arXiv subjects

Xinyi Li

Publications and source records attributed to Xinyi Li.

At least 19 recordsLinked to original sources

arXiv2026

HalifaxDT: A Wireless Digital Twin from Open Geospatial Data

Wireless digital twins can support site-specific analysis, planning, and experimentation for future wireless networks. However, constructing them at city scale remains challenging when accurate 3D city models are unavailable. This paper presents HalifaxDT, a wireless digital twin of the Halifax Peninsula in Nova Scotia, Canada, constructed from heterogeneous open geospatial and spectrum data. HalifaxDT combines building footprints, LiDAR elevation products, building metadata, and spectrum licensing records through a workflow that reconciles multiple sources of building-height information while preserving the provenance of geometry decisions. The resulting terrain, buildings, and gateway metadata are integrated into Sionna RT for wireless simulation. We evaluate HalifaxDT through two use cases. The first compares its coverage predictions with a reference derived from field measurements and with an analytical baseline. The second uses the digital twin to predict the received signal under normal operating conditions and detect interference. The coverage results show that HalifaxDT better preserves the spatial structure of the measured radio map than the analytical baseline. The interference study shows that deviations from the predicted reference can reveal interference that is difficult to detect from received power alone. We also identify current fidelity limitations, including simplified material representation, missing vegetation, and the need for broader RF calibration and synchronization with live measurements.

cs.NI↗

arXiv2026

Efficient and Robust Absolute Pose Estimation via Gravity-Prior-Driven Transformation Decoupling and Pose Refinement

Estimation of the absolute pose of an object is an essential task for various robotic applications. Recently, incorporating gravity direction as prior information has emerged as a popular approach to simplify absolute pose estimation. However, developing a robust and efficient algorithm to solve this challenging problem remains a difficult question due to large amounts of mismatches. In addition, obtaining an accurate pose solution from selected inlier correspondences with gravity prior is still a research gap. In this paper, we propose a novel transformation strategy that exploits geometric relations derived from the gravity prior. Through transformation decoupling, the original 6 degrees of freedom (DoF) absolute pose estimation problem is simplified into a 4-DoFs problem: 1-DoF for the rotation angle and 3-DoFs for translation, significantly improving the efficiency. For the 1-DoF rotation angle, we apply a one-dimensional global voting algorithm for optimal estimation. Once the optimal rotation is obtained, the mismatched correspondences are preliminarily filtered, and translation estimation, a linear problem, can be easily solved. Furthermore, to obtain accurate pose results, we introduce a novel pose refinement algorithm to enhance the accuracy of both rotation and translation. Extensive experiments on synthetic data and three publicly available real-world datasets (TUM RGB-D, ETH3D, and RobotCar) demonstrate that the proposed method achieves stronger performance compared to existing state-of-the-art (SOTA) approaches. To further validate our method, we integrated it into ORB-SLAM2. The results on the KITTI dataset show it effectively reduces drift and improves trajectory alignment during relocalization. The source code will be released upon acceptance.

cs.CV↗

arXiv2026

Beyond Legal Spacing: A Residual-Aware Characterization of Entangling-Zone Spacing in Neutral-Atom Compilation

Neutral-atom processors rely on spatially arranged qubit arrays and parallel Rydberg entangling gates for scalable execution. Their compilers enforce geometric spacing rules for simultaneous gates, yet legal separation does not make residual van der Waals coupling disappear. This paper studies that gap between geometric legality and residual noise by treating entangling-zone spacing as a cross-layer reliability-parallelism variable anchored to experimental neutral-atom geometry. We combine fixed-schedule residual replay, surface-code simulation with matched correlated decoding, and fresh recompilation to connect spacing to physical residual exposure, logical reliability, and makespan cost. The results show that near-floor spacing can produce structured correlated exposure that is visible both at the physical layer and, in the tightest case, after quantum error correction (QEC). Modest geometric slack strongly suppresses this residual contribution, but the timing cost of looser spacing is mediated by placement and scheduling rather than by a simple monotonic slowdown. These findings distinguish hardware legality from residual-noise safety and motivate spacing-aware compiler evaluations that report physical geometry, QEC absorption, and scheduling cost together.

