arXiv subjects

Wei Sun

Publications and source records attributed to Wei Sun.

At least 19 recordsLinked to original sources

arXiv2026

Look Before You Launch: Learned Throughput Prediction for Hardware Selection in a Production Training Fleet

In mixed-vendor accelerator fleets, many training jobs can run on several hardware types, and the best choice depends on each job's throughput on each type. Routing a job needs a throughput estimate for every candidate type before the job runs. We present LEAP, a machine learning performance prediction service that predicts each training job's throughput on every candidate hardware type at submission time and ranks the candidates by predicted throughput per unit cost. It routes tens of thousands of training job attempts a day across hundreds of different model types and more than 20 hardware types in a production fleet. LEAP predicts throughput from submission-time training configuration, including each job's prior throughput. LEAP reaches an accuracy of 77% over 30 nightly retrains, higher than the lookup table it replaced and at least as high as every alternative model we tested, including fine-tuned language models. Over six months, the share of eligible training GPU-hours that received a ranking by LEAP grew from 58% to 95%, and we estimate that it saves about 3% of training computational cost. In this paper, we identify the drivers that actually move accuracy and those that do not. We offer these as lessons for teams building performance predictors at fleet scale.

cs.AI↗

arXiv2026

Contrastive Retrieval Heads for Improved Attention-Based Reranking

The strong zero-shot and long-context capabilities of recent Large Language Models (LLMs) have paved the way for highly effective re-ranking systems. Attention-based re-rankers leverage attention weights from transformer heads to produce relevance scores, but not all heads are created equally: many contribute noise and redundancy, thus limiting performance. To address this, we introduce CoRe heads, a small set of retrieval heads identified via a contrastive scoring metric that explicitly rewards high attention heads that correlate with relevant documents, while downplaying nodes with higher attention that correlate with irrelevant documents. This relative ranking criterion isolates the most discriminative heads for re-ranking and yields a state-of-the-art list-wise re-ranker. Extensive experiments with three LLMs show that aggregated signals from CoRe heads, constituting less than 1% of all heads, substantially improve re-ranking accuracy over strong baselines. We further find that CoRe heads are concentrated in middle layers, and pruning the computation of final 50% of model layers preserves accuracy while significantly reducing inference time and memory usage.

cs.IR↗

arXiv2026

Optimizing Effective Training Time for Large-Scale Recommendation Systems

Lifecycle overhead silently consumes accelerator capacity across large-scale recommendation training fleets. Our largest recommendation workloads process tens of billions train- ing examples per day on thousands of GPUs. Before this work, only 50-60% of their end-to-end wall time advanced training on new data. We present a fleet-scale study of this lifecycle overhead and a set of optimizations spanning the full training stack. We use Effective Training Time (ETT%) as an operational framework to instrument lost time, localize it to independently owned infrastructure components, and expose work repeated across job restarts. This analysis guides optimizations like communication elimination and pipeline overlap during trainer initialization; dynamic-shape handling, autotuning pruning, and reusable Py- Torch 2 compilation caches; asynchronous checkpointing; stan- dalone model publishing; and reductions in recovery cost. We evaluate the optimizations on representative models and measure their impacts in our training fleet. ETT% improves on every benchmark, by 15.5% on average, and reaches 85% on our largest workload. Fleet-wide ETT% rose from about 80% to above 90% after deployment.

cs.IR↗

arXiv2026

Where Did We Go Wrong? Localizing Errors to Repair Multi-Turn Conversations

Large language model (LLM) assistants are increasingly used in multi-turn conversations, where the user's intent is often unclear at first and revealed only gradually over time. The assistant must therefore interpret each exchange and form inferences about the user's intent, using them to update its working state. These updates usually refine the assistant's understanding, but a misinterpretation can instead corrupt its working state by ruling out the true intent, influencing subsequent turns and ultimately causing the conversation to fail. Efficient recovery then requires locating and revising the misinterpretation without discarding information that remains valid. We introduce Loci (Localization Of Corrupted Inferences), a framework for probabilistically localizing misinterpretations and selectively repairing the assistant's working state. Given an observed failure, Loci reasons over possible user intents and assigns each update a probability of having excluded the true intent. These probabilities rank suspected updates and quantify the model's confidence in the localization of the underlying misinterpretations. Loci then uses this information in a repair policy that decides whether to retain, verify, or remove suspected updates, enabling recovery from corrupted inferences while preserving useful information. To evaluate Loci's ability to identify misinterpretations, we first test it in a controlled setup with injected errors. We find that Loci improves localization over the baselines, with its advantage becoming larger when multiple errors occur. We then evaluate it in free-form conversations, where misinterpretations arise naturally, and measure the interactions needed to complete the task after an initial failure. Using localization confidence to guide repair, Loci reduces mean recovery cost by 46-79% relative to the strongest baseline across multiple domains.

