arXiv subjects

Wei Fu

Publications and source records attributed to Wei Fu.

At least 19 recordsLinked to original sources

arXiv2026

Embracing Intra-Class Heterogeneity for Semi-Supervised Medical Image Segmentation: From Diversity to Precision

Due to the scarcity of expert-annotated data, Semi-Supervised Medical Image Segmentation (SSMIS) has emerged as a promising approach. Many anatomical structures in medical images exhibit significant intra-class heterogeneity, with different regions showing heterogeneous intensity patterns within the same structure. However, existing methods inadequately exploit this intensity-manifested intra-class heterogeneity, resulting in uniform structural representations and imprecise segmentation. Furthermore, the scarcity of labeled data makes it more difficult to effectively capture such complex heterogeneity. To address this, we propose Multiple Prototype Contrastive Learning (MPCL), an SSMIS framework that possesses better diversity and better precision. It consists of three novel designs: First, we provide structural representations with better diversity and propose Intensity-aligned Heterogeneous Prototype Generation (IHPG) that effectively models intra-class heterogeneity by generating multiple prototypes aligned with intensity characteristics. Second, we further enhance more diverse structural representations and build a solid foundation for more precise segmentation through Prototypical Space Optimization (PSO) that systematically optimizes a more discriminative and generalizable prototypical space. Finally, we achieve segmentation results with better precision through Dual-branch Knowledge Alignment (DKA) that efficiently promotes intra-class heterogeneity knowledge transfer from prototypical space to the segmentation network. Extensive experiments on three medical image datasets with significant intra-class heterogeneity demonstrate that MPCL significantly outperforms existing methods, especially under extremely limited labeled data.

cs.CV↗

arXiv2026

Next-Generation Agentic Reinforcement Learning Systems Enable Self-Evolving Agents

LLM agents are rapidly being deployed in production, including coding assistants, customer-support chatbots, and scientific research assistants, yet they remain fundamentally static in enterprise deployment. The LLM weights, system prompts, tool repertoires, and in-context harnesses are frozen at deployment time, and any improvement requires a manual loop of human-curated data collection, offline fine-tuning, modification of the agentic paradigm, and re-deployment. Recent work on self-evolving agents, such as OpenClaw for individual users, indicates that the next leap in agent capability will come from agents that continually learn from their own experience. In this paper, we argue that this vision for self-evolving agent deployment is being held back for enterprise-level large-scale agentic service not by reinforcement learning (RL) algorithms but by agentic online RL systems. Specifically, current agentic RL systems and the surrounding observability software stack are inadequate along three essential aspects: (i) there is no standardized agent trajectory data protocol capable of carrying RL learning signals at step granularity across heterogeneous agent paradigms; (ii) there is no enterprise-grade comprehensive data proxy that converts real workloads into governed learning substrates; and (iii) there is no unified agent evolution control plane that automatically decides, based on trajectory statistics, when to update policy weights or evolve the in-context harness. The next generation of agentic RL systems must be co-designed around these three pillars, and we sketch concrete architectures, case studies, and counter-arguments. We instantiate one branch through AReaL2.0, reorganizing existing RL infrastructure into an agent-oriented online RL loop for policy weight updates from deployed workloads.

cs.DC↗

arXiv2026

Cladding Layer Enhanced GHz Bulk Acoustic Wave Resonance in Sodium Niobate Thin Films on Silicon

Bulk acoustic wave resonators (BAWR) and bandpass filters operating at GHz frequency are the workhorse of (Vo-)LTE telecommunication and broadband internet. In line with the Singapore Green Plan 2030 for innovating environmentally friendly products, we fabricated lead-free BAWR with sodium niobate (NaNbO3) piezoelectric on silicon with a high electromechanical coupling factor up to 31.3% operating at ~4 GHz. We disclose our crucial strategy where the NaNbO3 layer is cladded between two thin layers of high band gap insulators, which satisfies two primary objectives, i.e. leakage current mitigation and crack avoidance. In addition, we also verified the efficacy of reducing lattice parameters of the cladding layers in promoting vertically distorted tetragonal phase NaNbO3 and producing stronger BAWR signals.

