arXiv subjects

Jiayi Lu

Publications and source records attributed to Jiayi Lu.

15 recordsLinked to original sources

arXiv2026

HAFI-VLM: A Frequency Perspective for Diagnosing and Enhancing Visual Perception in Vision-Language Models

Vision-language models (VLMs) remain unreliable when predictions require fine-grained visual evidence. We identify a previously overlooked cause: spectral response rigidity. Despite substantial frequency variation across images and tasks, pretrained vision encoders exhibit persistent, encoder-specific layerwise spectral profiles that change only marginally under downstream fine-tuning. Since pretrained vision encoders only receive images, they cannot adapt spectral extraction to the evidence required by the current query. We therefore propose HAFI-VLM, which introduces a task-conditioned frequency pathway while preserving the pretrained semantic representation. Hierarchical Adaptive Frequency Injection (HAFI) retrieves complementary low-, mid-, and high-frequency evidence at multiple encoder depths using text-modulated, spatially aligned cross-attention. A Visual Enrichment Layer Adapter further recalibrates shallow LLM attention to effectively utilize the enriched visual tokens. Experiments on LLaVA-1.5 and Qwen2.5-VL demonstrate consistent improvements in general VQA, text-rich understanding, and hallucination robustness, outperforming representation-level enhancement methods and most resolution- or cropping-based approaches without additional high-resolution encoding. Mechanistic analyses show that HAFI restores task-dependent spectral allocation while retaining semantic attention, establishing frequency enrichment as a distinct and effective route for improving VLM perception.

cs.CV↗

arXiv2026

"The Whole Is Greater Than the Sum of Its Parts": A Compatibility-Aware Multi-Teacher CoT Distillation Framework

Chain-of-Thought (CoT) reasoning empowers Large Language Models (LLMs) with remarkable capabilities but typically requires prohibitive parameter scales. CoT distillation has emerged as a promising paradigm to transfer reasoning prowess into compact Student Models (SLMs), but existing approaches often rely on a solitary teacher, capping the student's potential since individual LLMs often exhibit distinct capability biases and may suffer from catastrophic forgetting. While leveraging diverse teachers seems appealing, effectively fusing their supervisions remains challenging: teacher-student incompatibility risks amplifying hallucinations, and passive supervision fails to ensure genuine logic internalization. To address this, we introduce COMPACT, a framework that adaptively fuses supervisions from different teachers by dynamically weighting teacher gradients based on the student's real-time compatibility evaluated by a multi-dimensional metric: (1) Graph-based Consensus to filter misleading rationales by identifying mainstream reasoning paths; (2) Mutual-Information-based Adaptability to detect "epiphany moments" for genuinely understanding the reasoning process rather than merely imitating; and (3) Loss-based Difficulty to assess student receptivity to the teacher's guidance and prevent negative transfer. Extensive experiments and latent space analysis demonstrate that COMPACT effectively integrates diverse reasoning capabilities without damaging the model's original knowledge structure, achieving state-of-the-art performance on various benchmarks while mitigating catastrophic forgetting.

cs.CL↗

arXiv2026

ECHO: Continuous Hierarchical Memory for Vision-Language-Action Models

Memory capacity is a critical factor determining the performance of Vision-Language-Action (VLA) models in long-horizon manipulation tasks. Existing memory-augmented architectures primarily rely on linear or flat storage, lacking structural priors for manipulation categories and hierarchical organization. This deficiency hinders efficient experience retrieval and limits generalization to unseen long-horizon task compositions. Inspired by the hierarchical organization of human experience, we propose ECHO (Experience Consolidation and Hierarchical Organization), a novel memory framework operating within a Continuous Hierarchical Space. By employing a hyperbolic autoencoder, ECHO maps VLA hidden states into this space. Leveraging hyperbolic metrics and entailment constraint mechanisms, experience vectors are organized into a semantic memory tree that supports efficient top-down retrieval. In parallel, a background consolidation mechanism continuously refines the memory tree through geometric interpolation and structural splitting, supporting virtual memory synthesis in the continuous space. We integrate ECHO into the $π_0$ foundation model. Evaluations on LIBERO and preliminary real-world experiments demonstrate the effectiveness of our approach, notably achieving a 12.8% absolute improvement in execution success rate over the $π_0$ baseline on LIBERO-Long, while improving compositional generalization on cross-suite unseen long-horizon tasks.

