Rui Yang

Hao Bai, Rui Yang, Chenlu Ye et al.

Training vision-language web agents with multi-step RL is compute-intensive, with two dominant forms of inefficiency: idle GPUs in synchronous RL, and trajectories that use more steps and tokens than necessary. We present AsyncWebRL, which addresses both. On the system side, an asynchronous design overlaps rollout, gradient update, and policy refresh across iterations, paired with two web-agent-specific adaptations, namely an everlasting rollout pool and lightweight screenshot handling, that together deliver up to a $2.9\times$ end-to-end training-throughput speedup over the previously fastest open synchronous pipeline (WebGym). On the algorithmic side, we identify the per-trajectory normalizer $1/|τ_i|$ in multi-step GRPO as the root cause of trajectory-level and token-level inefficiency: because failures are systematically longer than successes, it down-weights the negative gradient on failed tokens, so the policy keeps producing verbose memory schemas. Replacing $1/|τ_i|$ with a constant $1/k$ breaks this coupling, contracting trajectories while preserving aggregate success. Together, these contributions set a new open-source state of the art on the WebGym out-of-distribution test split (+5.8% relative over the 42.9% prior best), with the largest gains on the harder slices (+42% relative on Medium, +48% relative on Hard).

Rui Yang, Qianhui Wu, Yuxi Chen et al.

Building capable visual web agents requires long-horizon reasoning, precise grounding, and robust interaction with dynamic real-world websites. Despite rapid progress, the strongest systems remain largely proprietary, while open agents still depend heavily on supervised post-training over large collections of curated web trajectories. This dependence creates a major scalability bottleneck: high-quality demonstrations are expensive to collect, and static datasets offer limited coverage of the diverse, ever-changing open web. Although online RL has shown promise for text-based agents, its potential for training visual web agents directly on live websites remains largely underexplored. In this paper, we introduce OpenWebRL, an open framework for training visual web agents with online multi-turn RL on real websites. OpenWebRL covers the full training pipeline, including scalable live-browser infrastructure, supervised initialization, multimodal context management, trajectory-level success judging, and efficient multi-turn policy optimization. Using this framework, we train OpenWebRL-4B, which establishes a new open-source state of the art on challenging live-web benchmarks. With only 0.4K initialization trajectories and 2.2K open-ended RL training tasks, OpenWebRL-4B achieves 67.0% success on Online-Mind2Web and 64.0% on DeepShop, outperforming prior open agents of similar or larger scale and remaining competitive with proprietary systems including OpenAI CUA and Gemini CUA. Beyond strong benchmark performance, we systematically study the key design choices that make online RL effective for visual web agents, and analyze how RL improves agentic reasoning. Overall, our work offers a practical path toward building more capable, reproducible, and cost-efficient open web agents. We will release our training data, models, and code to support future research.

Hantao Zhou, Rui Yang, Yachao Zhang et al. · tencent-ai

The detection head constitutes a pivotal component within object detectors, tasked with executing both classification and localization functions. Regrettably, the commonly used parallel head often lacks omni perceptual capabilities, such as deformation perception, global perception and cross-task perception. Despite numerous methods attempting to enhance these abilities from a single aspect, achieving a comprehensive and unified solution remains a significant challenge. In response to this challenge, we develop an innovative detection head, termed UniHead, to unify three perceptual abilities simultaneously. More precisely, our approach (1) introduces deformation perception, enabling the model to adaptively sample object features; (2) proposes a Dual-axial Aggregation Transformer (DAT) to adeptly model long-range dependencies, thereby achieving global perception; and (3) devises a Cross-task Interaction Transformer (CIT) that facilitates interaction between the classification and localization branches, thus aligning the two tasks. As a plug-and-play method, the proposed UniHead can be conveniently integrated with existing detectors. Extensive experiments on the COCO dataset demonstrate that our UniHead can bring significant improvements to many detectors. For instance, the UniHead can obtain +2.7 AP gains in RetinaNet, +2.9 AP gains in FreeAnchor, and +2.1 AP gains in GFL. The code is available at https://github.com/zht8506/UniHead.

Shali Jiang, Hua Zheng, Boyang Liu et al.

Knowledge distillation (KD) transfers a single scalar prediction from a large foundation model (FM) to compact vertical models (VMs), suffering from diminishing transfer ratio -- the fraction of FM improvement captured by the VM -- as a single scalar cannot convey the rich intermediate knowledge that larger FMs learn. To address this bottleneck, we propose LoopFM (Learning frOm HistOrical ReP*resentations of FM), a framework that opens a high-bandwidth transfer channel by structuring FM intermediate embeddings as input features (e.g., user history sequence) for downstream VMs, without requiring real-time FM inference at serving and architectural coupling between FM and VM. We provide a theoretical framework for LoopFM with a gain decomposition and transfer-ratio analysis. On three public benchmarks, LoopFM demonstrates strong AUC improvements (e.g., 6\%+ on TaobaoAd) and complementary knowledge transfer capability with KD. On industrial-scale systems (billions of examples, trillion-parameter FMs), LoopFM approximately doubles the knowledge transfer ratio on top of KD, delivering a +0.5\% conversion improvement in Y1H1, and a +1.03\% and +1.22\% conversion improvement from two individual launches respectively in Y1H2.

Xinyi Chen, Elad Hazan, Tongyang Li et al. · princeton

The problem of efficient quantum state learning, also called shadow tomography, aims to comprehend an unknown $d$-dimensional quantum state through POVMs. Yet, these states are rarely static; they evolve due to factors such as measurements, environmental noise, or inherent Hamiltonian state transitions. This paper leverages techniques from adaptive online learning to keep pace with such state changes. The key metrics considered for learning in these mutable environments are enhanced notions of regret, specifically adaptive and dynamic regret. We present adaptive and dynamic regret bounds for online shadow tomography, which are polynomial in the number of qubits and sublinear in the number of measurements. To support our theoretical findings, we include numerical experiments that validate our proposed models.

Hao Sun, Lei Han, Rui Yang et al. · cambridge

In this work, we study the simple yet universally applicable case of reward shaping in value-based Deep Reinforcement Learning (DRL). We show that reward shifting in the form of the linear transformation is equivalent to changing the initialization of the $Q$-function in function approximation. Based on such an equivalence, we bring the key insight that a positive reward shifting leads to conservative exploitation, while a negative reward shifting leads to curiosity-driven exploration. Accordingly, conservative exploitation improves offline RL value estimation, and optimistic value estimation improves exploration for online RL. We validate our insight on a range of RL tasks and show its improvement over baselines: (1) In offline RL, the conservative exploitation leads to improved performance based on off-the-shelf algorithms; (2) In online continuous control, multiple value functions with different shifting constants can be used to tackle the exploration-exploitation dilemma for better sample efficiency; (3) In discrete control tasks, a negative reward shifting yields an improvement over the curiosity-based exploration method.

Rui Yang, Lin Song, Yixiao Ge et al. · tencent-ai

Box-supervised instance segmentation has gained much attention as it requires only simple box annotations instead of costly mask or polygon annotations. However, existing box-supervised instance segmentation models mainly focus on mask-based frameworks. We propose a new end-to-end training technique, termed BoxSnake, to achieve effective polygonal instance segmentation using only box annotations for the first time. Our method consists of two loss functions: (1) a point-based unary loss that constrains the bounding box of predicted polygons to achieve coarse-grained segmentation; and (2) a distance-aware pairwise loss that encourages the predicted polygons to fit the object boundaries. Compared with the mask-based weakly-supervised methods, BoxSnake further reduces the performance gap between the predicted segmentation and the bounding box, and shows significant superiority on the Cityscapes dataset. The code has been available publicly.