quant-ph↗

arXiv2026

Where Atom Loss Lands Matters: Decoder-Aware Risk Deposition in Neutral-Atom QEC

Neutral-atom arrays are emerging as a leading platform for scalable quantum error correction (QEC). Qubits are routed and reused across the array, while detected loss is reported to the decoder as erasure information. Existing neutral-atom compilers optimize this movement, including routing, shuttling, and reuse, and often model loss through scalar exposure costs. Yet total exposure is an incomplete statistic for erasure-corrected QEC. It captures how much loss occurs, but not where it lands on the code, which we call its deposition. Under the same expected atom-loss budget, different deposition patterns over a code patch induce substantially different logical error rates (LER). We formalize this as decoder-aware risk deposition and present CAST, a compiler-side optimization pass that overlays a code-topology sensitivity map on a role-indexed exposure ledger and minimizes a decoder-weighted harm objective under a comparable-exposure constraint, using only local route, role, and seam-cooling actions. Across surface-code memory, physical-scale architecture models, lattice surgery, and decoder-mismatch checks, CAST lowers LER relative to topology-blind exposure minimization, improving on it in 35 of 48 physical-scale settings and by as much as 5.3x where exposure is heterogeneous and routing has slack. The largest gains occur when high exposure and high decoder sensitivity are initially misaligned, giving CAST room to redirect risk toward lower-impact code roles. CAST shows that decoder-aware atom-loss risk deposition can be optimized as a compiler-side pass in neutral-atom QEC.

quant-ph↗

arXiv2026

Beyond Myopic World Models: Long-Horizon End-to-End Training for Direct Future Prediction

World models are expected to support imagination over extended temporal horizons, yet most are still trained through local few-step prediction objectives and deployed by recursively rolling out their own predictions. This creates a fundamental mismatch: few-step losses optimize local transition fidelity, while long-horizon prediction depends on how errors and gradients propagate through the entire trajectory. As a result, transitions with different downstream influence on the endpoint are treated uniformly during training, and small local errors are amplified through recursive inference. We argue that long-horizon accuracy is better achieved by optimizing directly, through an end-to-end endpoint prediction objective. To instantiate this paradigm, we introduce the Direct Prediction World Model (DPWM), a non-recursive architecture that compresses an action sequence of arbitrary length into a single embedding and predicts the endpoint observation in a single forward pass. This design avoids recurrent rollout in both prediction and gradient propagation, making long-horizon end-to-end training practical at horizons where unrolled autoregressive training becomes unstable. Empirically, DPWM substantially improves long-horizon endpoint prediction over recursive world-model baselines on continuous-control and pixel-based benchmarks, with larger gains as the prediction horizon increases. We further show that recurrent baselines benefit similarly when retrained with the same long-horizon endpoint objective, supporting our central claim that the training objective, rather than the particular backbone choice, is the main driver of long-horizon prediction accuracy. Our results suggest that world models can benefit from being trained and evaluated at the temporal scales where they are ultimately used, shifting the focus from local transition modeling toward long-horizon predictive accuracy.

cs.LG↗

arXiv2026

Industrial Dexterity Benchmark: A Hardware-Software Benchmarking Platform for Industrial Dexterous Manipulation

Dexterous manipulation remains a critical bottleneck in industrial automation; tasks such as cable routing, connector insertion, and precision assembly still rely heavily on manual labor despite decades of robotics research. This work presents a progression from classical, modular robotics pipelines toward an end-to-end multimodal imitation-learning framework for industrial dexterous manipulation. As a part of this work, we introduce three key contributions: a set of Industrial Dexterity Benchmark (IDB) boards aimed to mimic datacenter cable management, automotive cable harnesses, and gearbox assembly tasks; a scalable imitation learning framework (DAG-ROS); and a multimodal diffusion-based policy framework (AG-iDP3) that creates models fusing RGB images, point clouds, joint positions, and wrist-frame wrench data. Focusing on the datacenter cable manipulation board, we evaluate the performance of a task involving cleaning a single cable over variations of an end-to-end AI policy using 48 trials per configuration. The best performing configuration, a multimodal expansion Diffusion Policy (DP), includes a multi-view RGB image source passed through an R3M encoder and reaches a 78% grasp and insert combined task success rate. This performance marks a significant improvement over the 36% observed from the single-camera RGB DP baseline. Each of the tested configurations requires only approximately 100 teleoperated demonstrations per task phase. These results indicate that the correct learned policy can outperform classical vision and control robotic methods in robustness, generalization, and deployment efficiency, justifying a shift toward scalable robotic automation for high up-time industrial environments.

cs.RO↗

arXiv2026

An Ontology-Guided Multi-Anchor Graph Retrieval Framework for Traffic Legal Liability Determination

