goods without going out. These demands motivate us to develop an autonomous vehicle, named as Hercules, for contact-less goods transportation during the COVID-19 pandemic. The vehicle is evaluated through real-world delivering tasks under various traffic conditions.

Zhao Kang, Chong Peng, Jie Cheng et al.

Low-rank matrix is desired in many machine learning and computer vision problems. Most of the recent studies use the nuclear norm as a convex surrogate of the rank operator. However, all singular values are simply added together by the nuclear norm, and thus the rank may not be well approximated in practical problems. In this paper, we propose to use a log-determinant (LogDet) function as a smooth and closer, though non-convex, approximation to rank for obtaining a low-rank representation in subspace clustering. Augmented Lagrange multipliers strategy is applied to iteratively optimize the LogDet-based non-convex objective function on potentially large-scale data. By making use of the angular information of principal directions of the resultant low-rank representation, an affinity graph matrix is constructed for spectral clustering. Experimental results on motion segmentation and face clustering data demonstrate that the proposed method often outperforms state-of-the-art subspace clustering algorithms.

Jie Cheng, Tianxi Li, Elizaveta Levina et al.

While graphical models for continuous data (Gaussian graphical models) and discrete data (Ising models) have been extensively studied, there is little work on graphical models linking both continuous and discrete variables (mixed data), which are common in many scientific applications. We propose a novel graphical model for mixed data, which is simple enough to be suitable for high-dimensional data, yet flexible enough to represent all possible graph structures. We develop a computationally efficient regression-based algorithm for fitting the model by focusing on the conditional log-likelihood of each variable given the rest. The parameters have a natural group structure, and sparsity in the fitted graph is attained by incorporating a group lasso penalty, approximated by a weighted $\ell_1$ penalty for computational efficiency. We demonstrate the effectiveness of our method through an extensive simulation study and apply it to a music annotation data set (CAL500), obtaining a sparse and interpretable graphical model relating the continuous features of the audio signal to categorical variables such as genre, emotions, and usage associated with particular songs. While we focus on binary discrete variables, we also show that the proposed methodology can be easily extended to general discrete variables.

Jie Cheng, Russell Greiner

In this paper, we empirically evaluate algorithms for learning four types of Bayesian network (BN) classifiers - Naive-Bayes, tree augmented Naive-Bayes, BN augmented Naive-Bayes and general BNs, where the latter two are learned using two variants of a conditional-independence (CI) based BN-learning algorithm. Experimental results show the obtained classifiers, learned using the CI based algorithms, are competitive with (or superior to) the best known classifiers, based on both Bayesian networks and other formalisms; and that the computational time for learning and using these classifiers is relatively small. Moreover, these results also suggest a way to learn yet more effective classifiers; we demonstrate empirically that this new algorithm does work as expected. Collectively, these results argue that BN classifiers deserve more attention in machine learning and data mining communities.

Jie Cheng, Elizaveta Levina, Pei Wang et al.

There has been a lot of work fitting Ising models to multivariate binary data in order to understand the conditional dependency relationships between the variables. However, additional covariates are frequently recorded together with the binary data, and may influence the dependence relationships. Motivated by such a dataset on genomic instability collected from tumor samples of several types, we propose a sparse covariate dependent Ising model to study both the conditional dependency within the binary data and its relationship with the additional covariates. This results in subject-specific Ising models, where the subject's covariates influence the strength of association between the genes. As in all exploratory data analysis, interpretability of results is important, and we use L1 penalties to induce sparsity in the fitted graphs and in the number of selected covariates. Two algorithms to fit the model are proposed and compared on a set of simulated data, and asymptotic results are established. The results on the tumor dataset and their biological significance are discussed in detail.