arXiv subjects

Yifan Wang

Publications and source records attributed to Yifan Wang.

At least 109 records · Page 6Linked to original sources

arXiv2026

Forward Modeling of the $δ$ Sct Star V1790 Ori: $Δν$, $Ω$, Resolution and Non-adiabatic Effects

We investigate the role of large separation, rotational correction order, structural resolution, and non-adiabatic effects in modelling the rotating $δ$ Scuti star V1790 Ori. From TESS data, we extract 69 frequencies and determine $Δν\simeq 82$ $μ$Hz. Rotating MESA models are computed at low and high resolution; their pulsation frequencies are calculated with GYRE (adiabatic/non-adiabatic, first-order rotation) and FILOU (adiabatic, second-order rotation). Using $Δν$ as a structural constraint is necessary to reduce model degeneracy. For the selected minimum-misfit reference model, considering only the 40 modes with consistent $(n,\ell,m)$ labels, the RMS$_{40}$ theoretical frequency differences are 0.442 $μ$Hz (resolution), 0.062 $μ$Hz (non-adiabatic), and 2.962 $μ$Hz (GYRE vs FILOU); including all 48 frequencies gives RMS$_{48}$ values of 1.033, 2.326, and 3.931 $μ$Hz. Relative to observations, higher resolution reduces residuals from 4.457 to 4.387 $μ$Hz (RMS$_{40}$) and from 4.715 to 4.682 $μ$Hz (RMS$_{48}$); non-adiabatic effects change them marginally to 4.381 and 4.673 $μ$Hz. FILOU gives the largest residuals: 5.331 $μ$Hz (RMS$_{40}$) and 5.270 $μ$Hz (RMS$_{48}$). Second-order rotation produces the largest frequency shifts, but improving agreement with observations requires denser grids and self-consistent FILOU optimisation. The 260.672 $μ$Hz peak -- previously identified as the fundamental radial mode -- shows uncertain identification. The results should be interpreted as diagnostics of modelling systematics and mode-identification robustness.

astro-ph.SR↗

arXiv2026

DECIS: Dual-Evidence Corrective Verification for Interpretable Strabismus Diagnostic Decision-Making

Strabismus is a common ocular disorder that requires fine-grained subtype diagnosis for individualized treatment planning. However, existing deep learning methods mainly provide diagnostic predictions without transparent reasoning, while recent large vision-language models (LVLMs), although promising for joint image understanding and report generation, remain highly prone to hallucination in this evidence-sensitive and rule-driven medical task. To address these challenges, we propose DECIS, a Dual-Evidence Corrective verification for Interpretable Strabismus diagnostic decision-making framework. DECIS transforms black-box end-to-end generation into a structured diagnostic process consisting of candidate hypothesis generation, dual-evidence constrained context, evidence-based corrective verification, and report generation. Specifically, we introduce a Dual-Evidence Constrained Context (DECC) mechanism that jointly organizes visual evidence from the photograph of the nine cardinal positions of gaze and evidence-based clinical diagnostic rules into a constrained context for reliable diagnostic reasoning. We further develop an Evidence-Based Corrective Verification (EBCV) mechanism that verifies whether the current diagnostic hypothesis is supported by visual evidence, heatmap-based visual cues, and evidence-based clinical diagnostic rules. Hypothesis refinement is triggered when inconsistency is detected. Experiments on a fine-grained strabismus benchmark demonstrate that DECIS not only outperforms other state-of-the-art diagnostic systems, improving the weighted F1 score from 72.0% to 91.3%, but also improves the clinical reliability (consistency, alignment, and completeness) of generated diagnostic reports. These results demonstrate that DECIS provides an effective solution for building accurate, evidence-based, and clinically interpretable strabismus diagnosis systems.