cond-mat.mtrl-sci↗

arXiv2026

Vector Map as Language: Toward Unified Remote Sensing Vector Mapping

Remote sensing vector mapping aims to generate structured maps of geospatial entities, such as buildings, roads, and water bodies, from remote sensing imagery. In practice, vector maps usually contain multiple category layers and heterogeneous entity structures, requiring a unified model for diverse mapping needs. However, existing methods typically represent vector objects as polygons or graphs, making them suitable only for specific categories: polygons poorly capture topological relations, while graphs often blur instance boundaries. We observe that language, as a natural medium for human communication, offers a flexible and expressive representation that can accommodate heterogeneous map elements, including geometry, semantics, and topolog. Motivated by this insight, we propose Vector Map as Language (VecLang), a unified paradigm that reformulates multiclass vector mapping as structured text generation. VecLang encodes the common elements of different geospatial entities into a GeoJSON-like vector language, enabling cross-category modeling within a shared textual format. To generate this language reliably, we design a progressive vision-language mapping framework that first localizes vectorization units and then generates structured map elements. We further introduce Hierarchical Vector Language Optimization, which uses reinforcement learning to improve syntax validity, content fidelity, and map executability. We also build VecMap-Bench with 54K images and 800K instances, supporting training and evaluation across standard and generalization settings. Extensive experiments demonstrate that VecLang handles both single-class and multiclass vector mapping while achieving strong cross-dataset and open-vocabulary generalization. The model and dataset are publicly available at https://github.com/yyyyll0ss/VecLang.

cs.CV↗

arXiv2026

SCKAN: Structural Consensus-based KAN Prototype Learning for Semi-Supervised Pancreas Segmentation

Accurate pancreas segmentation is critical for early cancer diagnosis, where annotation scarcity necessitates Semi-Supervised Learning (SSL). However, due to significant inter-sample morphological variability, existing SSL methods face severe generalizability limitations under sparse supervision, leading to the Supervision Bias problem. To address this, we propose Structural Consensus-based KAN Prototype Learning (SCKAN), which constructs the first cross-sample structural consensus learning with Kolmogorov-Arnold Networks (KANs), to achieve more generalizable and accurate segmentation. Specifically, SCKAN contains two key designs: Structure-constrained Prototype Consistency Learning (SPCL), which prompts unbiased structural representation by enforcing cross-sample consistency via prototype-level contrastive optimization, and Consensus-based Kolmogorov-Arnold Fusion (CKaF), which reduces morphology-specific bias by aggregating stable consensus and filtering sample-wise noise via KAN's adaptive B-spline nonlinearity. Extensive experiments on two public pancreas datasets demonstrate the effectiveness of SCKAN. Code is at https://github.com/rhodaliu17/SCKAN.

cs.CV↗

arXiv2026

MBD: A Model-Based Debiasing Framework Across User, Content, and Model Dimensions

Modern recommendation systems rank candidates by aggregating multiple behavioral signals through a value model. However, many commonly used signals are inherently affected by heterogeneous biases. For example, watch time naturally favors long-form content, loop rate favors short - form content, and comment probability favors videos over images. Such biases introduce two critical issues: (1) value model scores may be systematically misaligned with users' relative preferences - for instance, a seemingly low absolute like probability may represent exceptionally strong interest for a user who rarely engages; and (2) changes in value modeling rules can trigger abrupt and undesirable ecosystem shifts. In this work, we ask a fundamental question: can biased behavioral signals be systematically transformed into unbiased signals, under a user - defined notion of ``unbiasedness'', that are both personalized and adaptive? We propose a general, model-based debiasing (MBD) framework that addresses this challenge by augmenting it with distributional modeling. By conditioning on a flexible subset of features (partial feature set), we explicitly estimate the contextual mean and variance of the engagement distribution for arbitrary cohorts (e.g., specific video lengths or user regions) directly alongside the main prediction. This integration allows the framework to convert biased raw signals into unbiased representations, enabling the construction of higher-level, calibrated signals (such as percentiles or z - scores) suitable for the value model. Importantly, the definition of unbiasedness is flexible and controllable, allowing the system to adapt to different personalization objectives and modeling preferences. Crucially, this is implemented as a lightweight, built-in branch of the existing MTML ranking model, requiring no separate serving infrastructure.

