arXiv subjects

Xiaorong Wang

Publications and source records attributed to Xiaorong Wang.

At least 19 recordsLinked to original sources

arXiv2026

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

Chain-of-Thought (CoT) improves the reasoning ability of Large Language Models (LLMs) but incurs substantial computation and context costs. Existing methods either lose intermediate information through hard pruning or lack a principled criterion for continuous compression. We present A*-Thought-V2, a geometric dynamics of LLM guided framework that models CoT as a hidden-state trajectory and replaces hard deletion with an explicit-implicit interleaved latent architecture. After projecting question, step, and solution representations into a 3D PCA space, it measures alignment between each local transition and global question-to-solution direction. Aligned steps remain explicit text, whereas deviating steps are compressed into continuous latent tokens. Directional angles capture both local semantics and reasoning dynamics: small angles indicate direct execution and answer formation, while large angles more frequently involve checking, correction, and branch exploration; their temporal variation reveals exploration, convergence, and refinement stages. To train this architecture, we introduce stepwise embedding forcing, which pools each redundant step into a single latent embedding, and label forcing, which supervises that latent token with a soft multi-modal vocabulary distribution instead of a hard one-hot label. Experiments on Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks show that A*-Thought-V2 improves average accuracy by up to 2.6% while reducing response length by up to half, increasing Accuracy per Computation Unit by 2.29$\times$, and reducing preprocessing and training time by 94.6% and up to 80.3%, respectively. Representation analyses suggest that latent states form a compact region distinct from textual states, while higher entropy at latent-token positions reflects broader soft targets that encourage richer step-level feature learning.

cs.CL↗

arXiv2026

Beyond Teacher Likelihood: Group-Calibrated On-Policy Distillation for Long-Context Reasoning

On-policy distillation (OPD) trains a student on its own responses using dense token-level guidance from a stronger teacher. In long-context tasks, however, token-level teacher support can favor locally plausible responses that omit evidence distributed across the input or violate global task constraints. Task-specific verifiers, in contrast, evaluate task completion at the response level and may return graded rewards that reflect partial success. We diagnose this mismatch on fixed responses from two representative long-context evidence-aggregation tasks. Across longer input ranges, trajectory-level OPD scores become progressively less aligned with verifier rewards, indicating teacher-verifier disagreement. Motivated by this observation, we introduce Group-Calibrated On-Policy Distillation (GC-OPD). GC-OPD separately normalizes verifier rewards and trajectory-level OPD scores within each rollout group and uses their difference as a signed teacher-verifier disagreement residual. Relative-advantage-based credit assignment (RACA) distributes this trajectory-level residual across tokens according to their relative OPD advantages while preserving the original OPD signal. Across five long-context benchmarks, post-training with GC-OPD raises the five-benchmark averages of the official Qwen3-4B and Qwen3-8B checkpoints from 29.08 to 40.47 and from 35.12 to 44.65, respectively. Vanilla OPD reaches 39.31 and 43.56 under the same setup. Controlled ablations show that the signed residual is more effective than either an additional OPD-derived term or direct group-normalized verifier reward addition, while RACA further improves over uniform token allocation. Together, these results demonstrate that group-relative residual calibration can incorporate verifier outcomes without discarding dense token-level guidance. Code is available at https://github.com/SolereZhang/GC-OPD.

cs.LG↗

arXiv2026

Beyond Reward Engineering: A Data Recipe for Long-Context Reinforcement Learning

Long-context reasoning is an essential capability for large language models, particularly when they are deployed as autonomous agents that must reason over lengthy trajectories. Reinforcement learning (RL) has recently emerged as a dominant paradigm for improving this ability, yet existing work largely focuses on reward engineering while diverse training data remains scarce. We revisit this problem from a data-centric perspective and show that a simple yet effective data recipe alone, paired with a minimal outcome-based GRPO setup, suffices to substantially improve long-context reasoning. Our recipe targets three complementary task families -- retrieval, multi-evidence synthesis, and reasoning -- for which we construct and curate eight datasets totaling ~14K examples. Experiments on three models (Qwen3-4B/8B/30B-A3B) yield average gains of +7.2/+3.2/+6.4 points across seven long-context benchmarks, surpassing prior RL training sets. We further demonstrate that these gains transfer to agentic tasks, where continuing RL training on an agent-tuned model with our data recipe improves GAIA by +4.8 and BrowseComp by +7.0 points. We will release our datasets to facilitate future research.