cs.CV↗

arXiv2026

Conan-embedding-v3: Fusing Modality-Specific Models for Omni-Modal Embedding

Omni-modal retrieval promises a single embedding space for text, image, video, document, and audio inputs, but building such a unified retriever is difficult since these modalities differ in data distribution, architecture, and optimization dynamics. In this work, we present Conan-embedding-v3, a decouple--fuse--recover framework for omni-modal retrieval. Conan-embedding-v3 first trains modality specialists independently and fuses their task vectors into a single dense backbone, a strategy we call Decoupled Specialist Fusion. We show that this fusion composes visual, video, and document retrieval capabilities, but also exposes a failure mode for projector-based modalities: when audio is attached through an external encoder and projector, fusing the backbone leaves the projector calibrated to the audio-specialist backbone, causing a large audio retrieval regression despite copying all audio-specific modules unchanged. We call this failure Projector Drift. To repair it, Conan-embedding-v3 applies Projector Recovery (i.e., full-parameter fine-tuning of the projector while keeping the backbone frozen) followed by balanced multi-modal rehearsal. The resulting model supports these retrieval pathways in one backbone, achieving 74.9 scores on MMEB while obtaining 55.61 on the 30-task MAEB audio suite.

cs.MM↗

arXiv2026

From Interpretation to Compilation: Compilation-Based Execution of Semantic Operators [Vision]

Semantic operator systems extend data processing with natural-language interfaces, supporting operations such as semantic filtering, mapping, and joining. Existing systems commonly execute these operators through interpretation-based execution: for each row, record, or candidate pair, an LLM is invoked to interpret the semantic intent and produce an output. Although expressive, this places expensive LLM calls inside the data-processing loop, causing high latency, monetary cost, and limited scalability. We propose compilation-based execution of semantic operators. Instead of using an LLM as a runtime interpreter for every data item, we invoke it once during compilation to translate a semantic operator specification into deterministic executable code. The generated code serves as a compiled physical operator that approximates the behavior of the original LLM-based operator and runs locally over the dataset without per-row or per-pair LLM calls. We instantiate this approach for semantic filter, semantic map, and semantic join operators, compare it with LLM-based interpreted execution, and integrate it into an existing semantic operator system. Preliminary results show that compilation-based execution substantially reduces execution time and LLM calls while preserving much of the output quality. More broadly, we argue that semantic operator systems should treat LLMs not only as runtime executors, but also as semantic compilers that generate efficient executable plans, opening new research opportunities at the intersection of database query processing, program synthesis, and LLM-powered data systems.

cs.DB↗

arXiv2026

Direct Numerical Simulation of Fully Developed Turbulent Channel Flow Based on the Corrected Navier-Stokes Equations

Direct numerical simulations (DNS) of fully developed turbulent channel flows at Re_tau = 550 were performed to investigate the corrected Navier-Stokes (CNS) equations. Grounded in the fluid kinematics of Rortex, the CNS abandons the Stokes' isotropic hypothesis and applies the shearing-only constitutive relation by explicitly eliminating the controversial stretching terms in the stress tensor. Comparisons with the DNS data from the traditional Navier-Stokes (TNS) suggest that the CNS inherently rectifies the near-wall momentum transport. The removal of stretching-induced dissipation shifts the inner- and buffer-layer boundaries towards the wall and effectively suppresses the overshoot in the mean velocity profile in TNS. The turbulence statistics demonstrate a multiscale kinetic energy redistribution that intensifies the near-wall production-dissipation cycle. Furthermore, the topological delineation of instantaneous coherent structures, namely the discovery and definition of rotational/non-rotational interface (RNRI) via the velocity gradient tensor (VGT) discriminant (Delta = 0), confirms that the CNS is capable of capturing the highly complex and interwoven vortical structures. Ultimately, the spectral proper orthogonal decomposition (SPOD) unveils and elucidates that the shearing-only mechanisms intrinsically modulate the spatiotemporal energy cascade, promoting denser and more inclined vortices while enhancing turbulence intermittency by fragmenting the coherent packets. Overall, by isolating and detecting the shearing-only mechanism with physics purity, the DNS based on the CNS provides a more refined perception of the intrinsic dynamics of wall-bounded turbulence, offering a more physical soundness model to capture the interactions among the multiscale coherent structures and the improved capability in predicting the wall-bounded turbulence.