Weijian Huang, Hao Yang, Cheng Li et al.

Multi-modal representation methods have achieved advanced performance in medical applications by extracting more robust features from multi-domain data. However, existing methods usually need to train additional branches for downstream tasks, which may increase the model complexities in clinical applications as well as introduce additional human inductive bias. Besides, very few studies exploit the rich clinical knowledge embedded in clinical daily reports. To this end, we propose a novel medical generalist agent, MGA, that can address three kinds of common clinical tasks via clinical reports knowledge transformation. Unlike the existing methods, MGA can easily adapt to different tasks without specific downstream branches when their corresponding annotations are missing. More importantly, we are the first attempt to use medical professional language guidance as a transmission medium to guide the agent's behavior. The proposed method is implemented on four well-known X-ray open-source datasets, MIMIC-CXR, CheXpert, MIMIC-CXR-JPG, and MIMIC-CXR-MS. Promising results are obtained, which validate the effectiveness of our proposed MGA. Code is available at: https://github.com/SZUHvern/MGA

Jiawei Xu, Rui Yang, Shuang Qiu et al. · tencent-ai, tsinghua

Learning policy from offline datasets through offline reinforcement learning (RL) holds promise for scaling data-driven decision-making while avoiding unsafe and costly online interactions. However, real-world data collected from sensors or humans often contains noise and errors, posing a significant challenge for existing offline RL methods, particularly when the real-world data is limited. Our study reveals that prior research focusing on adapting predominant offline RL methods based on temporal difference learning still falls short under data corruption when the dataset is limited. In contrast, we discover that vanilla sequence modeling methods, such as Decision Transformer, exhibit robustness against data corruption, even without specialized modifications. To unlock the full potential of sequence modeling, we propose Robust Decision Rransformer (RDT) by incorporating three simple yet effective robust techniques: embedding dropout to improve the model's robustness against erroneous inputs, Gaussian weighted learning to mitigate the effects of corrupted labels, and iterative data correction to eliminate corrupted data from the source. Extensive experiments on MuJoCo, Kitchen, and Adroit tasks demonstrate RDT's superior performance under various data corruption scenarios compared to prior methods. Furthermore, RDT exhibits remarkable robustness in a more challenging setting that combines training-time data corruption with test-time observation perturbations. These results highlight the potential of sequence modeling for learning from noisy or corrupted offline datasets, thereby promoting the reliable application of offline RL in real-world scenarios. Our code is available at https://github.com/jiawei415/RobustDecisionTransformer.

Sheng Jin, Minghao Liu, Yunze Xiao et al.

Knowledge benchmarks for LLMs face three issues: scaling-driven designs that do not operationalize disciplinary representativeness; flat-payment annotation that permits lazy consensus; and unaudited ranking instability under bounded test budgets. We introduce KINA, an 899-item benchmark across 261 fine-grained disciplines, with two formal results. First, we cast representativeness as a coverage-style objective over expert-elicited anchors and operationalize disciplinary representativeness through a proxy, yielding a (1-1/e) greedy approximation (Proposition 1); the guarantee applies to the proxy, not to population representativeness. Second, we prove a bonus-on-bar tournament weakly FOSD-dominates flat payment in released-review quality, with incentive-compatibility threshold B > Delta C / Delta p_min (Theorem 1). Evaluating 42 models from 13 labs, the top model, Gemini-3.1-Pro-Preview, reaches 53.17%, followed by Claude-Opus-4.6 at 49.92% and GPT-5.4 at 48.55%, leaving substantial headroom below saturation. The full leaderboard shows a tiered structure rather than a smooth total order: a small frontier tier lies above 48%, a dense strong-model tier spans roughly 38-45%, and low-performing models remain only modestly above the 10% chance baseline. Tool augmentation adds up to 5.17 points across the five tool-use evaluations, with gains varying substantially across models. We report bootstrap ranking-stability statistics to make bounded-budget variance explicit and to discourage over-interpretation of adjacent ranks.

Yifei He, Rui Yang, Hao Bai et al.

Computer use agents (CUAs) have shown strong potential for automating complex digital workflows, yet their training remains constrained by costly live environment interaction and limited high-quality supervision. Existing filtered behavior cloning pipelines suffer from imitation bottlenecks, including distribution shift from the expert demonstration and the absence of negative learning signals. Meanwhile, standard trajectory-level reinforcement learning struggles with sparse rewards, ambiguous credit assignment, and high infrastructure costs for long-horizon GUI interaction. In this work, we propose PRO-CUA, a process-reward optimization framework for training CUAs with iterative step-level reinforcement learning. PRO-CUA decouples on-policy environment interaction from policy optimization: the current policy collects states through live rollouts, generates diverse candidate actions for each state, receives step-level feedback from a process reward model (PRM), and is optimized with group-relative advantages. This design enables dense and flexible credit assignment without relying on golden answers or offline expert trajectories, while reducing distribution shift by training on the agent's own execution states. Experiments on live web benchmarks demonstrate the effectiveness of PRO-CUA and the reliability of PRM-guided step-level training.

Rui Yang, Qingcheng Zeng, Keen You et al.

This study introduces Ascle, a pioneering natural language processing (NLP) toolkit designed for medical text generation. Ascle is tailored for biomedical researchers and healthcare professionals with an easy-to-use, all-in-one solution that requires minimal programming expertise. For the first time, Ascle evaluates and provides interfaces for the latest pre-trained language models, encompassing four advanced and challenging generative functions: question-answering, text summarization, text simplification, and machine translation. In addition, Ascle integrates 12 essential NLP functions, along with query and search capabilities for clinical databases. The toolkit, its models, and associated data are publicly available via https://github.com/Yale-LILY/MedGen.

Boka Qin, Rui Yang

We introduce the Polymarket-v1 Database: the complete on-chain trade archive of Polymarket's first-generation CTF Exchange on Polygon, spanning 2022-11-21 to 2026-04-28 and covering the full contract lifecycle from first settlement to natural termination. The dataset comprises 1.20 billion trade records across 1.30 million markets with $61 billion in nominal volume. Its defining feature is 100% ground-truth aggressor direction derived from the blockchain settlement layer, a property unavailable in existing prediction market archives, which rely on heuristic inference. We use this truth-aligned archive to benchmark standard microstructure tools and document three findings. First, the tick rule and bulk volume classification achieve near-random aggregate accuracy (49.83% and 50.51%), but this masks a systematic, correctable price-level gradient driven by positive trade direction autocorrelation and concentrated market-making -- two structural features of prediction markets that violate the mean-reversion assumption embedded in classical classifiers. Second, these classification errors propagate into downstream metrics: inferred VPIN diverges substantially from ground-truth VPIN, and OFI estimates are directionally biased, with material consequences for Transaction Cost Analysis. Third, ground-truth microstructure quality predicts forecasting performance in ways that classification-based proxies cannot recover: True VPIN positively predicts Brier scores, while Gibbs spread negatively predicts them -- a selection effect reflecting that high-spread niche markets attract informed specialists rather than noise traders. Replacing ground-truth metrics with classified proxies attenuates both relationships, illustrating that measurement accuracy at the transaction level is a prerequisite for reliable inference about prediction market design and probability calibration.