Traffic law liability determination is critical for assigning legal penalties, requiring the simultaneous identification of interdependent statutory provisions across multiple legal dimensions. However, existing retrieval-augmented generation methods suffer from a multi-dimensional retrieval bottleneck: single axis architectures compress complex legal queries into a single pathway, causing interdependent statutory dimensions to be overlooked. To address this, we propose OMAGR, an ontology-guided framework that decomposes queries into ontology-aligned anchors and executes parallel graph retrieval across each dimension, ensuring independent retrieval across dimensions before fusion. To evaluate the proposed method, we created the TrafficLaw-QA dataset, an expert-validated benchmark dataset containing 200 questions and 527 legal provisions. Results show that TrafficOmni-RAG outperforms baselines on Context Precision and Faithfulness metrics. The findings demonstrate that parallel multi-anchor retrieval effectively resolves the multi-dimensional retrieval bottleneck, offering a promising direction for traffic law liability determination research.

cs.CL↗

arXiv2026

Bridging Semantics and Physical Execution: A Neuro-Symbolic Framework for Multi-Pair Robotic Assembly

Multi-pair robotic assembly in unstructured environments faces spatial interference and contact uncertainties. Existing paradigms fail to bridge cognitive decision-making and physical execution, as they either encounter state-space explosion and knowledge bottlenecks or suffer from logical hallucinations and topological conflicts. We propose an end-to-end neuro-symbolic framework that solves the challenge hierarchically: generating optimal subgraphs for each pair, decoupling generality from edge cases, and then resolving cross-pair interferences. Given an eye-on-hand RGB-D assembly scene, the framework extracts semantic instance identity and state while quantifying the scene for divergence calculation. For each pair, optimal subgraph is generated via LLM using barely basic actions to mitigate hallucinations. Supportive actions for edge cases are reasoned and inserted with a lightweight discriminator. Driven by the divergence between the quantified baseline and current scene, it is easily extensible at low cost. Augmented subgraphs are topologically coordinated into global sequences while preserving internal behavioral coherence. Dynamic behavior trees embedding atomic skills close the force-aware execution loop. Offline evaluation on 100 real-world scenes achieves 97.00% global executability, outperforming classical and state-of-the-art planners. Real-robot deployment on a UR3 arm attains 90% success rate with 0.5 mm tolerance under strong interference, demonstrating a unified and verifiable solution for complex autonomous assembly.

cs.RO↗

arXiv2026

"Where is this coming from?" Uncovering Trustworthiness Ideals in AI-powered Peripartum Information Seeking

AI-powered tools increasingly promise to fill information gaps in health, especially in domains like maternal and reproductive health that demand timely, accurate, and actionable information. This is extremely important, as the United States leads peer nations in preventable deaths, with stark racial disparities. However, current AI and NLP-powered systems aim to improve access to vetted maternal health information by routing user queries to a factual response while under-specifying the socio-technical governance structures that shape trust, use, and harm in practice. We report findings from four synchronous focus groups ($n=24$) with three stakeholder groups central to peripartum information support: birthing people, clinicians, and health workers (e.g., doulas, social workers, community health workers) exploring topics around information seeking, experience with current clinical infrastructure, misinformation, and an AI-enabled factual answering tool design probe. Our inductive analysis surfaces a central finding: in high-stakes health contexts shaped by historical inequities, trustworthiness must be inspectable and not asserted. While stakeholders diverge on what makes information credible, they converge on the need for transparency, recourse, and ecosystem complementarity. Based on the discussions, we identify four themes and governance requirements: (1) support for social and identity-based sensemaking, (2) pluralistic verification practices, (3) inspectable governance with recourse mechanisms, and (4) ecosystem-aware integration that avoids shifting burden. Building on these findings, we propose design artifacts that are mistrust-aware and promote principled governance mechanisms for transparent, pluralistic AI systems. Finally, we discuss the implications of our findings for expanding human-AI evaluations and improving the transparency of deployed AI systems.

cs.CY↗

arXiv2026

OpenHalDet: A Unified Benchmark for Hallucination Detection across Diverse Generation Scenarios