physics.flu-dyn↗

arXiv2026

Multimodal 3D LUT Generation via StatLUT with Statistical Features for Photorealistic Style Transfer

Photorealistic Style Transfer (PST) aims to transfer the color and tonal style of a reference to a content image while strictly preserving its structural integrity. However, existing deep learning-based methods inherently suffer from semantic entanglement caused by pre-trained image encoders, leading to unnatural spatial distortions. Moreover, current pixel-level mapping paradigms often ignore color gamut topology, resulting in color banding, while also lacking the multimodal capability for intuitive text-driven control. To address these bottlenecks, we propose StatLUT, an innovative multimodal framework for 3D LUT generation. First, we bypass traditional encoders and introduce a Lab-Extractor to derive spatially-agnostic statistical features, fundamentally decoupling color distributions from structural semantics to ensure artifact-free rendering. Second, we formulate LUT generation as a Transformer-based Seq2Seq translation task, utilizing a Multi-dimensional Residual Mapper (MR-Mapper) to predict topologically smooth 3D LUTs. Finally, to break the single-modal barrier, we propose the H-Diffuser, a lightweight Diffusion Transformer that directly synthesizes statistical features from natural language prompts, enabling flexible text-driven color grading. Extensive experiments on standard benchmarks demonstrate that StatLUT significantly outperforms state-of-the-art methods in both visual quality and quantitative metrics, pioneering a highly robust and flexible paradigm for multimodal photorealistic style transfer.

cs.CV↗

arXiv2026

NAMD: Virtual Follow-up Computed Tomography Synthesis via Nodule-Aligned Multimodal Diffusion Models for Early Lung Cancer Diagnosis

Lung cancer remains the leading cause of cancer-related mortality worldwide, with survival outcomes critically dependent on early and accurate detection. When low-dose computed tomography (LDCT) findings are indeterminate, clinicians typically defer diagnosis pending follow-up CT imaging obtained up to 12 months later, inevitably delaying treatment for patients with malignant nodules. To address this clinical gap, we propose Nodule-Aligned Multimodal (Latent) Diffusion (NAMD), a novel generative framework that synthesizes one-year follow-up nodule CT images conditioned on the baseline CT scan, quantitative nodule biomarkers, and patient-level Electronic Health Records (EHR), enabling timely prediction of nodule malignant progression without requiring actual follow-up scans. NAMD introduces two key contributions: (i) a nodule-aligned latent space regularized so that embedding distances reflect clinically meaningful biomarker changes, and (ii) an LLM-driven multimodal conditioning mechanism encoding heterogeneous EHR data into the diffusion backbone. Evaluated on the National Lung Screening Trial (NLST), our method's synthetic follow-up images achieve an AUROC of 0.805 and an AUPRC of 0.346 for lung nodule malignancy prediction, outperforming both the baseline LDCT performance without virtual follow-up generation, and existing state-of-the-art conditional generation methods, while maintaining competitive image quality. These findings suggest that NAMD enables earlier and more accurate lung cancer diagnosis by capturing clinically meaningful features of nodule progression.

cs.CV↗

arXiv2026

OpenSPM: An Environment-Transferable Robotic Key Spatial Pose Memory and Closed-Loop High-Frequency Flow-Matching Action Generation Model