cs.CL↗

arXiv2026

MetaEncoder: Exploring the Limit of Bi-Encoders for Multimodal System One Decision Making with Natural Language Interface

System One models output constrained decisions and probability distributions rather than free-form text generation. While prevailing paradigms rely on structured schema objects to encode state, intent, and candidate choices, we revisit a fully natural language-based System One interface. In this framework, both the user request and each candidate option are expressed in natural language, supported by multimodal (image and video) auxiliary inputs. We introduce MetaEncoder, which fine-tunes a pre-trained Muse-Glimmer 30B decoder into an instruction-following decision-making encoder. To scale effectively across both small closed-set (< 256) and massive open-set (millions) candidate spaces, MetaEncoder employs a bi-encoder architecture trained via unidirectional contrastive learning for request-candidate alignment. We conduct extensive evaluations across 11 benchmark suites and 190 tasks spanning multimodal decision-making, understanding (closed-set) and retrieval (open-set), highlighting where MetaEncoder beats SOTA multimodal encoders, as well as its current limits on reasoning-intensive tasks.

cs.CL↗

arXiv2026

Performance Portable $\mathrm{SU}(N)$ Lattice Gauge Theory Simulation with Kokkos

The increasing diversity of high performance computing systems makes separate, architecture specific implementations of lattice gauge theory algorithms costly to maintain. We present \texttt{kwqft}, a performance portable Kokkos implementation of Wilson pure gauge Monte Carlo simulation for $\mathrm{SU}(N)$ Yang-Mills theory in an arbitrary number of space-time dimensions. The gauge group order $N$ and the dimension $D$ are compile time parameters. A single source targets the Serial, OpenMP, CUDA, HIP, and SYCL execution spaces, with MPI halo exchange overlapped with interior updates. The implementation reproduces the exact two-dimensional plaquette and published three and four dimensional values for gauge groups up to $\mathrm{SU}(17)$. On an NVIDIA A100 the Kokkos CUDA backend is competitive with a native CUDA code, SIMD acceleration improves the OpenMP path on Armv9 processors, and a large scale speedup is demonstrated for an $\mathrm{SU}(4)$ lattice on the LineShine supercomputer, currently ranked first on the TOP500 list.

hep-lat↗

arXiv2026

A photonic integrated comb engine for ultracold quantum gases

Cold atoms underpin quantum sensing, simulation and computation, but their coherent control demands highly stable optical fields whose generation, referencing and power scaling remain formidable integration challenges. While photonic integrated circuits have yielded compact visible lasers and high-$Q$ microresonators have enabled chip-scale optical frequency combs, these crucial technologies have largely remained functionally fragmented. Consequently, the coherent manipulation of ultracold quantum gases using a fully integrated laser-comb source has yet to be realized. Here we demonstrate a scalable, hybrid-integrated microcomb engine at 780 nm that seamlessly bridges frequency synthesis, atomic referencing and power amplification to achieve quantum state control of a Bose--Einstein condensate. By self-injection locking of electrically driven III--V lasers to high-$Q$ Si$_3$N$_4$ microresonators, we generate coherent platicon microcombs featuring 20- and 100-GHz mode spacings. Absolute referencing of the microcomb to an $^{85}$Rb transition actively suppresses long-term frequency drift from over 200 MHz to the 100-kHz level across 2,000 s. A selected comb tooth is subsequently injection-amplified to 102 mW, entirely preserving the microcomb's pristine coherence and stability. We utilize this synthesized field to construct an optical lattice, drive coherent two-photon Raman transitions, and prepare stationary spin--orbit-coupled and Raman-lattice states within an $^{87}$Rb condensate. By providing a synchronized optical grid, this atom-referenced microcomb can allow multiple optical-control channels to scale without a proportional multiplication of independent frequency references. Our work establishes a fully integrated frequency-synthesis architecture essential for realizing deployable, large-scale atomic quantum systems.