Hallucination detection is essential for the reliable deployment of large language models (LLMs). However, existing evaluations face two core challenges: inconsistent inference configuration and evaluation, and limited coverage of downstream domains and tasks. Consequently, reported detector performance is often difficult to compare, reproduce, and generalize beyond specific experimental settings. We introduce OpenHalDet, a unified benchmark for hallucination detection across diverse generation scenarios. OpenHalDet standardizes the evaluation pipeline, from prompt construction and response generation to truthfulness annotation, detector scoring, and metric computation. It supports heterogeneous detector families under different access settings, including black-box methods that use only generated outputs, gray-box methods that rely on probability-based signals, and white-box methods that exploit internal model signals. By bringing diverse tasks, models, and detectors into a shared framework, OpenHalDet enables controlled comparison and provides a systematic view of how different detection paradigms behave in LLM applications. We release OpenHalDet as an open and extensible codebase to facilitate reproducible evaluation and future development of hallucination detection methods. The code and datasets are available at https://github.com/Nellie179/Hallucination-Detection.

cs.CL↗

arXiv2026

TrafficRAG: A Multimodal RAG Framework for Traffic Accident Liability Determination

Traffic accident liability analysis is a critical yet challenging task in intelligent transportation and legal assistance. Existing methods often suffer from low efficiency, subjective judgment, and inconsistent analysis results. Meanwhile, large language models are constrained by noisy video inputs and insufficient legal domain knowledge. To address these issues, this work presents TrafficRAG, a multimodal retrieval-augmented framework for automated traffic accident analysis and report generation. Specifically, the proposed framework first adopts a vision-language model to produce structured textual descriptions of accident scenarios, which serve as accurate retrieval queries. Based on these textual queries, a hybrid retrieval strategy integrating BM25 sparse retrieval and dense embedding retrieval is employed to fetch relevant traffic regulations and similar historical cases. Finally, the large language model incorporates retrieved legal knowledge and multimodal accident evidence for comprehensive reasoning, and generates standardized, legally grounded liability analysis reports. Extensive experiments show that TrafficRAG consistently outperforms baseline methods, achieving 77.32% Legal Norm Adaptation Accuracy, 81.71% Factual Faithfulness, and a Liability Ratio MAE of 5.48%. The results validate that integrating multimodal factual evidence with legal clauses via retrieval augmentation can effectively improve the reliability and accuracy of traffic accident liability determination.

cs.AI↗

arXiv2026

EvoGens: A Population-Based Heuristic Search Framework for Scientific Idea Generation

Generating novel research ideas is fundamental to scientific progress. While Large Language Models (LLMs) show promise in assisting this process, existing approaches often exhibit semantic convergence, resulting in limited diversity and novelty. To address this, we introduce EvoGens, an evolution-inspired framework that recasts scientific idea generation as an evolutionary search over a population of ideas. EvoGens iteratively applies rank-based mutation with differentiated retrieval planning to incorporate external knowledge, and semantic-aware crossover to fuse complementary concepts for conceptual reorganization. A lightweight evaluation signal guides the selection process, encouraging sustained exploration while mitigating premature convergence. Extensive experiments demonstrate that EvoGens substantially enhances exploration capabilities compared to state-of-the-art baselines. Specifically, it improves the Novelty from 0.1 to 0.4 and the Diversity from 0.24 to 0.55, while maintaining comparable idea quality under the current automatic evaluation protocol. These findings suggest that evolutionary mechanisms can serve as a useful framework for exploration-oriented research ideation, especially for broadening the novelty and diversity of candidate ideas under a shared automatic evaluation setting.

cs.CL↗

arXiv2026

LLMs Need Encoders for Semantic IDs Too

Multimodal LLMs use dedicated encoders to bridge non-language modalities (vision encoders for images, depth models for audio codec tokens) because raw token embeddings alone cannot capture modality-specific structure. We argue that Semantic IDs (SIDs), the hierarchical codes used in generative recommendation, constitute another such modality: a SID level token's meaning depends on its prefix context, yet current systems simply add SID tokens to the vocabulary and rely on training to learn these context-dependent meanings from scratch. We propose PrefixMem, a lightweight SID encoder based on prefix n-gram memory tables that provides the LLM with structured, prefix-conditioned representations at SID token positions. Like vision encoders in multimodal LLMs, PrefixMem can be pre-trained independently and then attached to any LLM for joint training. We evaluate on large-scale data from Pinterest across multiple LLM families and show that PrefixMem improves deepest-level SID accuracy by up to 46% relative and full-SID retrieval recall by up to 22% relative at matched training compute. The encoder's benefit concentrates on hard examples where greedy decoding fails, with up to 77% relative accuracy gains, confirming that SID tokens benefit from a dedicated encoder just as other non-language modalities do.