Open-environment tabletop robotic manipulation requires systems to possess semantic understanding, precise geometric pose estimation, and high-frequency action generation. While end-to-end vision-language-action (VLA) models excel at semantic generalization, they often lack explicit geometric constraints for fine-grained tasks and require costly training. To bridge the gap between high-level semantics and low-level physical execution, we propose OpenSPM, an open environment spatial persistent memory framework consisting of spatial pose memory and flow-matching action generation model. OpenSPM first leverages semantically conditioned 3D perception and Kalman filtering to track continuous 6D poses. It then extracts key spatial poses from human demonstrations, keeping them as transferable, object-centric spatial persistent memory entries. During inference, OpenSPM retrieves relevant memory entries in terms of natural language instructions, transfers the spatial poses to new scenes using SE(3) transformations, and generates high-frequency action chunks via a lightweight conditional flow-matching model. Combined with real-time proprioceptive state feedback and terminal residual correction, the system effectively suppresses trajectory error accumulation. Evaluated on ten LIBERO-GOAL tasks, OpenSPM achieves an 85.6% success rate and an equivalent control frequency of 1033.3 Hz, while requiring minimal inference AI computing power. Extensive ablations illustrate that structured spatial persistent memory and closed-loop residual correction play a crucial role in reliable, high-frequency robotic manipulation.

cs.RO↗

arXiv2026

CertMix: Certified, Data-Efficient Metamaterial Design by Affine Mixing of Aligned Neural-Implicit Weight Spaces

Inverse design of mechanical metamaterials seeks a periodic unit cell whose homogenized elastic properties meet a prescribed target, but current learning-based methods are data-hungry, mostly interpolative, and provide no guarantee that the generated design satisfies the specification. We introduce CertMix, a data-efficient framework that represents each exemplar unit cell as a small periodic neural implicit field, specifically a SIREN signed-distance decoder overfit from a shared anchor, so that exemplar weight vectors become aligned and directly comparable. The key observation is that, in this aligned weight space, the homogenized elasticity tensor is approximately linear in the mixing coefficients. Targeted design therefore reduces to a small constrained affine-mixing problem solved with a differentiable periodic homogenizer in the loop. Negative coefficients enable extrapolation beyond the exemplar range, a linearity-mismatch trust region keeps blends valid, and split-conformal calibration converts the mismatch signal into a distribution-free certificate on achieved-property error. From as few as 50 exemplars, CertMix attains a scaled property error of $10^{-4}$, roughly two to three orders of magnitude below conditional generative baselines trained on 1000 cells. It remains accurate far outside the exemplar range, is $57\times$ faster than per-target topology optimization while avoiding checkerboards and enclosed voids, and extends to spatially graded fields, 3D triply periodic surfaces, and a certified running-shoe midsole application.

cs.LG↗

arXiv2026

ThermoForce: A Physics-Structured Interventional World Model for Building HVAC Control

Model predictive control (MPC) of building HVAC systems needs thermal models that answer a causal question: what indoor temperature, energy use, and comfort will result if a control action is applied? Time-series foundation models (TSFMs) can forecast passive building trajectories with strong zero-shot skill, but high factual accuracy does not imply valid response to control interventions. We show that an observational grey-box model with the best passive accuracy predicts cooling effects with the wrong sign, and that adding control and weather covariates to a TSFM does not fix intervention response. We introduce ThermoForce, a control-ready interventional thermal world model that keeps a TSFM frozen as a passive free-response prior and learns a compact, physics-structured forced-response operator for the causal effect of HVAC actuation. The operator is monotone in the control input by construction, is identified from one to three days of control excitation, and composes with the free response into a counterfactual-capable world model. Across paired EnergyPlus heating and cooling interventions, ThermoForce attains the lowest intervention-effect error and correct effect sign where covariate-TSFM, observational grey-box, and distillation baselines fail. Embedded in MPC on the BOPTEST benchmark, it reduces thermal discomfort by 33--84\% relative to the native controller across three two-week windows while simultaneously reducing energy, using a frozen backbone, 195 trainable parameters, and CPU-only computation. ThermoForce reframes foundation models for building control: passive prediction and forced intervention response must be structurally separated for a model to be control-ready.