cs.RO↗

arXiv2026

Towards Real-World Industrial-Scale Verification: LLM-Driven Theorem Proving on seL4

Formal methods (FM) are reliable but costly to apply, often requiring years of expert effort in industrial-scale projects such as seL4, especially for theorem proving. Recent advances in large language models (LLMs) have made automated theorem proving increasingly feasible. However, most prior work focuses on mathematics-oriented benchmarks such as miniF2F, with limited evaluation on real-world verification projects. The few studies that consider industrial-scale verification mostly rely on closed-source models with hundreds of billions of parameters, which cannot be locally deployed and incur substantial usage costs. In this paper, we propose AutoReal, an LLM-driven theorem proving method for real-world industrial-scale systems with support for lightweight local deployment. We evaluate AutoReal on the seL4-Isabelle verification project as a representative and challenging case study. AutoReal incorporates two key improvements: (1) chain-of-thought (CoT)-based proof training, which teaches the LLM the reasoning behind proof steps and enables step-wise explanations alongside proofs, and (2) context augmentation, which leverages proof context from the project to enhance LLM-driven proving. Based on the AutoReal methodology, we fine-tune a base model to obtain AutoReal-Prover, a compact 7B-scale prover for industrial-scale theorem proving. AutoReal-Prover achieves a 51.67% proof success rate on 660 theorems from seL4-designated Important Theories across all 10 seL4 proof categories, substantially outperforming prior attempts on seL4 (27.06%). To evaluate generalization, we further apply AutoReal-Prover to three security-related projects from the Archive of Formal Proofs (AFP), covering all 451 theorems and achieving a proof success rate of 53.88%. Overall, this work advances the application of LLM-driven theorem proving in real-world industrial-scale verification.

cs.CR↗

arXiv2026

MIND: From Passive Mimicry to Active Reasoning through Capability-Aware Multi-Perspective CoT Distillation

While Large Language Models (LLMs) have emerged with remarkable capabilities in complex tasks through Chain-of-Thought reasoning, practical resource constraints have sparked interest in transferring these abilities to smaller models. However, achieving both domain performance and cross-domain generalization remains challenging. Existing approaches typically restrict students to following a single golden rationale and treat different reasoning paths independently. Due to distinct inductive biases and intrinsic preferences, alongside the student's evolving capacity and reasoning preferences during training, a teacher's "optimal" rationale could act as out-of-distribution noise. This misalignment leads to a degeneration of the student's latent reasoning distribution, causing suboptimal performance. To bridge this gap, we propose MIND, a capability-adaptive framework that transitions distillation from passive mimicry to active cognitive construction. We synthesize diverse teacher perspectives through a novel "Teaching Assistant" network. By employing a Feedback-Driven Inertia Calibration mechanism, this network utilizes inertia-filtered training loss to align supervision with the student's current adaptability, effectively enhancing performance while mitigating catastrophic forgetting. Extensive experiments demonstrate that MIND achieves state-of-the-art performance on both in-distribution and out-of-distribution benchmarks, and our sophisticated latent space analysis further confirms the mechanism of reasoning ability internalization.

cs.CL↗

arXiv2025

MindDrive: An All-in-One Framework Bridging World Models and Vision-Language Model for End-to-End Autonomous Driving