Chengxuan Qin, Rui Yang, Wenlong You et al.

The increasing number of dispersed EEG dataset publications and the advancement of large-scale Electroencephalogram (EEG) models have increased the demand for practical tools to manage diverse EEG datasets. However, the inherent complexity of EEG data, characterized by variability in content data, metadata, and data formats, poses challenges for integrating multiple datasets and conducting large-scale EEG model research. To tackle the challenges, this paper introduces EEGUnity, an open-source tool that incorporates modules of 'EEG Parser', 'Correction', 'Batch Processing', and 'Large Language Model Boost'. Leveraging the functionality of such modules, EEGUnity facilitates the efficient management of multiple EEG datasets, such as intelligent data structure inference, data cleaning, and data unification. In addition, the capabilities of EEGUnity ensure high data quality and consistency, providing a reliable foundation for large-scale EEG data research. EEGUnity is evaluated across 25 EEG datasets from different sources, offering several typical batch processing workflows. The results demonstrate the high performance and flexibility of EEGUnity in parsing and data processing. The project code is publicly available at github.com/Baizhige/EEGUnity.

Haonan Han, Rui Yang, Shuyan Li et al.

Interactive devices with touch screen have become commonly used in various aspects of daily life, which raises the demand for high production quality of touch screen glass. While it is desirable to develop effective defect detection technologies to optimize the automatic touch screen production lines, the development of these technologies suffers from the lack of publicly available datasets. To address this issue, we in this paper propose a dedicated touch screen glass defect dataset which includes seven types of defects and consists of 2504 images captured in various scenarios.All data are captured with professional acquisition equipment on the fixed workstation. Additionally, we benchmark the CNN- and Transformer-based object detection frameworks on the proposed dataset to demonstrate the challenges of defect detection on high-resolution images. Dataset and related code will be available at https://github.com/Yangr116/SSGDataset.

Fan Gao, Hang Jiang, Rui Yang et al. · mit

Educational materials such as survey articles in specialized fields like computer science traditionally require tremendous expert inputs and are therefore expensive to create and update. Recently, Large Language Models (LLMs) have achieved significant success across various general tasks. However, their effectiveness and limitations in the education domain are yet to be fully explored. In this work, we examine the proficiency of LLMs in generating succinct survey articles specific to the niche field of NLP in computer science, focusing on a curated list of 99 topics. Automated benchmarks reveal that GPT-4 surpasses its predecessors, inluding GPT-3.5, PaLM2, and LLaMa2 by margins ranging from 2% to 20% in comparison to the established ground truth. We compare both human and GPT-based evaluation scores and provide in-depth analysis. While our findings suggest that GPT-created surveys are more contemporary and accessible than human-authored ones, certain limitations were observed. Notably, GPT-4, despite often delivering outstanding content, occasionally exhibited lapses like missing details or factual errors. At last, we compared the rating behavior between humans and GPT-4 and found systematic bias in using GPT evaluation.

Rui Yang, Chenjia Bai, Xiaoteng Ma et al.

Offline reinforcement learning (RL) provides a promising direction to exploit massive amount of offline data for complex decision-making tasks. Due to the distribution shift issue, current offline RL algorithms are generally designed to be conservative in value estimation and action selection. However, such conservatism can impair the robustness of learned policies when encountering observation deviation under realistic conditions, such as sensor errors and adversarial attacks. To trade off robustness and conservatism, we propose Robust Offline Reinforcement Learning (RORL) with a novel conservative smoothing technique. In RORL, we explicitly introduce regularization on the policy and the value function for states near the dataset, as well as additional conservative value estimation on these states. Theoretically, we show RORL enjoys a tighter suboptimality bound than recent theoretical results in linear MDPs. We demonstrate that RORL can achieve state-of-the-art performance on the general offline RL benchmark and is considerably robust to adversarial observation perturbations.

Keming Wu, Man Yao, Yuhong Chou et al.

Spiking Neural Networks (SNNs) have received widespread attention due to their unique neuronal dynamics and low-power nature. Previous research empirically shows that SNNs with Poisson coding are more robust than Artificial Neural Networks (ANNs) on small-scale datasets. However, it is still unclear in theory how the adversarial robustness of SNNs is derived, and whether SNNs can still maintain its adversarial robustness advantage on large-scale dataset tasks. This work theoretically demonstrates that SNN's inherent adversarial robustness stems from its Poisson coding. We reveal the conceptual equivalence of Poisson coding and randomized smoothing in defense strategies, and analyze in depth the trade-off between accuracy and adversarial robustness in SNNs via the proposed Randomized Smoothing Coding (RSC) method. Experiments demonstrate that the proposed RSC-SNNs show remarkable adversarial robustness, surpassing ANNs and achieving state-of-the-art robustness results on large-scale dataset ImageNet. Our open-source implementation code is available at this https URL: https://github.com/KemingWu/RSC-SNN.

Yimin Ou, Rui Yang, Lufan Ma et al.

Existing instance segmentation methods have achieved impressive performance but still suffer from a common dilemma: redundant representations (e.g., multiple boxes, grids, and anchor points) are inferred for one instance, which leads to multiple duplicated predictions. Thus, mainstream methods usually rely on a hand-designed non-maximum suppression (NMS) post-processing step to select the optimal prediction result, which hinders end-to-end training. To address this issue, we propose a box-free and NMS-free end-to-end instance segmentation framework, termed UniInst, that yields only one unique representation for each instance. Specifically, we design an instance-aware one-to-one assignment scheme, namely Only Yield One Representation (OYOR), which dynamically assigns one unique representation to each instance according to the matching quality between predictions and ground truths. Then, a novel prediction re-ranking strategy is elegantly integrated into the framework to address the misalignment between the classification score and the mask quality, enabling the learned representation to be more discriminative. With these techniques, our UniInst, the first FCN-based box-free and NMS-free instance segmentation framework, achieves competitive performance, e.g., 39.0 mask AP using ResNet-50-FPN and 40.2 mask AP using ResNet-101-FPN, against mainstream methods on COCO test-dev 2017. Moreover, the proposed instance-aware method is robust to occlusion scenes, outperforming common baselines by remarkable mask AP on the heavily-occluded OCHuman benchmark. Code is available at https://github.com/b03505036/UniInst.

Rui Yang, Jie Wang, Zijie Geng et al.

Generalization across different environments with the same tasks is critical for successful applications of visual reinforcement learning (RL) in real scenarios. However, visual distractions -- which are common in real scenes -- from high-dimensional observations can be hurtful to the learned representations in visual RL, thus degrading the performance of generalization. To tackle this problem, we propose a novel approach, namely Characteristic Reward Sequence Prediction (CRESP), to extract the task-relevant information by learning reward sequence distributions (RSDs), as the reward signals are task-relevant in RL and invariant to visual distractions. Specifically, to effectively capture the task-relevant information via RSDs, CRESP introduces an auxiliary task -- that is, predicting the characteristic functions of RSDs -- to learn task-relevant representations, because we can well approximate the high-dimensional distributions by leveraging the corresponding characteristic functions. Experiments demonstrate that CRESP significantly improves the performance of generalization on unseen environments, outperforming several state-of-the-arts on DeepMind Control tasks with different visual distractions.