physics.optics↗

arXiv2026

Nuclear mass table in deformed relativistic Hartree-Bogoliubov theory in continuum, III: nuclei with $8 \leq Z \leq 120$

The mass table in the deformed relativistic Hartree-Bogoliubov theory in continuum (DRHBc) with the PC-PK1 density functional has been established for nuclei with $8 \leq Z \leq 120$, extended from the previous works for even-even nuclei [Zhang et al. (DRHBc mass table collaboration), At. Data Nucl. Data Tables 144, 101488 (2022)] and for even-$Z$ nuclei [Guo et al. (DRHBc mass table collaboration), At. Data Nucl. Data Tables 158, 101661 (2024)]. The calculated binding energies, two- and one-nucleon separation energies, root-mean-square (rms) radii of neutron, proton, matter, and charge distributions, quadrupole deformations, neutron and proton Fermi surfaces, and the blocked neutron (proton) orbitals of odd-$N$ ($Z$) nuclei are tabulated and compared with the available experimental data. A total of 9495 nuclei are predicted to be bound, with an rms deviation of 1.444 MeV from the 2380 mass data. Good agreement with the available experimental pairing gaps, $α$ decay energies, and charge radii is also achieved. The accuracies of the calculated nuclear masses and nucleon separation energies as well as the prediction for drip lines are compared with those obtained by other relativistic and nonrelativistic density functional calculations. It turns out that the DRHBc theory with PC-PK1 provides one of the best microscopic descriptions for nuclear masses. The systematics of nucleon separation energies, pairing gaps, pairing energies, two-nucleon gaps, $α$ decay energies, rms radii, quadrupole deformations, potential energy curves, neutron density distributions, and neutron mean-field potentials are discussed.

nucl-th↗

arXiv2026

GenCOPE: Syn2Real Generalized Category-Level Object Pose Estimation for Robotic Picking

Category-level object pose estimation (COPE), capable of generalizing to intra-class unknown objects, has become a core technique for robotic 3D scene understanding. However, existing COPE methods still require labor-intensive recollection of real-world training data for novel object categories, which limits their scalability in practical applications. This paper aims to achieve synthetic-to-real (Syn2Real) generalized COPE, where a model is trained solely on rendered synthetic data and directly generalized to real-world deployments. The central challenge lies in the significant domain gap between synthetic and real-world data, particularly in texture appearance. To address this, we aim to enhance domain generalization by learning domain-invariant representations that capture semantic commonalities among objects within the same category. We introduce 2D and 3D semantic consistency constraints to reduce the sensitivity of feature encoders to domain-specific features. In addition, we propose an end-to-end pose regression framework that performs 2D-3D cross consistency learning, leveraging dense cross-modality fusion to further refine pose estimation. Since simplicity and effectiveness are essential for real-world robotic deployment, our model operates exclusively on global features, yielding a highly lightweight and efficient architecture. Extensive experiments on the REAL275 and Wild6D benchmarks, as well as real-world robotic manipulation scenes, show superior Syn2Real generalization performance of our paradigm. Code and demos are released at https://paperreview99.github.io/GenCOPE/.

cs.CV↗

arXiv2026

An Unbounded Archive-based Transfer Strategy for Dynamic Multi-Objective Optimization with a Changing Number of Objectives

Dynamic multi-objective optimization with a variable number of objectives is difficult because objective-dimensional variations may significantly change the Pareto front and degrade algorithm adaptability. This paper proposes an unbounded archive-based transfer strategy (UATS), which maintains an unbounded archive of offspring solutions within each environment stage and extracts feasible nondominated solutions as transferable elites when objective changes occur. UATS is embedded into SPEA2SDE to construct UATS-SPEA2SDE, enabling the algorithm to reuse historical evolutionary information while retaining the convergence and diversity advantages of shift-based density estimation. Experiments are conducted on four benchmark problems under three objective-changing settings, where UATS-SPEA2SDE is compared with a restart-based SPEA2SDE baseline and four representative dynamic multi-objective optimization algorithms. The results indicate that the archive-guided transfer improves recovery after environmental changes and enhances adaptability to objective-number variations.