End-to-End autonomous driving (E2E-AD) has emerged as a new paradigm, where trajectory planning plays a crucial role. Existing studies mainly follow two directions: trajectory generation oriented, which focuses on producing high-quality trajectories with simple decision mechanisms, and trajectory selection oriented, which performs multi-dimensional evaluation to select the best trajectory yet lacks sufficient generative capability. In this work, we propose MindDrive, a harmonized framework that integrates high-quality trajectory generation with comprehensive decision reasoning. It establishes a structured reasoning paradigm of "context simulation - candidate generation - multi-objective trade-off". In particular, the proposed Future-aware Trajectory Generator (FaTG), based on a World Action Model (WaM), performs ego-conditioned "what-if" simulations to predict potential future scenes and generate foresighted trajectory candidates. Building upon this, the VLM-oriented Evaluator (VLoE) leverages the reasoning capability of a large vision-language model to conduct multi-objective evaluations across safety, comfort, and efficiency dimensions, leading to reasoned and human-aligned decision making. Extensive experiments on the NAVSIM-v1 and NAVSIM-v2 benchmarks demonstrate that MindDrive achieves state-of-the-art performance across multi-dimensional driving metrics, significantly enhancing safety, compliance, and generalization. This work provides a promising path toward interpretable and cognitively guided autonomous driving.

cs.CV↗

arXiv2025

Near-field Spatial-domain Channel Extrapolation for XL-MIMO Systems

Extremely large-scale multiple-input multiple-output (XL-MIMO) systems are pivotal to next-generation wireless communications, where dynamic RF chain architectures offer enhanced performance. However, efficient precoding in such systems requires accurate channel state information (CSI) obtained with low complexity. To address this challenge, spatial-domain channel extrapolation has attracted growing interest. Existing methods often overlook near-field spherical wavefronts or rely heavily on sparsity priors, leading to performance degradation. In this paper, we propose an adaptive near-field channel extrapolation framework for multi-subcarrier XL-MIMO systems, leveraging a strategically selected subset of antennas. Subsequently, we develop both on-grid and off-grid algorithms, where the latter refines the former's estimates for improved accuracy. To further reduce complexity, a cross-validation (CV)-based scheme is introduced. Additionally, we analytically formulate the mutual coherence of the sensing matrix and propose a coherence-minimizing-based random pattern to ensure robust extrapolation. Numerical results validate that the proposed algorithms significantly outperform existing methods in both extrapolation accuracy and achievable rate, while maintaining low computational complexity. In particular, our proposed CV ratio offers a flexible trade-off between accuracy and efficiency, and the corresponding off-grid algorithm achieves high accuracy with complexity comparable to conventional on-grid methods.

eess.SP↗

arXiv2025

AetherCode: Evaluating LLMs' Ability to Win In Premier Programming Competitions

Competitive programming has emerged as a critical benchmark for evaluating the reasoning and coding capabilities of Large Language Models (LLMs). Despite impressive progress on existing benchmarks, we argue that current evaluations overstate model proficiency, masking a substantial gap between LLMs and elite human programmers. This gap arises from two key limitations: insufficient difficulty and scope of benchmark problems, and evaluation bias from low-quality test cases. To address these shortcomings, we present AetherCode, a new benchmark that draws problems from premier programming competitions such as IOI and ICPC, offering broader coverage and higher difficulty. AetherCode further incorporates comprehensive, expert-validated test suites built through a hybrid of automated generation and human curation, ensuring rigorous and reliable assessment. By combining challenging problem design with robust evaluation, AetherCode provides a more faithful measure of LLM capabilities and sets a new standard for future research in code reasoning.

cs.SE↗

arXiv2025

Multi-User Generative Semantic Communication with Intent-Aware Semantic-Splitting Multiple Access

With the booming development of generative artificial intelligence (GAI), semantic communication (SemCom) has emerged as a new paradigm for reliable and efficient communication. This paper considers a multi-user downlink SemCom system, using vehicular networks as the representative scenario for multi-user content dissemination. To address diverse yet overlapping user demands, we propose a multi-user Generative SemCom-enhanced intent-aware semantic-splitting multiple access (SS-MGSC) framework. In the framework, we construct an intent-aware shared knowledge base (SKB) that incorporates prior knowledge of semantic information (SI) and user-specific preferences. Then, we designate the common SI as a one-hot semantic map that is broadcast to all users, while the private SI is delivered as personalized text for each user. On the receiver side, a diffusion model enhanced with ControlNet is adopted to generate high-quality personalized images. To capture both semantic relevance and perceptual similarity, we design a novel semantic efficiency score (SES) metric as the optimization objective. Building on this, we formulate a joint optimization problem for multi-user semantic extraction and beamforming, solved using a reinforcement learning-based algorithm due to its robustness in high-dimensional settings. Simulation results demonstrate the effectiveness of the proposed scheme.