Tao Yan, Rui Yang, Ziyang Zheng et al.

Photonic neural networks perform brain-inspired computations using photons instead of electrons that can achieve substantially improved computing performance. However, existing architectures can only handle data with regular structures, e.g., images or videos, but fail to generalize to graph-structured data beyond Euclidean space, e.g., social networks or document co-citation networks. Here, we propose an all-optical graph representation learning architecture, termed diffractive graph neural network (DGNN), based on the integrated diffractive photonic computing units (DPUs) to address this limitation. Specifically, DGNN optically encodes node attributes into strip optical waveguides, which are transformed by DPUs and aggregated by on-chip optical couplers to extract their feature representations. Each DPU comprises successive passive layers of metalines to modulate the electromagnetic optical field via diffraction, where the metaline structures are learnable parameters shared across graph nodes. DGNN captures complex dependencies among the node neighborhoods and eliminates the nonlinear transition functions during the light-speed optical message passing over graph structures. We demonstrate the use of DGNN extracted features for node and graph-level classification tasks with benchmark databases and achieve superior performance. Our work opens up a new direction for designing application-specific integrated photonic circuits for high-efficiency processing of large-scale graph data structures using deep learning.

Peiyu Zhuang, Haodong Li, Rui Yang et al.

With the spread of tampered images, locating the tampered regions in digital images has drawn increasing attention. The existing image tampering localization methods, however, suffer from severe performance degradation when the tampered images are subjected to some post-processing, as the tampering traces would be distorted by the post-processing operations. The poor robustness against post-processing has become a bottleneck for the practical applications of image tampering localization techniques. In order to address this issue, this paper proposes a novel restoration-assisted framework for image tampering localization (ReLoc). The ReLoc framework mainly consists of an image restoration module and a tampering localization module. The key idea of ReLoc is to use the restoration module to recover a high-quality counterpart of the distorted tampered image, such that the distorted tampering traces can be re-enhanced, facilitating the tampering localization module to identify the tampered regions. To achieve this, the restoration module is optimized not only with the conventional constraints on image visual quality but also with a forensics-oriented objective function. Furthermore, the restoration module and the localization module are trained alternately, which can stabilize the training process and is beneficial for improving the performance. The proposed framework is evaluated by fighting against JPEG compression, the most commonly used post-processing. Extensive experimental results show that ReLoc can significantly improve the robustness against JPEG compression. The restoration module in a well-trained ReLoc model is transferable. Namely, it is still effective when being directly deployed with another tampering localization module.

Tencent Hunyuan Team, Ao Liu, Botong Zhou et al. · tencent-ai

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Zunhai Su, Hengyuan Zhang, Wei Wu et al.

As the foundational architecture of modern machine learning, Transformers have driven remarkable progress across diverse AI domains. Despite their transformative impact, a persistent challenge across various Transformers is Attention Sink (AS), in which a disproportionate amount of attention is focused on a small subset of specific yet uninformative tokens. AS complicates interpretability, significantly affecting the training and inference dynamics, and exacerbates issues such as hallucinations. In recent years, substantial research has been dedicated to understanding and harnessing AS. However, a comprehensive survey that systematically consolidates AS-related research and offers guidance for future advancements remains lacking. To address this gap, we present the first survey on AS, structured around three key dimensions that define the current research landscape: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Our work provides a pivotal contribution by clarifying key concepts and guiding researchers through the evolution and trends of the field. We envision this survey as a definitive resource, empowering researchers and practitioners to effectively manage AS within the current Transformer paradigm, while simultaneously inspiring innovative advancements for the next generation of Transformers. The paper list of this work is available at https://github.com/ZunhaiSu/Awesome-Attention-Sink.

Ziyang Zheng, Zhengyang Duan, Hang Chen et al.

Photonic neural network (PNN) is a remarkable analog artificial intelligence (AI) accelerator that computes with photons instead of electrons to feature low latency, high energy efficiency, and high parallelism. However, the existing training approaches cannot address the extensive accumulation of systematic errors in large-scale PNNs, resulting in a significant decrease in model performance in physical systems. Here, we propose dual adaptive training (DAT) that allows the PNN model to adapt to substantial systematic errors and preserves its performance during the deployment. By introducing the systematic error prediction networks with task-similarity joint optimization, DAT achieves the high similarity mapping between the PNN numerical models and physical systems and high-accurate gradient calculations during the dual backpropagation training. We validated the effectiveness of DAT by using diffractive PNNs and interference-based PNNs on image classification tasks. DAT successfully trained large-scale PNNs under major systematic errors and preserved the model classification accuracies comparable to error-free systems. The results further demonstrated its superior performance over the state-of-the-art in situ training approaches. DAT provides critical support for constructing large-scale PNNs to achieve advanced architectures and can be generalized to other types of AI systems with analog computing errors.

Qiyuan Liu, Qi Zhou, Rui Yang et al.

Recent work has shown that representation learning plays a critical role in sample-efficient reinforcement learning (RL) from pixels. Unfortunately, in real-world scenarios, representation learning is usually fragile to task-irrelevant distractions such as variations in background or viewpoint. To tackle this problem, we propose a novel clustering-based approach, namely Clustering with Bisimulation Metrics (CBM), which learns robust representations by grouping visual observations in the latent space. Specifically, CBM alternates between two steps: (1) grouping observations by measuring their bisimulation distances to the learned prototypes; (2) learning a set of prototypes according to the current cluster assignments. Computing cluster assignments with bisimulation metrics enables CBM to capture task-relevant information, as bisimulation metrics quantify the behavioral similarity between observations. Moreover, CBM encourages the consistency of representations within each group, which facilitates filtering out task-irrelevant information and thus induces robust representations against distractions. An appealing feature is that CBM can achieve sample-efficient representation learning even if multiple distractions exist simultaneously.Experiments demonstrate that CBM significantly improves the sample efficiency of popular visual RL algorithms and achieves state-of-the-art performance on both multiple and single distraction settings. The code is available at https://github.com/MIRALab-USTC/RL-CBM.

Mingxuan Ye, Yufei Kuang, Jie Wang et al.

While deep reinforcement learning (RL) has been demonstrated effective in solving complex control tasks, sample efficiency remains a key challenge due to the large amounts of data required for remarkable performance. Existing research explores the application of representation learning for data-efficient RL, e.g., learning predictive representations by predicting long-term future states. However, many existing methods do not fully exploit the structural information inherent in sequential state signals, which can potentially improve the quality of long-term decision-making but is difficult to discern in the time domain. To tackle this problem, we propose State Sequences Prediction via Fourier Transform (SPF), a novel method that exploits the frequency domain of state sequences to extract the underlying patterns in time series data for learning expressive representations efficiently. Specifically, we theoretically analyze the existence of structural information in state sequences, which is closely related to policy performance and signal regularity, and then propose to predict the Fourier transform of infinite-step future state sequences to extract such information. One of the appealing features of SPF is that it is simple to implement while not requiring storage of infinite-step future states as prediction targets. Experiments demonstrate that the proposed method outperforms several state-of-the-art algorithms in terms of both sample efficiency and performance.