eess.SY↗

arXiv2026

ProLaViT: Learning Progressive Latent Visual Thoughts in Structured Latent Space

Multimodal Large Language Models (MLLMs) have achieved remarkable progress but still struggle with complex visual reasoning tasks requiring multi-step perception and logical deduction. While explicit visual generation incurs prohibitive computational costs, existing latent approaches often rely on external experts or lack rigorous cognitive logic. In this paper, we introduce ProLaViT (Progressive Latent Visual Thought), a framework empowering MLLMs to perform structured visual derivation in the continuous latent space. Unlike works dependent on heterogeneous external models, ProLaViT leverages an endogenous self-distillation mechanism, utilizing the model's own visual encoder to supervise latent thoughts. To facilitate this, we construct a scalable programmatic synthesis pipeline enabling the model to internalize algorithmic precision without inference time tools. We design two reasoning paradigms: (1) Coarse-to-Fine Causal Chain for spatial tasks, guiding attention from global context to local targets. (2) Dialectical Reasoning Chain for logical tasks, incorporating counter-factual thinking for verification. Furthermore, we propose a Distance-Weighted Diversity Loss to impose topology-aware constraints, preventing feature degeneration by enforcing semantic distinctiveness. Extensive experiments demonstrate that ProLaViT outperforms baselines on vision-centric benchmarks, achieving superior accuracy and interpretability with high efficiency.

cs.CV↗

arXiv2026

Beyond Heuristics: A Standardized Real2Sim Pipeline for Physical Human Robot Interaction in Human-in-the-Loop Simulation

The aging global population drives demand for assistive robots, yet the safety risks and costs of physical testing make Human-in-the-Loop (HITL) simulation an attractive alternative. Its fidelity for coupled systems, however, is limited by interaction models whose impedance parameters are tuned heuristically rather than identified from data. We present a Real2Sim pipeline that identifies the coupled Physical Human-Robot Interaction (pHRI) dynamics of a pelvis--strap interface on an overground mobile balance assistant. The interface is modeled as a 6-DoF viscoelastic mechanism whose 12 directional stiffness and damping parameters are identified per subject via Covariance Matrix Adaptation Evolution Strategy (CMA-ES), using the user's ``Safe \& Comfortable'' feedback as a reproducible operating point that resolves harness-tightness ambiguity across anthropometrics. An intraclass-correlation analysis over a five-subject cohort separates shareable from subject-specific parameters, yielding a set of prior parameters derived from the existing data. Deploying this prior configures a previously unseen subject by refining only 5 of the 12 parameters. The calibrated model then reproduces the real interaction envelope and induces biomechanically accurate gait adaptations in the Human Digital Twin (HDT). Overly compliant and overly stiff settings, by contrast, fail as extreme settings, confirming a correct operating point that no heuristic tuning procedure can reliably select. The pipeline thus improves HITL simulation fidelity and supports the Human Digital Twin as a predictive tool for pre-clinical verification of personalized controllers.

cs.RO↗

arXiv2026

DRBA: Dynamic Robotic Balance Assistant -- An assist-as-needed gait and balance rehabilitation robot for versatile training

The decline of human balance control due to aging and pathological conditions increases fall risk, a major concern in geriatric care and rehabilitation. Gait training is essential for balance recovery, enhancing walking ability and postural control. However, existing overground robotic gait trainers have limitations: body weight support systems are bulky and impractical for daily use, while end-effector-based systems often compromise transparency, altering natural gait dynamics. This paper presents the Dynamic Robotic Balance Assistant (DRBA), a novel gait trainer providing assist-as-needed body weight and balance support for various training scenarios. DRBA integrates a 3-degree-of-freedom (3-DoF) robotic arm for pelvic support with flexible motion, a compact sit-to-stand assistance module, and user-following and fall detection algorithms to ensure minimal interference and responsive support. Experimental results demonstrated high transparency, with minimal impact on natural gait dynamics. A patient trial with nine elderly patients with varying medical conditions and balance impairments (ranging from severe to mild) further validated DRBA's effectiveness. The results showed that DRBA-assisted training increased step length and walking speed compared to therapist-assisted gait training. Additionally, DRBA enabled users to perform tasks beyond their unaided ability, expanding rehabilitation possibilities. These findings highlight DRBA's potential to enhance rehabilitation outcomes by facilitating higher training intensity and enabling task-oriented exercises.