cs.CL↗

arXiv2026

GenTS: A Comprehensive Benchmark Library for Generative Time Series Models

Generative models have demonstrated remarkable potential in time series analysis tasks, like synthesis, forecasting, imputation, etc. However, offering limited coverage for generative models, existing time series libraries are mainly engineered for discriminative models, with standardized workflows for specific tasks, such as optimizing Mean Squared Errors for time series forecasting. This rigid structure is fundamentally incompatible with the distinct and often complex paradigms of generative models (e.g., adversarial training, diffusion processes), which learn the underlying data distribution rather than a direct input-output mapping. To this end, we proposed GenTS, a comprehensive and extensible benchmark library designed for systematic assessment on generative time series models. GenTS features a unified data preprocessing pipeline, a collection of versatile models, and panoramic evaluation metrics. Its modular design also enables the researchers to flexibly customize beyond our built-in datasets and models. Based on GenTS, we conducted benchmarking experiments under diverse tasks, accordingly offering suggestions for model selection and identifying potential directions for future research. Our codes are open-source at https://github.com/WillWang1113/GenTS. The official tutorials and document are available at https://willwang1113.github.io/GenTS/.

cs.LG↗

arXiv2026

MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling

The evolution of large language models (LLMs) towards applications with ultra-long contexts faces challenges posed by the high computational and memory costs of the Transformer architecture. While existing sparse and linear attention mechanisms attempt to mitigate these issues, they typically involve a trade-off between memory efficiency and model performance. This paper introduces MiniCPM-SALA, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention (InfLLM-V2) with the global efficiency of linear attention (Lightning Attention). By employing a layer selection algorithm to integrate these mechanisms in a 1:3 ratio and utilizing a hybrid positional encoding (HyPE), the model maintains efficiency and performance for long-context tasks. Furthermore, we introduce a cost-effective continual training framework that transforms pre-trained Transformer-based models into hybrid models, which reduces training costs by approximately 75% compared to training from scratch. Extensive experiments show that MiniCPM-SALA maintains general capabilities comparable to full-attention models while offering improved efficiency. On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens and supports context lengths of up to 1M tokens, a scale where traditional full-attention 8B models fail because of memory constraints.

cs.CL↗

arXiv2026

AgentCPM-Report: Interleaving Drafting and Deepening for Open-Ended Deep Research

Generating deep research reports requires large-scale information acquisition and the synthesis of insight-driven analysis, posing a significant challenge for current language models. Most existing approaches follow a plan-then-write paradigm, whose performance heavily depends on the quality of the initial outline. However, constructing a comprehensive outline itself demands strong reasoning ability, causing current deep research systems to rely almost exclusively on closed-source or online large models. This reliance raises practical barriers to deployment and introduces safety and privacy concerns for user-authored data. In this work, we present AgentCPM-Report, a lightweight yet high-performing local solution composed of a framework that mirrors the human writing process and an 8B-parameter deep research agent. Our framework uses a Writing As Reasoning Policy (WARP), which enables models to dynamically revise outlines during report generation. Under this policy, the agent alternates between Evidence-Based Drafting and Reasoning-Driven Deepening, jointly supporting information acquisition, knowledge refinement, and iterative outline evolution. To effectively equip small models with this capability, we introduce a Multi-Stage Agentic Training strategy, consisting of cold-start, atomic skill RL, and holistic pipeline RL. Experiments on DeepResearch Bench, DeepConsult, and DeepResearch Gym demonstrate that AgentCPM-Report outperforms leading closed-source systems, with substantial gains in Insight.