cs.NI↗

arXiv2025

On Path to Multimodal Historical Reasoning: HistBench and HistAgent

Recent advances in large language models (LLMs) have led to remarkable progress across domains, yet their capabilities in the humanities, particularly history, remain underexplored. Historical reasoning poses unique challenges for AI, involving multimodal source interpretation, temporal inference, and cross-linguistic analysis. While general-purpose agents perform well on many existing benchmarks, they lack the domain-specific expertise required to engage with historical materials and questions. To address this gap, we introduce HistBench, a new benchmark of 414 high-quality questions designed to evaluate AI's capacity for historical reasoning and authored by more than 40 expert contributors. The tasks span a wide range of historical problems-from factual retrieval based on primary sources to interpretive analysis of manuscripts and images, to interdisciplinary challenges involving archaeology, linguistics, or cultural history. Furthermore, the benchmark dataset spans 29 ancient and modern languages and covers a wide range of historical periods and world regions. Finding the poor performance of LLMs and other agents on HistBench, we further present HistAgent, a history-specific agent equipped with carefully designed tools for OCR, translation, archival search, and image understanding in History. On HistBench, HistAgent based on GPT-4o achieves an accuracy of 27.54% pass@1 and 36.47% pass@2, significantly outperforming LLMs with online search and generalist agents, including GPT-4o (18.60%), DeepSeek-R1(14.49%) and Open Deep Research-smolagents(20.29% pass@1 and 25.12% pass@2). These results highlight the limitations of existing LLMs and generalist agents and demonstrate the advantages of HistAgent for historical reasoning.

cs.AI↗

arXiv2025

FocalAD: Local Motion Planning for End-to-End Autonomous Driving

In end-to-end autonomous driving,the motion prediction plays a pivotal role in ego-vehicle planning. However, existing methods often rely on globally aggregated motion features, ignoring the fact that planning decisions are primarily influenced by a small number of locally interacting agents. Failing to attend to these critical local interactions can obscure potential risks and undermine planning reliability. In this work, we propose FocalAD, a novel end-to-end autonomous driving framework that focuses on critical local neighbors and refines planning by enhancing local motion representations. Specifically, FocalAD comprises two core modules: the Ego-Local-Agents Interactor (ELAI) and the Focal-Local-Agents Loss (FLA Loss). ELAI conducts a graph-based ego-centric interaction representation that captures motion dynamics with local neighbors to enhance both ego planning and agent motion queries. FLA Loss increases the weights of decision-critical neighboring agents, guiding the model to prioritize those more relevant to planning. Extensive experiments show that FocalAD outperforms existing state-of-the-art methods on the open-loop nuScenes datasets and closed-loop Bench2Drive benchmark. Notably, on the robustness-focused Adv-nuScenes dataset, FocalAD achieves even greater improvements, reducing the average colilision rate by 41.9% compared to DiffusionDrive and by 15.6% compared to SparseDrive.

cs.AI↗

arXiv2025

$3d$ flat bands and coupled $4f$ moments in the kagome-honeycomb permanent magnet Sm$_{2}$Co$_{17}$