cs.IR↗

arXiv2026

UniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale

Modern recommendation systems predominantly train retrieval and ranking as separate models despite both increasingly relying on large transformers encoding the same user behavior data, duplicating parameters, compute, and serving cost. Prior work unifies the model architecture but not the full pipeline: input formats, training procedures, and serving stacks remain fragmented across stages. We present UniPinRec, which achieves full-stack unification of retrieval and ranking at Pinterest: one input format, one model, one training stage, deployed within existing serving infrastructure. A shared transformer encodes the user action sequence into candidate-independent representations that branch into retrieval (ANN dot-product) and ranking (cross-attention) via task-specific heads. Three ideas make this work: (1) Masked Action Modeling (MAM) eliminates interleaving, enabling weight sharing without doubling context length; (2) Blended training examples pair action sequences with feedview impression slates to satisfy both objectives jointly; (3) Cross-stage KV cache sharing reuses user-history computation from retrieval for ranking, reducing total FLOPs versus serving two independent models. Deployed in the Pinterest core surfaces, UniPinRec delivers approximately +1% online engagement lift while cutting end-to-end serving latency by 11.1% and lifting QPS by 63.6%. To our knowledge, this is the first full-stack unification of retrieval and ranking, covering inputs, model, training and serving, deployed in a production recommendation system.

cs.IR↗

arXiv2026

Ultra-Confinement of Polaritons in Single Atomic Layer Ag Photonic Quantum Dots

Light scattering by two-dimensional (2D) van der Waals heterostructures (vdWHs) is immense, especially given their infinitesimal volume, thus enabling strong light-matter interactions. Surface 2D polariton waves manifest through large concentration of electromagnetic field in vertical direction, normal to their propagation. By confining vdWH materials into 2D photonic shapes, one can manipulate and compress light in lateral directions. Scattering-type scanning near-field optical microscopy is a perfect tool for direct imaging of the propagating polaritons and studying the properties of confined polaritons in nanostructures. Though, thus far the quantitative analysis, such the wavelength extraction, has been challenged for confined polaritons by incapability of mapping of the wave period on sub-wavelength scale and difficulty of identifying an adequate substrate's "background" to subtract. Here, an analytical approach is developed to reveal the local propagation constant of confined polaritons under abovementioned constraints and map it with the sub-wavelength resolution. Applied to analysis of the SiC/2D-Ag/EG (epitaxial graphene) photonic nanostructures, the technique uncovered that the polaritons are highly confined in both vertical ($\sim\lambda$/50) and lateral directions ($\sim\lambda$/40) by 2D metal.

cond-mat.mtrl-sci↗

arXiv2026

Tail exponents of the three-dimensional uniform spanning tree and Abelian sandpile

We study the local geometry of the three-dimensional uniform spanning tree and its connection with the Abelian sandpile model. We obtain sharp tail exponents, up to subpolynomial errors, for the past of the origin in the three-dimensional UST and for the $0$-tree of the $0$-wired uniform spanning forest. As a principal application, we prove the corresponding three-dimensional Abelian sandpile avalanche exponents: the avalanche-cluster radius has tail exponent $1$, while both the avalanche-cluster size and the total number of topplings have tail exponent $1/3$. These results identify the leading power-law behaviour of three-dimensional sandpile avalanches and improve previously known bounds.

math.PR↗

arXiv2026

Generalized intersection exponents and local cut points for three-dimensional Brownian loop soup

We study generalized non-intersection probabilities for the three-dimensional Brownian loop soup at subcritical intensities. We establish the existence of generalized intersection exponents (GIE) and prove an up-to-constants estimate for these probabilities by means of a separation lemma tailored to this setting. We also relate the Hausdorff dimension of the set of local cut points of the three-dimensional Brownian loop soup to the GIE, and show that the GIE is continuous at intensity zero, where it reduces to the classical Brownian intersection exponent. In particular, this implies that, for sufficiently small intensity parameters, the set of local cut points has Hausdorff dimension strictly larger than $1$.

math.PR↗

arXiv2026

Loop pruning and downward deviations for maximum local time of discrete-time simple random walks

We study downward deviations of the maximum local time of the discrete-time simple random walk on $\mathbb{Z}^d$, $d\ge 3$. In our previous paper \cite{li2026ldmaxlocal}, the corresponding upper bound was established, while the matching lower bound was left open. In the present paper, we prove this lower bound and hence obtain the sharp asymptotic formula for the downward-deviation probability. To provide a discrete-time analogue of the jump-chain/holding-time structure used in the continuous-time argument, we introduce a new random structure which we name as {\it loop-pruned random walk} and the associated loop-pruning decomposition, which is also of independent interest.

math.PR↗