Alphan Sahin, Rui Yang

Communication and computation are often viewed as separate tasks. This approach is very effective from the perspective of engineering as isolated optimizations can be performed. However, for many computation-oriented applications, the main interest is a function of the local information at the devices, rather than the local information itself. In such scenarios, information theoretical results show that harnessing the interference in a multiple access channel for computation, i.e., over-the-air computation (OAC), can provide a significantly higher achievable computation rate than separating communication and computation tasks. Moreover, the gap between OAC and separation in terms of computation rate increases with more participating nodes. Given this motivation, in this study, we provide a comprehensive survey on practical OAC methods. After outlining fundamentals related to OAC, we discuss the available OAC schemes with their pros and cons. We provide an overview of the enabling mechanisms for achieving reliable computation in the wireless channel. Finally, we summarize the potential applications of OAC and point out some future directions.

Rui Yang, Hailong Ma, Jie Wu et al.

The vanilla self-attention mechanism inherently relies on pre-defined and steadfast computational dimensions. Such inflexibility restricts it from possessing context-oriented generalization that can bring more contextual cues and global representations. To mitigate this issue, we propose a Scalable Self-Attention (SSA) mechanism that leverages two scaling factors to release dimensions of query, key, and value matrices while unbinding them with the input. This scalability fetches context-oriented generalization and enhances object sensitivity, which pushes the whole network into a more effective trade-off state between accuracy and cost. Furthermore, we propose an Interactive Window-based Self-Attention (IWSA), which establishes interaction between non-overlapping regions by re-merging independent value tokens and aggregating spatial information from adjacent windows. By stacking the SSA and IWSA alternately, the Scalable Vision Transformer (ScalableViT) achieves state-of-the-art performance in general-purpose vision tasks. For example, ScalableViT-S outperforms Twins-SVT-S by 1.4% and Swin-T by 1.8% on ImageNet-1K classification.

Jie Wang, Rui Yang, Zijie Geng et al.

Generalization in partially observed markov decision processes (POMDPs) is critical for successful applications of visual reinforcement learning (VRL) in real scenarios. A widely used idea is to learn task-relevant representations that encode task-relevant information of common features in POMDPs, i.e., rewards and transition dynamics. As transition dynamics in the latent state space -- which are task-relevant and invariant to visual distractions -- are unknown to the agents, existing methods alternatively use transition dynamics in the observation space to extract task-relevant information in transition dynamics. However, such transition dynamics in the observation space involve task-irrelevant visual distractions, degrading the generalization performance of VRL methods. To tackle this problem, we propose the reward sequence distribution conditioned on the starting observation and the predefined subsequent action sequence (RSD-OA). The appealing features of RSD-OA include that: (1) RSD-OA is invariant to visual distractions, as it is conditioned on the predefined subsequent action sequence without task-irrelevant information from transition dynamics, and (2) the reward sequence captures long-term task-relevant information in both rewards and transition dynamics. Experiments demonstrate that our representation learning approach based on RSD-OA significantly improves the generalization performance on unseen environments, outperforming several state-of-the-arts on DeepMind Control tasks with visual distractions.

Samuel Bosch, Bobak Kiani, Rui Yang et al.

Quantum machine learning has the potential to enable advances in artificial intelligence, such as solving problems intractable on classical computers. Some fundamental ideas behind quantum machine learning are similar to kernel methods in classical machine learning. Both process information by mapping it into high-dimensional vector spaces without explicitly calculating their numerical values. We explore a setup for performing classification on labeled classical datasets, consisting of a classical neural network connected to a quantum annealer. The neural network programs the quantum annealer's controls and thereby maps the annealer's initial states into new states in the Hilbert space. The neural network's parameters are optimized to maximize the distance of states corresponding to inputs from different classes and minimize the distance between quantum states corresponding to the same class. Recent literature showed that at least some of the "learning" is due to the quantum annealer, connecting a small linear network to a quantum annealer and using it to learn small and linearly inseparable datasets. In this study, we consider a similar but not quite the same case, where a classical fully-fledged neural network is connected with a small quantum annealer. In such a setting, the fully-fledged classical neural-network already has built-in nonlinearity and learning power, and can already handle the classification problem alone, we want to see whether an additional quantum layer could boost its performance. We simulate this system to learn several common datasets, including those for image and sound recognition. We conclude that adding a small quantum annealer does not provide a significant benefit over just using a regular (nonlinear) classical neural network.

Zunhai Su, Rui Yang, Chao Zhang et al.

The rapid advancement toward long-context reasoning and multi-modal intelligence has made the memory footprint of the Key-Value (KV) cache a dominant memory bottleneck for efficient deployment. While the established per-channel quantization effectively accommodates intrinsic channel-wise outliers in Key tensors, its efficacy diminishes under extreme compression. In this work, we revisit the inherent limitations of the per-channel quantization paradigm from both empirical and theoretical perspectives. Our analysis identifies Token Norm Imbalance (TNI) as the primary bottleneck to quantization fidelity. We demonstrate that TNI systematically amplifies errors when shared quantization parameters are required to span token groups exhibiting substantial norm disparities. Instead of relying on intricate quantization pipelines (e.g., TurboQuant), we propose OScaR (Omni-Scaled Canalized Rotation), an accurate and lightweight KV cache compression framework for X-LLMs (i.e., text-only, multi-modal, and omni-modal LLMs). Advancing the per-channel paradigm, OScaR employs Canalized Rotation followed by Omni-Token Scaling to mitigate TNI-induced sequence-dimensional variance both effectively and efficiently, further supported by our optimized system design and CUDA kernels. Extensive evaluations across X-LLMs show that OScaR consistently outperforms existing methods and achieves near-lossless performance under INT2 quantization, establishing it as a robust, low-complexity, and universal framework that defines a new Pareto front. Compared with the BF16 FlashDecoding-v2 baseline, our OScaR implementation achieves a notable up to 3.0x speedup in decoding, reduces memory footprint by 5.3x, and increases throughput by 4.1x. The code for OScaR is publicly available at https://github.com/ZunhaiSu/OScaR-KV-Quant.

Rui Yang, Qianhui Wu, Zhaoyang Wang et al.

Open-source native GUI agents still lag behind closed-source systems on long-horizon navigation tasks. This gap stems from two limitations: a shortage of high-quality, action-aligned reasoning data, and the direct adoption of generic post-training pipelines that overlook the unique challenges of GUI agents. We identify two fundamental issues in these pipelines: (i) standard SFT with CoT reasoning often hurts grounding, and (ii) step-wise RLVR-tyle training faces partial verifiability, where multiple actions can be correct but only a single demonstrated action is used for verification. This makes offline step-wise metrics weak predictors of online task success. In this work, we present GUI-Libra, a tailored training recipe that addresses these challenges. First, to mitigate the scarcity of action-aligned reasoning data, we introduce a data construction and filtering pipeline and release a curated 81K GUI reasoning dataset. Second, to reconcile reasoning with grounding, we propose action-aware SFT that mixes reasoning-then-action and direct-action data and reweights tokens to emphasize action and grounding. Third, to stabilize RL under partial verifiability, we identify the overlooked importance of KL regularization in RLVR and show that a KL trust region is critical for improving offline-to-online predictability; we further introduce success-adaptive scaling to downweight unreliable negative gradients. Across diverse web and mobile benchmarks, GUI-Libra consistently improves both step-wise accuracy and end-to-end task completion. Our results suggest that carefully designed post-training and data curation can unlock significantly stronger task-solving capabilities without costly online data collection. We release our dataset, code, and models to facilitate further research on data-efficient post-training for reasoning-capable GUI agents.