cs.AI↗

arXiv2025

The 2025 Roadmaps for the US Magnet Development Program

The US Physics community completed the Snowmass planning process in 2022, culminating in the HEPAP Particle Physics Project Prioritization Panel (P5) publishing its summary report at the end of 2023. Building on this, the US Magnet Development Program, a national accelerator magnet R&D program established by DOE-OHEP in 2016, has updated its strategic plan to align with the 2023 P5 report, resulting in this roadmap document.

physics.acc-ph↗

arXiv2025

Monocle: Hybrid Local-Global In-Context Evaluation for Long-Text Generation with Uncertainty-Based Active Learning

Assessing the quality of long-form, model-generated text is challenging, even with advanced LLM-as-a-Judge methods, due to performance degradation as input length increases. To address this issue, we propose a divide-and-conquer approach, which breaks down the comprehensive evaluation task into a series of localized scoring tasks, followed by a final global assessment. This strategy allows for more granular and manageable evaluations, ensuring that each segment of the text is assessed in isolation for both coherence and quality, while also accounting for the overall structure and consistency of the entire piece. Moreover, we introduce a hybrid in-context learning approach that leverages human annotations to enhance the performance of both local and global evaluations. By incorporating human-generated feedback directly into the evaluation process, this method allows the model to better align with human judgment. Finally, we develop an uncertainty-based active learning algorithm that efficiently selects data samples for human annotation, thereby reducing annotation costs in practical scenarios. Experimental results show that the proposed evaluation framework outperforms several representative baselines, highlighting the effectiveness of our approach.

cs.CL↗

arXiv2025

LLM$\times$MapReduce-V2: Entropy-Driven Convolutional Test-Time Scaling for Generating Long-Form Articles from Extremely Long Resources

Long-form generation is crucial for a wide range of practical applications, typically categorized into short-to-long and long-to-long generation. While short-to-long generations have received considerable attention, generating long texts from extremely long resources remains relatively underexplored. The primary challenge in long-to-long generation lies in effectively integrating and analyzing relevant information from extensive inputs, which remains difficult for current large language models (LLMs). In this paper, we propose LLM$\times$MapReduce-V2, a novel test-time scaling strategy designed to enhance the ability of LLMs to process extremely long inputs. Drawing inspiration from convolutional neural networks, which iteratively integrate local features into higher-level global representations, LLM$\times$MapReduce-V2 utilizes stacked convolutional scaling layers to progressively expand the understanding of input materials. Both quantitative and qualitative experimental results demonstrate that our approach substantially enhances the ability of LLMs to process long inputs and generate coherent, informative long-form articles, outperforming several representative baselines. Both LLM$\times$MapReduce-V2 and SurveyEval are publicly available at https://github.com/thunlp/LLMxMapReduce .

cs.CL↗

arXiv2024

An energy-stable phase-field model for droplet icing simulations

A phase-field model for three-phase flows is established by combining the Navier-Stokes (NS) and the energy equations, with the Allen-Cahn (AC) and Cahn-Hilliard (CH) equations and is demonstrated analytically to satisfy the energy dissipation law. A finite difference scheme is then established to discretize the model and this numerical scheme is proved to be unconditionally stable. Based on this scheme, the droplet icing process with phase changing is numerically simulated and the pointy tip of the icy droplet is obtained and analyzed. The influence of the temperature of the supercooled substrate and the ambient air on the droplet freezing process is studied. The results indicate that the formation of the droplet pointy tip is primarily due to the expansion in the vertical direction during the freezing process. Lower substrate temperatures can accelerate this process. Changes in air temperature have a relatively minor impact on the freezing process, mainly affecting its early stages. Moreover, our results demonstrate that the ice front transitions from an approximately horizontal shape to a concave one. Dedicated physical experiments were conducted and the measured solidification process matches the results of the proposed phase-field method very well.