cs.LG↗

arXiv2026

AREAL-DTA: Dynamic Tree Attention for Efficient Reinforcement Learning of Large Language Models

Reinforcement learning (RL)-based post-training for large language models (LLMs) is computationally expensive, as it generates many rollout sequences that frequently share long token prefixes. Existing RL frameworks usually process these sequences independently during policy training, i.e., repeatedly recomputing identical prefixes in both the forward and backward passes of policy gradient computation, leading to substantial inefficiencies in computation resources and memory usage. Although prefix sharing naturally induces a tree structure over rollouts, packed tree-mask approaches scale poorly in RL settings. In this paper, we introduce AReaL-DTA, which efficiently exploits prefix sharing in RL training. AReaL-DTA employs a depth-first search (DFS)-based execution strategy that dynamically traverses the rollout prefix tree during both forward and backward computation, materializing only a single root-to-leaf path at a time. To further improve scalability, AReaL-DTA incorporates a load-balanced distributed batching mechanism that dynamically constructs and processes prefix trees across multiple GPUs. On $\tau^2$-bench, AReaL-DTA improves training throughput by up to $8.31\times$ over dense training and up to $1.70\times$ over sparse training. Our code is available at https://github.com/areal-project/AReaL/tree/feat/dta.

cs.LG↗

arXiv2025

AReaL-Hex: Accommodating Asynchronous RL Training over Heterogeneous GPUs

Maximizing training throughput and cost-efficiency of RL for LLMs is essential to democratize this advanced technique. One promising but challenging approach is to deploy such a computational workflow over heterogeneous GPUs. Unlike conventional large-scale LLM pretraining, RL training generally decomposes into three coupled stages, i.e., rollout generation, reward computation, and policy/value updates, which exhibit markedly different compute intensities, memory footprints, and communication patterns. Recent research shows that fully asynchronous RL training can disaggregate these stages across disjoint hardware pools without sacrificing training stability, creating a great opportunity for real-world heterogeneous deployment. To this end, we present AReaL-Hex, a heterogeneity-aware asynchronous RL training system that effectively schedules how to execute rollout generation and policy model training over heterogeneous GPUs while enforcing data staleness bounds. Concretely, we use a two-phase scheduler: (i) a constrained search with MILP to select per-stage parallelization strategies and workload assignments given a resource budget, and (ii) a graph-partitioning step that allocates heterogeneous GPUs and interconnects to maximize end-to-end throughput. Built atop a fully asynchronous RL architecture, AReaL-Hex maps HBM-I/O-bound generation and compute-bound optimization to more cost-efficient resources and balances their producer-consumer interactions to avoid both idleness and stale rollout trajectories. On the mathematical reasoning task with various model scales (1.5B, 7B, and 14B), compared to homogeneous deployments of state-of-the-art asynchronous RL systems: (i) When maintaining the same total budgets, AReaL-Hex delivers up to 1.50x higher training throughput; (ii) When achieving the same training throughput, AReaL-Hex results in up to 1.46x reduction in training cost.

cs.DC↗

arXiv2025

Beyond Ten Turns: Unlocking Long-Horizon Agentic Search with Large-Scale Asynchronous RL