cs.RO↗

arXiv2026

Mobility Safe Adaptive Reserve Certification for Electric Vehicle Hydrogen Bus and Building Resilience Hubs

Zero-emission mobility depots are becoming resilience assets because one site can host EV charging, hydrogen-bus operation, stationary conversion equipment, and nearby critical-building backup. The key question is not raw outage export capacity: hydrogen exported to buildings can strand buses, EV availability is stochastic, and building demand shifts seasonally. We introduce a mobility-safe reserve certification framework for a coupled EV, hydrogen-bus, and critical-building hub. It combines a physics-hybrid universal differential equation building-load twin, one-sided split conformal reserve calibration, adaptive conformal inference for seasonal drift, and a mobility-first scheduling rule that protects post-event bus service before assigning hydrogen to buildings. Evaluation uses 495,221 real EV charging sessions across eight regions, AC Transit GTFS-derived hydrogen-bus service days, and EnergyPlus 25.2 simulations under real TMY3 weather. Across 66,816 held-out outage scenarios, a mobility-blind hydrogen-export policy served 39.2\% of building demand but protected buses in 0\% of cases and caused a 426.7 kg mean bus-hydrogen shortfall. A nominal mean-resource promise delivered only 45.4\% of commitments. The certified mobility-first policy was the only tested policy to achieve 100\% commitment delivery, 100\% bus protection, and zero mean bus-hydrogen shortfall, while serving 20.5\% of critical-building demand. Under a summer-to-winter load shift, adaptive conformal inference raised late-period empirical coverage from 0.687 to 0.831 and reached 0.891 overall coverage against a 0.90 target with lower mean reserve than static split conformal. Across 12 building/seed drift runs, it kept low late-coverage variability and the lowest mean reserve. These results show that resilience value in shared zero-emission hubs depends on service-aware certification, not raw export capacity alone.

eess.SY↗

arXiv2026

Asymmetric Trading Prophets

The "Trading Prophet" problem challenges an online trader to maximize its profit by buying and selling assets under stochastic prices and capacity constraints, competing against an offline prophet with full foresight. In previous work, each arriving asset was assumed to have a single price $p_t$, and the trader was allowed to either buy a copy at this price (subject to having available capacity), or sell a copy (if it already held at least one copy in hand). However, this abstraction can fail to capture the structural asymmetry of decentralized dealer-based markets, where buying and selling opportunities could be distinct, and driven by individual preferences. To address this, we introduce the Asymmetric Trading Prophets problem, where at each timestep the trader observes a price tuple $(b_t, s_t)$ -- representing the cost to buy, and the revenue from selling at this timestep. Importantly, the $(b_t,s_t)$ tuple could be potentially arbitrarily correlated. We provide the first competitive analysis for this asymmetric trading prophets problem, characterizing the achievable profit based on the trader's capacity $B$ and initial inventory $B_0$. For the unit-capacity case of $B=1$, we design online algorithms that achieve constant competitive ratios for both i.i.d. and non-i.i.d. distributions on the price tuples, when the trader has one initial copy ($B_0=1$). For the general capacity case where $B$ can be large, we give algorithms for i.i.d. distributions that achieve a competitive ratio of $1 - Θ(\log B_0/\sqrt{B_0})$. Finally, for the symmetric case (where the price tuple satisfies $b_t=s_t$), we improve this to get a competitive ratio of $1 - O(\log B/\sqrt{B})$, demonstrating that the performance approaches optimality as the capacity increases. We show that both ratios are tight up to a logarithmic factor.