cs.NE↗

arXiv2026

SafetyFlow: An Agent-Flow System for Automated LLM Safety Benchmarking

The rapid proliferation of large language models (LLMs) has intensified the requirement for reliable safety evaluation to uncover model vulnerabilities. To this end, numerous LLM safety evaluation benchmarks are proposed. However, existing benchmarks generally rely on labor-intensive manual curation, which causes excessive time and resource consumption. They also exhibit significant redundancy and limited difficulty. To alleviate these problems, we introduce SafetyFlow, the first agent-flow system designed to automate the construction of LLM safety benchmarks. SafetyFlow can automatically build a comprehensive safety benchmark in only four days without any human intervention by orchestrating seven specialized agents, significantly reducing time and resource cost. Equipped with versatile tools, the agents of SafetyFlow ensure process and cost controllability while integrating human expertise into the automatic pipeline. The final constructed dataset, SafetyFlowBench, contains 23,446 queries with low redundancy and strong discriminative power. Our contribution includes the first fully automated benchmarking pipeline and a comprehensive safety benchmark. We evaluate the safety of 49 advanced LLMs on our dataset and conduct extensive experiments to validate our efficacy and efficiency.

cs.CL↗

arXiv2026

Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation

Generalizable embodied manipulation remains difficult to achieve through pretraining alone, due to unseen physical conditions in the real world. We argue that robots need to learn from their own physical interactions on the fly during real-world deployment and use this knowledge to inform subsequent actions. We present Zeva, the first framework that enables in-context learning from a robot's own physical interaction experience while keeping the policy model frozen. Zeva employs a Causal Interaction Extractor to encode an executed action and its induced state change into a causal interaction signal, which is stored in a dual-timescale causal memory. For subsequent actions, relevant causal interaction signals are retrieved from memory and injected into the frozen policy model as context. Experiments in simulation and real-world manipulation demonstrate that Zeva achieves the best performance among the compared frontier VLAs and WAMs and, more importantly, enables self-evolution during deployment without gradient updates. Its success rate continues to improve as the robot accumulates interaction experience. Furthermore, the acquired interaction experience can generalize across tasks.

cs.RO↗

arXiv2026

Zeva-Ego: Egocentric Mid-Training with In-Context Causal Learning for Robot Manipulation

Egocentric video offers a scalable source of physical interaction experience, yet translating it into robot-executable knowledge and enabling continual adaptation remain challenging. We introduce Zeva-Ego, a unified framework that learns physical priors from human experience and evolves through robot interaction. An Action-Centric Encoder (ACE) converts egocentric visual transitions into action-centered supervision for VLA mid-training, while In-Context Causal Learning (ICCL) enables parameter-free adaptation from action-effect feedback at deployment. Scaling Ego data to 10K hours improves RoboTwin success from 63.8% to 75.3%, matching 2K hours of robot demonstrations (74.7%), corresponding to an empirical data ratio of roughly 4-5:1. With accumulated interaction experience, ICCL further improves success from 58% to 89% within four attempts without parameter updates. These results demonstrate a scalable path toward embodied intelligence that learns from human experience and continuously improves through its own interaction.

cs.RO↗

arXiv2026

Neutron Star Merger Universality Relations for a Quark--Hadron Crossover Equation of State

A crossover transition from hadronic to quark matter during the merger of neutron stars can lead to interesting observational consequences in the emergent gravitational radiation. In particular, the increased pressure in the crossover density region ($2-5$ times the nuclear saturation density) can lead to an extended duration of high-frequency ($\sim 2-3.5$ kHz) gravitational wave emission during the post-merger epoch. We study the universality relations for a variety of equations of state with and without a crossover transition to a quark mater based upon the QHC19 and QHC21 EoSs. We find that the long-duration post-merger gravitational-wave emission is more pronounced in the QHC21 EoS. We then consider universality relations among various observable quantities that have been identified in simulations with hadronic equations of state. We show that there are deviations from the universality relations for some quantities that may ultimately provide an observational signal that such a transition has occurred.

astro-ph.HE↗

arXiv2026

RFID-based Relative Localization for Nanodrone Swarm Platooning

A nanodrone swarm is formulated by multiple lightweight and low-cost nanodrones to perform tasks in very challenging environments. Therefore, it is essential to recognize the relative positions of the nanodrones in the swarm for accurate and safe platooning. However, the camera or LiDAR sensors suffer from the line-of-sight (LoS) sensing. This paper presents the design, implementation, and evaluation of RFDrone, a system that can enable the relative positioning for the nanodrone swarm using the commodity passive RFID tags attached to each nanodrone for safe and accurate platooning. To do so, we propose a novel algorithm for spatial-temporal phase profiling based on the swarm platooning nature and further design a machine-learning model to predict the relative position of the tagged nanodrones in the swarm. Our experiments show that RFDrone can accurately estimate the relative positions of nanodrones in the swarm, whereby the 3D geometry of the nanodrone swarm pattern is easily obtained across different environments, tag diversity, and swarm pattern diversity.