physics.flu-dyn↗

arXiv2023

Energy and Carbon Considerations of Fine-Tuning BERT

Despite the popularity of the `pre-train then fine-tune' paradigm in the NLP community, existing work quantifying energy costs and associated carbon emissions has largely focused on language model pre-training. Although a single pre-training run draws substantially more energy than fine-tuning, fine-tuning is performed more frequently by many more individual actors, and thus must be accounted for when considering the energy and carbon footprint of NLP. In order to better characterize the role of fine-tuning in the landscape of energy and carbon emissions in NLP, we perform a careful empirical study of the computational costs of fine-tuning across tasks, datasets, hardware infrastructure and measurement modalities. Our experimental results allow us to place fine-tuning energy and carbon costs into perspective with respect to pre-training and inference, and outline recommendations to NLP researchers and practitioners who wish to improve their fine-tuning energy efficiency.

cs.CL↗

arXiv2022

High-performance chiral all-optical logic gate based on topological edge states of valley photonic crystal

For all-optical communication and information processing, it is necessary to develop all-optical logic gates based on photonic structures that can directly perform logic operations. All-optical logic gates have been demonstrated based on conventional waveguides and interferometry, as well as photonic crystal structures. Nonetheless, any defects in those structures will introduce high scattering loss, which compromises the fidelity and contrast ratio of the information process. Based on the spin-valley locking effect that can achieve defect-immune unidirectional transmission of topological edge states in valley photonic crystals (VPCs), we propose a high-performance all-optical logic OR gate based on a VPC structure. By tuning the working bandwidth of the two input channels, we prevent interference between the two channels to achieve a stable and high-fidelity output. The transmittance of both channels is higher than 0.8, and a high contrast ratio of 28.8 dB is achieved. Moreover, the chirality of the logic gate originated from the spin-valley locking effect allows using different circularly polarized light as inputs, representing "1" or "0", which is highly desired in quantum computing. The device's footprint is small, allowing high-density on-chip integration. In addition, this design can be experimentally fabricated using current nanofabrication techniques and will have potential applications in optical communication, information processing, and quantum computing.

physics.optics↗

arXiv2022

One Hyper-Initializer for All Network Architectures in Medical Image Analysis

Pre-training is essential to deep learning model performance, especially in medical image analysis tasks where limited training data are available. However, existing pre-training methods are inflexible as the pre-trained weights of one model cannot be reused by other network architectures. In this paper, we propose an architecture-irrelevant hyper-initializer, which can initialize any given network architecture well after being pre-trained for only once. The proposed initializer is a hypernetwork which takes a downstream architecture as input graphs and outputs the initialization parameters of the respective architecture. We show the effectiveness and efficiency of the hyper-initializer through extensive experimental results on multiple medical imaging modalities, especially in data-limited fields. Moreover, we prove that the proposed algorithm can be reused as a favorable plug-and-play initializer for any downstream architecture and task (both classification and segmentation) of the same modality.

cs.CV↗

arXiv2022

Contrastive Centroid Supervision Alleviates Domain Shift in Medical Image Classification

Deep learning based medical imaging classification models usually suffer from the domain shift problem, where the classification performance drops when training data and real-world data differ in imaging equipment manufacturer, image acquisition protocol, patient populations, etc. We propose Feature Centroid Contrast Learning (FCCL), which can improve target domain classification performance by extra supervision during training with contrastive loss between instance and class centroid. Compared with current unsupervised domain adaptation and domain generalization methods, FCCL performs better while only requires labeled image data from a single source domain and no target domain. We verify through extensive experiments that FCCL can achieve superior performance on at least three imaging modalities, i.e. fundus photographs, dermatoscopic images, and H & E tissue images.

cs.CV↗

arXiv2022

An Effective Transformer-based Solution for RSNA Intracranial Hemorrhage Detection Competition