Rui Yang, Jiahao Zhu, Jianping Man et al.

The design and development of text-based knowledge graph completion (KGC) methods leveraging textual entity descriptions are at the forefront of research. These methods involve advanced optimization techniques such as soft prompts and contrastive learning to enhance KGC models. The effectiveness of text-based methods largely hinges on the quality and richness of the training data. Large language models (LLMs) can utilize straightforward prompts to alter text data, thereby enabling data augmentation for KGC. Nevertheless, LLMs typically demand substantial computational resources. To address these issues, we introduce a framework termed constrained prompts for KGC (CP-KGC). This CP-KGC framework designs prompts that adapt to different datasets to enhance semantic richness. Additionally, CP-KGC employs a context constraint strategy to effectively identify polysemous entities within KGC datasets. Through extensive experimentation, we have verified the effectiveness of this framework. Even after quantization, the LLM (Qwen-7B-Chat-int4) still enhances the performance of text-based KGC methods \footnote{Code and datasets are available at \href{https://github.com/sjlmg/CP-KGC}{https://github.com/sjlmg/CP-KGC}}. This study extends the performance limits of existing models and promotes further integration of KGC with LLMs.

Baolin Peng, Wenlin Yao, Qianhui Wu et al.

Agentic modeling aims to transform LLMs into autonomous agents capable of solving complex tasks through planning, reasoning, tool use, and multi-turn interaction with environments. Despite major investment, open research remains constrained by infrastructure and training gaps. Many high-performing systems rely on proprietary codebases, models, or services, while most open-source frameworks focus on orchestration and evaluation rather than scalable agent training. We present Orchard, an open-source framework for scalable agentic modeling. At its core is Orchard Env, a lightweight environment service providing reusable primitives for sandbox lifecycle management across task domains, agent harnesses, and pipeline stages. On top of Orchard Env, we build three agentic modeling recipes. Orchard-SWE targets coding agents. We distill 107K trajectories from MiniMax-M2.5 and Qwen3.5-397B, introduce credit-assignment SFT to learn from productive segments of unresolved trajectories, and apply Balanced Adaptive Rollout for RL. Starting from Qwen3-30B-A3B-Thinking, Orchard-SWE achieves 64.3% on SWE-bench Verified after SFT and 67.5% after SFT+RL, setting a new state of the art among open-source models of comparable size. Orchard-GUI trains a 4B vision-language computer-use agent using only 0.4K distilled trajectories and 2.2K open-ended tasks. It achieves 74.1%, 67.0%, and 64.0% success rates on WebVoyager, Online-Mind2Web, and DeepShop, respectively, making it the strongest open-source model while remaining competitive with proprietary systems. Orchard-Claw targets personal assistant agents. Trained with only 0.2K synthetic tasks, it achieves 59.6% pass@3 on Claw-Eval and 73.9% when paired with a stronger ZeroClaw harness. Collectively, these results show that a lightweight, open, harness-agnostic environment layer enables reusable agentic data, training recipes, and evaluations across domains.

Yiran Qiao, Yiren Lu, Yunlai Zhou et al.

3D Gaussian Splatting (3DGS) is increasingly recognized as a powerful paradigm for real-time, high-fidelity 3D reconstruction. However, its per-scene optimization pipeline limits scalability and generalization, and prevents efficient inference. Recently emerged feed-forward 3DGS models address these limitations by enabling fast reconstruction from a few input views after large-scale pretraining, without scene-specific optimization. Despite their advantages and strong potential for commercial deployment, the use of neural networks as the backbone also amplifies the risk of adversarial manipulation. In this paper, we introduce AdvSplat, the first systematic study of adversarial attacks on feed-forward 3DGS. We first employ white-box attacks to reveal fundamental vulnerabilities of this model family. We then develop two improved, practically relevant, query-efficient black-box algorithms that optimize pixel-space perturbations via a frequency-domain parameterization: one based on gradient estimation and the other gradient-free, without requiring any access to model internals. Extensive experiments across multiple datasets demonstrate that AdvSplat can significantly disrupt reconstruction results by injecting imperceptible perturbations into the input images. Our findings surface an overlooked yet urgent problem in this domain, and we hope to draw the community's attention to this emerging security and robustness challenge.

Minze Chen, Zhenxiang Tao, Weitong Tang et al.

Emergency management urgently requires comprehensive knowledge while having a high possibility to go beyond individuals' cognitive scope. Therefore, artificial intelligence(AI) supported decision-making under that circumstance is of vital importance. Recent emerging large language models (LLM) provide a new direction for enhancing targeted machine intelligence. However, the utilization of LLM directly would inevitably introduce unreliable output for its inherent issue of hallucination and poor reasoning skills. In this work, we develop a system called Enhancing Emergency decision-making with Knowledge Graph and LLM (E-KELL), which provides evidence-based decision-making in various emergency stages. The study constructs a structured emergency knowledge graph and guides LLMs to reason over it via a prompt chain. In real-world evaluations, E-KELL receives scores of 9.06, 9.09, 9.03, and 9.09 in comprehensibility, accuracy, conciseness, and instructiveness from a group of emergency commanders and firefighters, demonstrating a significant improvement across various situations compared to baseline models. This work introduces a novel approach to providing reliable emergency decision support.

Liujie Zhang, Benzhe Ning, Rui Yang et al.

Reinforcement learning (RL) post-training has proven effective at unlocking reasoning, self-reflection, and tool-use capabilities in large language models. As models extend to omni-modal inputs and agentic multi-turn workflows, RL training systems face three interdependent challenges: heterogeneous data flows, operational robustness at scale, and the staleness -- throughput tradeoff. We present \textbf{Relax} (Reinforcement Engine Leveraging Agentic X-modality), an open-source RL training engine that addresses these challenges through three co-designed architectural layers. First, an \emph{omni-native architecture} builds multimodal support into the full stack -- from data preprocessing and modality-aware parallelism to inference generation -- rather than retrofitting it onto a text-centric pipeline. Second, each RL role runs as an independent, fault-isolated service that can be scaled, recovered, and upgraded without global coordination. Third, service-level decoupling enables asynchronous training via the TransferQueue data bus, where a single staleness parameter smoothly interpolates among on-policy, near-on-policy, and fully asynchronous execution. Relax achieves a 1.20$\times$ end-to-end speedup over veRL on Qwen3-4B on-policy training. Its fully async mode delivers a 1.76$\times$ speedup over colocate on Qwen3-4B and a 2.00$\times$ speedup on Qwen3-Omni-30B, while all modes converge to the same reward level. Relax supports R3 (Rollout Routing Replay)~\cite{ma2025r3} for MoE models with only 1.9\% overhead, compared to 32\% degradation in veRL under the same configuration. It further demonstrates stable omni-modal RL convergence on Qwen3-Omni across image, text, and audio, sustaining over 2{,}000 steps on video without degradation. Relax is available at https://github.com/rednote-ai/Relax.