Recent advancements in LLM-based agents have demonstrated remarkable capabilities in handling complex, knowledge-intensive tasks by integrating external tools. Among diverse choices of tools, search tools play a pivotal role in accessing vast external knowledge. However, open-source agents still fall short of achieving expert-level Search Intelligence, the ability to resolve ambiguous queries, generate precise searches, analyze results, and conduct thorough exploration. Existing approaches fall short in scalability, efficiency, and data quality. For example, small turn limits in existing online RL methods, e.g. <=10, restrict complex strategy learning. This paper introduces ASearcher, an open-source project for large-scale RL training of search agents. Our key contributions include: (1) Scalable fully asynchronous RL training that enables long-horizon search while maintaining high training efficiency. (2) A prompt-based LLM agent that autonomously synthesizes high-quality and challenging QAs, creating a large-scale QA dataset. Through RL training, our prompt-based QwQ-32B agent achieves substantial improvements, with 78.0% and 34.3% Avg@4 gains on xBench and GAIA, respectively. Notably, our agent exhibits extreme long-horizon search, with tool calls exceeding 100 turns and output tokens exceeding 400k during training time. With a simple agent design and no external LLMs, ASearcher-Web-QwQ achieves Avg@4 scores of 51.1 on xBench and 58.7 on GAIA, surpassing existing open-source 32B agents. Finally, we also show that ASearcher-Web-QwQ could achieve performance of commercial systems using external summary tool in a zero-shot transfer manner and test-time search. We open-source our models, training data, and codes in https://github.com/inclusionAI/ASearcher.

cs.CL↗

arXiv2025

How Far Are We from Optimal Reasoning Efficiency?

Large Reasoning Models (LRMs) demonstrate remarkable problem-solving capabilities through extended Chain-of-Thought (CoT) reasoning but often produce excessively verbose and redundant reasoning traces. This inefficiency incurs high inference costs and limits practical deployment. While existing fine-tuning methods aim to improve reasoning efficiency, assessing their efficiency gains remains challenging due to inconsistent evaluations. In this work, we introduce the reasoning efficiency frontiers, empirical upper bounds derived from fine-tuning base LRMs across diverse approaches and training configurations. Based on these frontiers, we propose the Reasoning Efficiency Gap (REG), a unified metric quantifying deviations of any fine-tuned LRMs from these frontiers. Systematic evaluation on challenging mathematical benchmarks reveals significant gaps in current methods: they either sacrifice accuracy for short length or still remain inefficient under tight token budgets. To reduce the efficiency gap, we propose REO-RL, a class of Reinforcement Learning algorithms that minimizes REG by targeting a sparse set of token budgets. Leveraging numerical integration over strategically selected budgets, REO-RL approximates the full efficiency objective with low error using a small set of token budgets. Through systematic benchmarking, we demonstrate that our efficiency metric, REG, effectively captures the accuracy-length trade-off, with low-REG methods reducing length while maintaining accuracy. Our approach, REO-RL, consistently reduces REG by >=50 across all evaluated LRMs and matching Qwen3-4B/8B efficiency frontiers under a 16K token budget with minimal accuracy loss. Ablation studies confirm the effectiveness of our exponential token budget strategy. Finally, our findings highlight that fine-tuning LRMs to perfectly align with the efficiency frontiers remains an open challenge.

cs.CL↗

arXiv2025

AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning

Reinforcement learning (RL) has become a dominant paradigm for training large language models (LLMs), particularly for reasoning tasks. Effective RL for LLMs requires massive parallelization and poses an urgent need for efficient training systems. Most existing large-scale RL systems for LLMs are synchronous, alternating generation and training in a batch setting where rollouts in each training batch are generated by the same model. This approach stabilizes RL training but suffers from severe system-level inefficiency: generation must wait until the longest output in the batch is completed before model updates, resulting in GPU underutilization. We present AReaL, a fully asynchronous RL system that completely decouples generation from training. Rollout workers in AReaL continuously generate new outputs without waiting, while training workers update the model whenever a batch of data is collected. AReaL also incorporates a collection of system-level optimizations, leading to substantially higher GPU utilization. To stabilize RL training, AReaL balances the workload of rollout and training workers to control data staleness, and adopts a staleness-enhanced PPO variant to better handle outdated training samples. Extensive experiments on math and code reasoning benchmarks show that AReaL achieves up to 2.77$\times$ training speedup compared to synchronous systems with the same number of GPUs and matched or improved final performance. The code of AReaL is available at https://github.com/inclusionAI/AReaL/.

cs.LG↗

arXiv2025

Ground Calibration Result of the Wide-field X-ray Telescope (WXT) onboard the Einstein Probe