We present an effective method for Intracranial Hemorrhage Detection (IHD) which exceeds the performance of the winner solution in RSNA-IHD competition (2019). Meanwhile, our model only takes quarter parameters and ten percent FLOPs compared to the winner's solution. The IHD task needs to predict the hemorrhage category of each slice for the input brain CT. We review the top-5 solutions for the IHD competition held by the Radiological Society of North America(RSNA) in 2019. Nearly all the top solutions rely on 2D convolutional networks and sequential models (Bidirectional GRU or LSTM) to extract intra-slice and inter-slice features, respectively. All the top solutions enhance the performance by leveraging the model ensemble, and the model number varies from 7 to 31. In the past years, since much progress has been made in the computer vision regime especially Transformer-based models, we introduce the Transformer-based techniques to extract the features in both intra-slice and inter-slice views for IHD tasks. Additionally, a semi-supervised method is embedded into our workflow to further improve the performance. The code is available in the manuscript.

cs.CV↗

arXiv2022

REBCO -- a silver bullet for our next high-field magnet and collider budget?

High-field superconducting magnets with a dipole field of 16 T and above enable future energy-frontier circular particle colliders. Although we believe these magnets can be built, none exists today. They can also be a showstopper for future high-energy machines due to a prohibitively high price tag based on the current conductor and magnet fabrication cost. The high-temperature superconducting REBCO coated conductor can address both the technical and cost issues, a silver bullet to lay both monsters to rest. The challenges and unknowns, however, can be too arduous to make the silver bullet. We lay out a potential road forward and suggest key action items. As a contribution from the accelerator community, we attempt to clarify for our theorist and experimenter colleagues a few aspects about the future high-field superconducting magnets. We hope to stimulate an effective plan for the 2023 P5 process that can lead to a cost-effective high-field magnet technology for future colliders and the exciting physics they can steward.

physics.acc-ph↗

arXiv2022

Common Coil Dipole for High Field Magnet Design and R&D

The common coil geometry provides an alternate design to the conventional cosine theta dipoles. It allows a wider range of conductor and magnet technologies. It also facilitates a low-cost, rapid-turn-around design and R&D program. Recent studies carried out as a part of the US Magnet Development Program revealed that at high fields (20 T with 15% operating margin or more), the common coil design also uses significantly less conductor (particularly much less HTS), as compared to that in the other designs.

physics.acc-ph↗

arXiv2022

Fiber-optic diagnostic system for future accelerator magnets

The next generation high energy physics accelerators will require magnetic fields at ~20 T. HTS coils will be an essential component of future accelerator magnets and several efforts are currently dedicated on designing 20 T HTS- LTS hybrid magnets. Among the existing challenges, there is the lack of a robust quench detection system for hybrid magnet technology. Another big challenge is represented by the high number of training quenches required by Nb3Sn magnets to reach performance level. In this framework it is important to find a tool that allow local real-time monitoring of magnet strain and temperature. In this paper, we propose the use of fiber optics sensors for diagnostic and quench detection in future accelerator superconducting magnets. Discrete and distributed fiber optic sensors have demonstrated to be a promising tool. The goal is to instrument hundreds of accelerator superconducting magnets and to move beyond the proof-of-concept level. Significant developments are still needed. Here, we are going to present the most recent results and discuss the most urgent technical developments in order to make those sensors a robust and reliable diagnostic tool for accelerator superconducting magnets over the next 10 year. We foresee that discrete fiber sensors will be a stable diagnostic probe for superconducting magnets over the next 3 to 5 years. More R&D work will be necessary for distributed fibers. The most urgent needs are the increase of sample rate and sensitivity. Close collaboration with vendors will be necessary to improve mechanical properties and fabrication processes in order to produce hundreds of meters of fiber and instrument a large number of accelerator superconducting magnets. Those R&D efforts will last up to 10 years with a founding level that spans between 5-10 M$.

physics.acc-ph↗