Alphan Sahin, Rui Yang

In this study, a digital over-the-air computation (OAC) scheme for achieving continuous-valued gradient aggregation is proposed. It is shown that the average of a set of real-valued parameters can be calculated approximately by using the average of the corresponding numerals, where the numerals are obtained based on a balanced number system. By using this property, the proposed scheme encodes the local gradients into a set of numerals. It then determines the positions of the activated orthogonal frequency division multiplexing (OFDM) subcarriers by using the values of the numerals. To eliminate the need for a precise sample-level time synchronization, channel estimation overhead, and power instabilities due to the channel inversion, the proposed scheme also uses a non-coherent receiver at the edge server (ES) and does not utilize a pre-equalization at the edge devices (EDs). Finally, the theoretical mean squared error (MSE) performance of the proposed scheme is derived and its performance for federated edge learning (FEEL) is demonstrated.

Rui Yang, Edison Marrese-Taylor, Yuhe Ke et al.

Large language models (LLMs) have demonstrated powerful text generation capabilities, bringing unprecedented innovation to the healthcare field. While LLMs hold immense promise for applications in healthcare, applying them to real clinical scenarios presents significant challenges, as these models may generate content that deviates from established medical facts and even exhibit potential biases. In our research, we develop an augmented LLM framework based on the Unified Medical Language System (UMLS), aiming to better serve the healthcare community. We employ LLaMa2-13b-chat and ChatGPT-3.5 as our benchmark models, and conduct automatic evaluations using the ROUGE Score and BERTScore on 104 questions from the LiveQA test set. Additionally, we establish criteria for physician-evaluation based on four dimensions: Factuality, Completeness, Readability and Relevancy. ChatGPT-3.5 is used for physician evaluation with 20 questions on the LiveQA test set. Multiple resident physicians conducted blind reviews to evaluate the generated content, and the results indicate that this framework effectively enhances the factuality, completeness, and relevance of generated content. Our research demonstrates the effectiveness of using UMLS-augmented LLMs and highlights the potential application value of LLMs in in medical question-answering.

Yuanyuan Zhang, Runwei Guan, Lingxiao Li et al.

Radar-based contactless cardiac monitoring has become a popular research direction recently, but the fine-grained electrocardiogram (ECG) signal is still hard to reconstruct from millimeter-wave radar signal. The key obstacle is to decouple the cardiac activities in the electrical domain (i.e., ECG) from that in the mechanical domain (i.e., heartbeat), and most existing research only uses pure data-driven methods to map such domain transformation as a black box. Therefore, this work first proposes a signal model for domain transformation, and then a novel deep learning framework called radarODE is designed to fuse the temporal and morphological features extracted from radar signals and generate ECG. In addition, ordinary differential equations are embedded in radarODE as a decoder to provide morphological prior, helping the convergence of the model training and improving the robustness under body movements. After being validated on the dataset, the proposed radarODE achieves better performance compared with the benchmark in terms of missed detection rate, root mean square error, Pearson correlation coefficient with the improvement of 9%, 16% and 19%, respectively. The validation results imply that radarODE is capable of recovering ECG signals from radar signals with high fidelity and can be potentially implemented in real-life scenarios.

Pei Liu, Bo Fu, Feng Ye et al.

The cancer prognosis on gigapixel Whole-Slide Images (WSIs) has always been a challenging task. To further enhance WSI visual representations, existing methods have explored image pyramids, instead of single-resolution images, in WSIs. In spite of this, they still face two major problems: high computational cost and the unnoticed semantical gap in multi-resolution feature fusion. To tackle these problems, this paper proposes to efficiently exploit WSI pyramids from a new perspective, the dual-stream network with cross-attention (DSCA). Our key idea is to utilize two sub-streams to process the WSI patches with two resolutions, where a square pooling is devised in a high-resolution stream to significantly reduce computational costs, and a cross-attention-based method is proposed to properly handle the fusion of dual-stream features. We validate our DSCA on three publicly-available datasets with a total number of 3,101 WSIs from 1,911 patients. Our experiments and ablation studies verify that (i) the proposed DSCA could outperform existing state-of-the-art methods in cancer prognosis, by an average C-Index improvement of around 4.6%; (ii) our DSCA network is more efficient in computation -- it has more learnable parameters (6.31M vs. 860.18K) but less computational costs (2.51G vs. 4.94G), compared to a typical existing multi-resolution network. (iii) the key components of DSCA, dual-stream and cross-attention, indeed contribute to our model's performance, gaining an average C-Index rise of around 2.0% while maintaining a relatively-small computational load. Our DSCA could serve as an alternative and effective tool for WSI-based cancer prognosis.

Chenlu Ye, Rui Yang, Quanquan Gu et al.

We investigate the problem of corruption robustness in offline reinforcement learning (RL) with general function approximation, where an adversary can corrupt each sample in the offline dataset, and the corruption level $ζ\geq0$ quantifies the cumulative corruption amount over $n$ episodes and $H$ steps. Our goal is to find a policy that is robust to such corruption and minimizes the suboptimality gap with respect to the optimal policy for the uncorrupted Markov decision processes (MDPs). Drawing inspiration from the uncertainty-weighting technique from the robust online RL setting \citep{he2022nearly,ye2022corruptionrobust}, we design a new uncertainty weight iteration procedure to efficiently compute on batched samples and propose a corruption-robust algorithm for offline RL. Notably, under the assumption of single policy coverage and the knowledge of $ζ$, our proposed algorithm achieves a suboptimality bound that is worsened by an additive factor of $\mathcal{O}(ζ(C(\widehat{\mathcal{F}},μ)n)^{-1})$ due to the corruption. Here $\widehat{\mathcal{F}}$ is the confidence set, and the dataset $\mathcal{Z}_n^H$, and $C(\widehat{\mathcal{F}},μ)$ is a coefficient that depends on $\widehat{\mathcal{F}}$ and the underlying data distribution $μ$. When specialized to linear MDPs, the corruption-dependent error term reduces to $\mathcal{O}(ζd n^{-1})$ with $d$ being the dimension of the feature map, which matches the existing lower bound for corrupted linear MDPs. This suggests that our analysis is tight in terms of the corruption-dependent term.

Xiaoyu Wen, Xudong Yu, Rui Yang et al.

To obtain a near-optimal policy with fewer interactions in Reinforcement Learning (RL), a promising approach involves the combination of offline RL, which enhances sample efficiency by leveraging offline datasets, and online RL, which explores informative transitions by interacting with the environment. Offline-to-Online (O2O) RL provides a paradigm for improving an offline trained agent within limited online interactions. However, due to the significant distribution shift between online experiences and offline data, most offline RL algorithms suffer from performance drops and fail to achieve stable policy improvement in O2O adaptation. To address this problem, we propose the Robust Offline-to-Online (RO2O) algorithm, designed to enhance offline policies through uncertainty and smoothness, and to mitigate the performance drop in online adaptation. Specifically, RO2O incorporates Q-ensemble for uncertainty penalty and adversarial samples for policy and value smoothness, which enable RO2O to maintain a consistent learning procedure in online adaptation without requiring special changes to the learning objective. Theoretical analyses in linear MDPs demonstrate that the uncertainty and smoothness lead to a tighter optimality bound in O2O against distribution shift. Experimental results illustrate the superiority of RO2O in facilitating stable offline-to-online learning and achieving significant improvement with limited online interactions.

Ruoyu Yao, Ruiguo Zhong, Pei Liu et al.