We report on results of the on-ground X-ray calibration of the Wide-field X-ray Telescope (WXT) built from novel lobster-eye micro-pore optics, onboard the Einstein Probe (EP) satellite. To fully characterize the instrumental performance and properties, a series of tests and calibrations have been carried out at different levels of devices, assemblies and the complete module before the launch of EP. In this paper, we present the calibration results of three flight model modules (FM1, FM5 and FM11) obtained during their end-to-end module calibration experiments carried out at the 100-m X-ray Test Facility (100XF) of IHEP, CAS. Measurements of the Point Spread Function (PSF), effective area, and energy response were performed for multiple incident directions and several characteristic X-ray emission line energies. Specifically, the distributions of the PSF and effective areas are found to be roughly uniform across the FoV, in large agreement with the prediction of lobster-eye optics. Their energy dependence behavior aligns well with theoretical predictions and Monte Carlo simulations. At 1.25 keV, the full width at half maximum (FWHM) of the focal spot is in range of 3-7 arcmin (a median of 4.2) and the effective area in range of 2-3 $cm^2$. Noticeably, the flight model instruments demonstrate a $\sim1.5$ arcmin spatial resolution improvement over the previously launched Lobster Eye Imager for Astronomy. The properties of the complementary metal-oxide semiconductor (CMOS) sensors were also calibrated. The gain coefficients are in range of 6.4-6.9 eV/DN. The energy resolutions are in range of 120-140 eV at 1.25 keV, meeting design requirements. These calibration results have been ingested into the first version of calibration database (CALDB) and applied to the analysis of the scientific data acquired by WXT after the launch of EP.

astro-ph.IM↗

arXiv2025

Nuclear Physics at BRIF

The Beijing Radioactive Ion-beam Facility (BRIF), which is based on Isotope Separation On-Line (ISOL) technique, consists of a 100 MeV proton cyclotron as the driving accelerator, a two-stage ISOL system for ion separation, a 13-MV tandem accelerator for post-acceleration, a superconducting linac for further boosting beam energies. It is capable of providing ISOL beams in the energy range from 60 to 300 keV, and post-accelerated beams in the energy range from 3 to 10 MeV/u for nuclei with mass numbers of A < 80 by Isotope Separation On-Line (ISOL) technique. For nuclei with A up to 170, energies are still able to reach 3 MeV/u. This facility offers opportunities to address key questions of current interest in nuclear astrophysics, nuclear structure and reactions of unstable nuclei. In this review we present a comprehensive introduction to the BRIF and the typical experimental instruments installed on it, and then summarize current experimental results on unstable Na and Rb isotopes and future plan for further development of the BRIF to improve its performance.

physics.acc-ph↗

arXiv2024

On Designing Effective RL Reward at Training Time for LLM Reasoning

Reward models have been increasingly critical for improving the reasoning capability of LLMs. Existing research has shown that a well-trained reward model can substantially improve model performances at inference time via search. However, the potential of reward models during RL training time still remains largely under-explored. It is currently unclear whether these reward models can provide additional training signals to enhance the reasoning capabilities of LLMs in RL training that uses sparse success rewards, which verify the correctness of solutions. In this work, we evaluate popular reward models for RL training, including the Outcome-supervised Reward Model (ORM) and the Process-supervised Reward Model (PRM), and train a collection of LLMs for math problems using RL by combining these learned rewards with success rewards. Surprisingly, even though these learned reward models have strong inference-time performances, they may NOT help or even hurt RL training, producing worse performances than LLMs trained with the success reward only. Our analysis reveals that an LLM can receive high rewards from some of these reward models by repeating correct but unnecessary reasoning steps, leading to a severe reward hacking issue. Therefore, we introduce two novel reward refinement techniques, including Clipping and Delta. The key idea is to ensure the accumulative reward of any reasoning trajectory is upper-bounded to keep a learned reward model effective without being exploited. We evaluate our techniques with multiple reward models over a set of 1.5B and 7B LLMs on MATH and GSM8K benchmarks and demonstrate that with a carefully designed reward function, RL training without any additional supervised tuning can improve all the evaluated LLMs, including the state-of-the-art 7B LLM Qwen2.5-Math-7B-Instruct on MATH and GSM8K benchmarks.