Rare earth permanent magnets (REPMs) with both localized moments and itinerant conduction bands are not only important for fundamental research but also have significant technological applications. In particular, Sm$_{\rm 2}$Co$_{\rm 17}$ is a prototypical high-temperture REPM, where the Co atoms form a kagome-honeycomb stacked lattice. Here we report synthesis of epitaxial Sm$_{\rm 2}$Co$_{\rm 17}$ films using molecular beam epitaxy and measurements of their momentum-resolved electronic structure from \textit{in-situ} angle-resolved photoemission spectroscopy. Our results unveil two flat bands from Co $3d$ orbitals near the Fermi level ($E_F$), one at $\sim$\,--300\,meV and another right at $E_F$, which arise from orbital-selective destructive interference and strong electron correlations, respectively. In addition, our results reveal that Sm $4f$ states are far away from $E_F$ (hence mostly localized) and exhibit an anomalous temperature dependence, caused by the $3d$-$4f$ magnetic coupling. Our findings provide direct spectroscopic insights to understand the strong uniaxial ferromagnetism in Sm$_{\rm 2}$Co$_{\rm 17}$ (and REPMs alike). Our work also opens avenues to explore flat-band physics near $E_F$ and emergent phenomena in correlated kagome-honeycomb lattices.

cond-mat.str-el↗

arXiv2025

A Multi-modal Fusion Network for Terrain Perception Based on Illumination Aware

Road terrains play a crucial role in ensuring the driving safety of autonomous vehicles (AVs). However, existing sensors of AVs, including cameras and Lidars, are susceptible to variations in lighting and weather conditions, making it challenging to achieve real-time perception of road conditions. In this paper, we propose an illumination-aware multi-modal fusion network (IMF), which leverages both exteroceptive and proprioceptive perception and optimizes the fusion process based on illumination features. We introduce an illumination-perception sub-network to accurately estimate illumination features. Moreover, we design a multi-modal fusion network which is able to dynamically adjust weights of different modalities according to illumination features. We enhance the optimization process by pre-training of the illumination-perception sub-network and incorporating illumination loss as one of the training constraints. Extensive experiments demonstrate that the IMF shows a superior performance compared to state-of-the-art methods. The comparison results with single modality perception methods highlight the comprehensive advantages of multi-modal fusion in accurately perceiving road terrains under varying lighting conditions. Our dataset is available at: https://github.com/lindawang2016/IMF.

cs.AI↗

arXiv2024

Generative AI-Enhanced Multi-Modal Semantic Communication in Internet of Vehicles: System Design and Methodologies

Vehicle-to-everything (V2X) communication supports numerous tasks, from driving safety to entertainment services. To achieve a holistic view, vehicles are typically equipped with multiple sensors to compensate for undetectable blind spots. However, processing large volumes of multi-modal data increases transmission load, while the dynamic nature of vehicular networks adds to transmission instability. To address these challenges, we propose a novel framework, Generative Artificial intelligence (GAI)-enhanced multi-modal semantic communication (SemCom), referred to as G-MSC, designed to handle various vehicular network tasks by employing suitable analog or digital transmission. GAI presents a promising opportunity to transform the SemCom framework by significantly enhancing semantic encoding to facilitate the optimized integration of multi-modal information, enhancing channel robustness, and fortifying semantic decoding against noise interference. To validate the effectiveness of the G-MSC framework, we conduct a case study showcasing its performance in vehicular communication networks for predictive tasks. The experimental results show that the design achieves reliable and efficient communication in V2X networks. In the end, we present future research directions on G-MSC.

cs.NI↗

arXiv2023

Triangle Counting Rule: An Approach to Forecast the Magnetic Properties of Benzenoid Polycyclic Hydrocarbons

Open-shell benzenoid polycyclic hydrocarbons (BPHs) are promising materials for future quantum applications. However, the search and realization of open-shell BPHs with desired properties is a challenging task due to the gigantic chemical space of BPHs, requiring new strategies for both theoretical understanding and experimental advancement. In this work, by building a structure database of BPHs through graphical enumeration, performing data-driven analysis, and combining tight-binding and mean-field Hubbard calculations, we discovered that the number of the internal vertices of the BPH graphs is closely correlated to their open-shell characters. We further established a series of simple rules, the triangle counting rule (TCR), to predict the magnetic ground state of BPHs. These findings not only provide a database of open-shell BPHs, but also extend the well-known Lieb's theorem and Ovchinnikov's rule and provide a straightforward method for designing open-shell carbon nanostructures. These insights may aid in the exploration of emerging quantum phases and the development of magnetic carbon materials for technology applications.

physics.chem-ph↗