cs.DS↗

arXiv2026

VisReason: A Large-Scale Dataset for Visual Chain-of-Thought Reasoning

Chain-of-Thought (CoT) prompting has proven remarkably effective for eliciting complex reasoning in large language models (LLMs). Yet, its potential in multimodal large language models (MLLMs) remains largely untapped, hindered by the absence of large-scale datasets that capture the rich, spatially grounded reasoning intrinsic to visual understanding. Existing visual-CoT resources are typically small, domain-specific, or lack the human-like stepwise structure necessary for compositional visual reasoning. In this paper, we introduce VisReason, a large-scale dataset designed to advance visual Chain-of-Thought reasoning. VisReason comprises 489K annotated examples spanning four diverse domains, each featuring multi-round, human-like rationales that guide MLLMs through interpretable visual reasoning steps. Building upon this, we curate VisReason-Pro, a 165K subset produced with a stronger expert-level GPT annotator, enriched with detailed reasoning traces and 3D spatial grounding via depth-informed annotations. Fine-tuning the state-of-the-art Qwen2.5-VL model on VisReason and VisReason-Pro yields substantial improvements in step-by-step visual reasoning accuracy, interpretability, and cross-benchmark generalization. These results demonstrate that VisReason equips MLLMs with more systematic and generalizable reasoning capabilities. We envision VisReason as a cornerstone for cultivating human-like visual reasoning, paving the way toward the next generation of multimodal intelligence.

cs.CV↗

arXiv2026

APAO: Bridging the Training-Inference Gap in Generative Recommendation via Adaptive Prefix-Aware Optimization

Generative recommendation has recently emerged as a promising paradigm for sequential recommendation. It formulates the task as an autoregressive generation process, predicting tokens of the next item conditioned on user interaction histories. Existing generative recommendation models are typically trained with token-level likelihood objectives such as cross-entropy loss, while employing beam search during inference to generate ranked candidates. However, this leads to a fundamental training-inference inconsistency: standard training assumes ground-truth tokens are always available, while beam search prunes low-probability branches during inference, causing the correct item to be prematurely discarded when its prefixes receive low scores. To address this issue, we propose the Adaptive Prefix-Aware Optimization (APAO) framework, which introduces prefix-level optimization losses to better align the training objective with the inference setting. Furthermore, we design an adaptive worst-prefix optimization strategy that dynamically focuses on the most vulnerable prefixes during training, thereby enhancing the model's ability to retain correct candidates under beam search constraints. We provide theoretical analyses to demonstrate the effectiveness and efficiency of our framework. Extensive experiments show that APAO consistently alleviates the training-inference inconsistency and improves performance across generative recommendation backbones. The source code is publicly available at https://github.com/yuyq18/APAO.

cs.IR↗

arXiv2026

xGR: Efficient Generative Recommendation Serving at Scale

Recommendation system delivers substantial economic benefits by providing personalized predictions. Generative recommendation (GR) integrates LLMs to enhance the understanding of long user-item sequences. Despite employing attention-based architectures, GR's workload differs markedly from that of LLM serving. GR typically processes long prompt while producing short, fixed-length outputs, yet the computational cost of each decode phase is especially high due to the large beam width. Furthermore, since the beam search involves a vast item space, the sorting overhead becomes particularly time-consuming. We propose xGR, a GR-oriented serving system that meets strict low-latency requirements under high-concurrency scenarios. First, xGR unifies the processing of prefill and decode phases through staged computation and separated KV cache. Second, xGR enables early sorting termination and mask-based item filtering with data structure reuse. Third, xGR reconstructs the overall pipeline to exploit multi-level overlap and multi-stream parallelism. The experiments on real-world datasets demonstrate that xGR achieves at least 2.89x throughput compared to the state-of-the-art baseline under strict latency constraints.

cs.LG↗