cs.LG↗

arXiv2024

ReaL: Efficient RLHF Training of Large Language Models with Parameter Reallocation

Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for empowering large language model (LLM) applications. Compared with the supervised training process of LLMs, the RLHF training process is much more sophisticated, requiring a diverse range of computation workloads with intricate dependencies between multiple LLM instances. Therefore, simply adopting the fixed parallelization strategies from supervised training for LLMs can be insufficient for RLHF and result in low training efficiency. To overcome this limitation, we propose a novel technique named parameter ReaLlocation, which dynamically adapts the parallelization strategies for different workloads during training by redistributing LLM parameters across the training cluster. Building upon this idea, we introduce ReaL, a pioneering system for efficient RLHF training. ReaL introduces the concept of an execution plan, which defines a fine-grained resource allocation and parallelization strategy particularly designed for RLHF training. Based on this concept, ReaL employs a tailored search algorithm with a lightweight run-time estimator to automatically discover an efficient execution plan for an instance of RLHF experiment. Subsequently, the runtime engine deploys the selected plan by effectively parallelizing computations and redistributing parameters. We evaluate ReaL on the LLaMA models with up to 70 billion parameters and 128 GPUs. The experimental results demonstrate that ReaL achieves speedups of up to $3.58\times$ compared to baseline methods. Furthermore, the execution plans generated by ReaL exhibit an average of $81\%$ performance improvement over heuristic approaches based on Megatron-LM in the long-context scenario. The source code of ReaL is publicly available at https://github.com/openpsi-project/ReaLHF .

cs.DC↗

arXiv2024

Towards edge engineering of two-dimensional layered transition-metal dichalcogenides by chemical vapor deposition

The manipulation of edge configurations and structures in atomically thin transition metal dichalcogenides (TMDs) for versatile functionalization has attracted intensive interest in recent years. The chemical vapor deposition (CVD) approach has shown promise for TMD edge engineering of atomic edge configurations (1H, 1T or 1T'-zigzag or armchair edges), as well as diverse edge morphologies (1D nanoribbons, 2D dendrites, 3D spirals, etc). These rich-edge TMD layers offer versatile candidates for probing the physical and chemical properties, and exploring new applications in electronics, optoelectronics, catalysis, sensing and quantum field. In this review, we present an overview of the current state-of-the-art in the manipulation of TMD atomic edges and edge-rich structures using CVD. We highlight the vast range of unique properties associated with these edge configurations and structures and provide insights into the opportunities afforded by such edge-functionalized crystals. The objective of this review is to motivate further research and development efforts in using CVD as a scalable approach to harness the benefits of such crystal-edge engineering.

cond-mat.mtrl-sci↗

arXiv2024

Is DPO Superior to PPO for LLM Alignment? A Comprehensive Study

Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions. Our code is publicly available at https://github.com/openpsi-project/ReaLHF.

cs.CL↗

arXiv2024

Dielectrics for Two-Dimensional Transition Metal Dichalcogenide Applications

Despite over a decade of intense research efforts, the full potential of two-dimensional transition metal dichalcogenides continues to be limited by major challenges. The lack of compatible and scalable dielectric materials and integration techniques restrict device performances and their commercial applications Conventional dielectric integration techniques for bulk semiconductors are difficult to adapt for atomically thin two-dimensional materials. This review provides a brief introduction into various common and emerging dielectric synthesis and integration techniques and discusses their applicability for 2D transition metal dichalcogenides. Dielectric integration for various applications is reviewed in subsequent sections including nanoelectronics, optoelectronics, flexible electronics, valleytronics, biosensing, quantum information processing, and quantum sensing. For each application, we introduce basic device working principles, discuss the specific dielectric requirements, review current progress, present key challenges, and offer insights into future prospects and opportunities.

cond-mat.mtrl-sci↗