cs.HC↗

arXiv2026

Event Interaction in Low-Rank Bottlenecks for Temporal Relation Extraction

Temporal relation extraction determines whether an event occurs before, after, or simultaneously with another event, and therefore relies on accurately modeling how the two events interact. Mainstream systems achieve this by concatenating event spans or using shallow fusion, which works well when all model parameters are trainable. However, in parameter-efficient fine-tuning, low-rank bottlenecks restrict information flow and prevent these interaction signals from passing through, leading to clear performance drops. To address this limitation, we propose a theoretically grounded architecture, Convolutional Bottleneck Interaction (CBI), which first applies lightweight depthwise convolution to enhance event representations and then uses element-wise multiplication to capture effective event-event interactions inside the bottleneck. Across five datasets and seven backbone models in the Adapter and LoRA settings, CBI provides consistent and substantial gains, up to +31.7 micro F1, while adding minimal computational cost, showing that explicit interaction inside low-rank spaces is crucial for temporal relation extraction. The code is available at https://github.com/VRCMF/CIF.git.

cs.CL↗

arXiv2026

A Computational Pipeline for Patient-Specific Modeling of Thoracic Aortic Aneurysm: From Medical Image to Finite Element Analysis

The aorta is the body's largest arterial vessel, serving as the primary pathway for oxygenated blood within the systemic circulation. Aortic aneurysms consistently rank among the top twenty causes of mortality in the United States. Thoracic aortic aneurysm (TAA) arises from abnormal dilation of the thoracic aorta and remains a clinically significant disease, ranking as one of the leading causes of death in adults. A thoracic aortic aneurysm ruptures when the integrity of all aortic wall layers is compromised due to elevated blood pressure. Currently, three-dimensional computed tomography (3D CT) is considered the gold standard for diagnosing TAA. The geometric characteristics of the aorta, which can be quantified from medical imaging, and stresses on the aortic wall, which can be obtained by finite element analysis (FEA), are critical in evaluating the risk of rupture and dissection. Deep learning based image segmentation has emerged as a reliable method for extracting anatomical regions of interest from medical images. Voxel based segmentation masks of anatomical structures are typically converted into structured mesh representation to enable accurate simulation. Hexahedral meshes are commonly used in finite element simulations of the aorta due to their computational efficiency and superior simulation accuracy. Due to anatomical variability, patient specific modeling enables detailed assessment of individual anatomical and biomechanics behaviors, supporting precise simulations, accurate diagnoses, and personalized treatment strategies. Finite element (FE) simulations provide valuable insights into the biomechanical behaviors of tissues and organs in clinical studies. Developing accurate FE models represents a crucial initial step in establishing a patient-specific, biomechanically based framework for predicting the risk of TAA.

eess.IV↗

arXiv2026

Hosting Capacity Assessment of Data Centers with Voltage Ride-Through Capability in Power Systems

Large data centers are emerging as concentrated, power-electronic grid loads whose abrupt disconnection or transfer to on-site backup supply during voltage disturbances can remove large demand from the power system, and may create a system-level stability problem. Their interconnection feasibility therefore depends not only on steady-state thermal and voltage limits, but also on whether internal power-conditioning systems can maintain IT service while limiting customer-initiated load reduction. This paper presents a voltage ride-through (VRT)-aware data center and grid co-planning framework that couples transmission-level fault simulation with an internal data center ride-through model. Python-based dynamic simulations generate point-of-interconnection (POI) voltage trajectories under selected network faults, and the resulting waveforms drive an internal model incorporating IT and cooling-load dynamics, DC-link, Uninterruptible Power Supply (UPS) response, and converter apparent power limits. The IEEE 118-bus case study shows that internal VRT capability can become a binding interconnection constraint: steady-state planning alone can overestimate feasible data center capacity, whereas increased UPS converter headroom progressively restores hosting capacity. Under the reduced-order response models studied, the grid-forming mode provides greater ride-through margin than the current-limited grid-following mode under the same network fault conditions. The results further show that VRT constraints can materially change both the total hosting capacity of data centers and its spatial allocation across candidate interconnection buses.

eess.SY↗