Large Language Models (LLMs) and Multimodal LLMs (MLLMs) have demonstrated immense potential in autonomous driving (AD) by offering human-like reasoning and open-world generalization. However, the excessive computational overhead and high inference latency of these massive models severely hinder their deployment in resource-constrained AD systems. To address this challenge, we propose a novel decision-making framework utilizing a lightweight confidence-aware language model, which bridges the gap between complex multimodal intention reasoning and efficient inference. Specifically, we design a multi-agent collaborative workflow, comprising action voting, confidence assessment, and summarization agents, to generate high-quality, confidence-annotated decision demonstrations via explicit Chain-of-Thought (CoT) reasoning. These demonstrations are then distilled into a lightweight language model featuring a dual-head architecture, enabling the joint prediction of decision probabilities and the generation of textual rationales. The distillation is realized via a confidence-aware fine-tuning strategy coupled with Retrieval Augmented Generation (RAG) to enhance the model's adaptability and data efficiency. Comprehensive closed-loop experiments on the nuPlan benchmark demonstrate that our approach achieves state-of-the-art (SOTA) success rates in both regular and long-tail scenarios while maintaining low inference latency.

Yifan Zhang, Zunhai Su, Shuhao Hu et al.

While FP8 attention has shown substantial promise in innovations like FlashAttention-3, its integration into the decoding phase of the DeepSeek Multi-head Latent Attention (MLA) architecture presents notable challenges. These challenges include numerical heterogeneity arising from the decoupling of positional embeddings, misalignment of quantization scales in FP8 PV GEMM, and the need for optimized system-level support. In this paper, we introduce SnapMLA, an FP8 MLA decoding framework optimized to improve long-context efficiency through the following hardware-aware algorithm-kernel co-optimization techniques: (i) RoPE-Aware Per-Token KV Quantization, where the RoPE part is maintained in high precision, motivated by our comprehensive analysis of the heterogeneous quantization sensitivity inherent to the MLA KV cache. Furthermore, per-token granularity is employed to align with the autoregressive decoding process and maintain quantization accuracy. (ii) Quantized PV Computation Pipeline Reconstruction, which resolves the misalignment of quantization scale in FP8 PV computation stemming from the shared KV structure of the MLA KV cache. (iii) End-to-End Dataflow Optimization, where we establish an efficient data read-and-write workflow using specialized kernels, ensuring efficient data flow and performance gains. Extensive experiments on state-of-the-art MLA LLMs show that SnapMLA achieves up to a 1.91x improvement in throughput, with negligible risk of performance degradation in challenging long-context tasks, including mathematical reasoning and code generation benchmarks. Code is available at https://github.com/meituan-longcat/SGLang-FluentLLM.

Rui Yang, Han Zhong, Jiawei Xu et al.

Offline reinforcement learning (RL) presents a promising approach for learning reinforced policies from offline datasets without the need for costly or unsafe interactions with the environment. However, datasets collected by humans in real-world environments are often noisy and may even be maliciously corrupted, which can significantly degrade the performance of offline RL. In this work, we first investigate the performance of current offline RL algorithms under comprehensive data corruption, including states, actions, rewards, and dynamics. Our extensive experiments reveal that implicit Q-learning (IQL) demonstrates remarkable resilience to data corruption among various offline RL algorithms. Furthermore, we conduct both empirical and theoretical analyses to understand IQL's robust performance, identifying its supervised policy learning scheme as the key factor. Despite its relative robustness, IQL still suffers from heavy-tail targets of Q functions under dynamics corruption. To tackle this challenge, we draw inspiration from robust statistics to employ the Huber loss to handle the heavy-tailedness and utilize quantile estimators to balance penalization for corrupted data and learning stability. By incorporating these simple yet effective modifications into IQL, we propose a more robust offline RL approach named Robust IQL (RIQL). Extensive experiments demonstrate that RIQL exhibits highly robust performance when subjected to diverse data corruption scenarios.

Lei Zheng, Rui Yang, Minzhe Zheng et al.

Ensuring safety and driving consistency is a significant challenge for autonomous vehicles operating in partially observed environments. This work introduces a consistent parallel trajectory optimization (CPTO) approach to enable safe and consistent driving in dense obstacle environments with perception uncertainties. Utilizing discrete-time barrier function theory, we develop a consensus safety barrier module that ensures reliable safety coverage within the spatiotemporal trajectory space across potential obstacle configurations. Following this, a bi-convex parallel trajectory optimization problem is derived that facilitates decomposition into a series of low-dimensional quadratic programming problems to accelerate computation. By leveraging the consensus alternating direction method of multipliers (ADMM) for parallel optimization, each generated candidate trajectory corresponds to a possible environment configuration while sharing a common consensus trajectory segment. This ensures driving safety and consistency when executing the consensus trajectory segment for the ego vehicle in real time. We validate our CPTO framework through extensive comparisons with state-of-the-art baselines across multiple driving tasks in partially observable environments. Our results demonstrate improved safety and consistency using both synthetic and real-world traffic datasets.

Bohao Yang, Rui Yang, Joshua M. Biro et al.

Clinical communication is central to patient outcomes, yet large-scale human annotation of patient-provider conversation remains labor-intensive, inconsistent, and difficult to scale. Existing approaches based on large language models typically rely on single-task models that lack adaptability, interpretability, and reliability, especially when applied across various communication frameworks and clinical domains. In this study, we developed a Multi-framework Structured Agentic AI system for Clinical Communication (MOSAIC), built on a LangGraph-based architecture that orchestrates four core agents, including a Plan Agent for codebook selection and workflow planning, an Update Agent for maintaining up-to-date retrieval databases, a set of Annotation Agents that applies codebook-guided retrieval-augmented generation (RAG) with dynamic few-shot prompting, and a Verification Agent that provides consistency checks and feedback. To evaluate performance, we compared MOSAIC outputs against gold-standard annotations created by trained human coders. We developed and evaluated MOSAIC using 26 gold standard annotated transcripts for training and 50 transcripts for testing, spanning rheumatology and OB/GYN domains. On the test set, MOSAIC achieved an overall F1 score of 0.928. Performance was highest in the Rheumatology subset (F1 = 0.962) and strongest for Patient Behavior (e.g., patients asking questions, expressing preferences, or showing assertiveness). Ablations revealed that MOSAIC outperforms baseline benchmarking.

Mianchu Wang, Rui Yang, Xi Chen et al.

Offline Goal-Conditioned RL (GCRL) offers a feasible paradigm for learning general-purpose policies from diverse and multi-task offline datasets. Despite notable recent progress, the predominant offline GCRL methods, mainly model-free, face constraints in handling limited data and generalizing to unseen goals. In this work, we propose Goal-conditioned Offline Planning (GOPlan), a novel model-based framework that contains two key phases: (1) pretraining a prior policy capable of capturing multi-modal action distribution within the multi-goal dataset; (2) employing the reanalysis method with planning to generate imagined trajectories for funetuning policies. Specifically, we base the prior policy on an advantage-weighted conditioned generative adversarial network, which facilitates distinct mode separation, mitigating the pitfalls of out-of-distribution (OOD) actions. For further policy optimization, the reanalysis method generates high-quality imaginary data by planning with learned models for both intra-trajectory and inter-trajectory goals. With thorough experimental evaluations, we demonstrate that GOPlan achieves state-of-the-art performance on various offline multi-goal navigation and manipulation tasks. Moreover, our results highlight the superior ability of GOPlan to handle small data budgets and generalize to OOD goals.