arXiv subjects

Bowen Zeng

Publications and source records attributed to Bowen Zeng.

17 recordsLinked to original sources

arXiv2026

Multi-Faceted Evaluation and Mitigation of Emotion Hallucinations in MLLMs

Multimodal large language models (MLLMs) have shown strong potential in open-ended emotion understanding, yet they often generate emotion hallucinations. Evaluating such hallucinations is particularly challenging for two reasons. First, emotion understanding spans multiple cognitive facets, from multimodal perception to psychological reasoning. Second, emotional interpretations are expressed in free-form language, making existing closed-ended protocols insufficient for evaluation. To address these challenges, we introduce EHR (Emotion Hallucination Rate), an evaluator that quantifies emotion hallucinations across six facets: expression, action, audio, instinct, logic, and conclusion. Using EHR, we reveal that existing mitigation methods often reduce hallucinations in some facets while aggravating them in others, exposing the limitation of coarse-grained correction and the need for facet-aware localization and mitigation. Motivated by this finding, we propose HMER (Hallucination-aware Memory-guided Emotion Reasoning), a training-free framework for emotion hallucination mitigation. HMER maintains a Hallucination Memory that records localized hallucinated claims and enables targeted logit rectification, together with an Anchor Memory that preserves reliable intermediate reasoning states to stabilize subsequent generation. By selectively suppressing unreliable cues while preserving trustworthy reasoning context, HMER enables fine-grained mitigation across diverse hallucination facets. Extensive experiments on 19 MLLMs demonstrate the prevalence of emotion hallucinations and the effectiveness of our framework across diverse model architectures.

cs.MM↗

arXiv2026

HARD-KV: Head-Adaptive Regularization for Decoding-time KV Compression

Long-context LLM inference faces a fundamental conflict: head-adaptive compression algorithms (e.g., Top-$p$ nucleus sampling) offer superior accuracy by dynamically fluctuating memory budgets, yet modern inference engines (e.g., vLLM) demand rigid, static memory patterns to leverage CUDA Graphs and PagedAttention. We resolve this ``Static-Dynamic'' mismatch with HARD-KV, a unified framework that that bridges dynamic selection with rigid system constraints. HARD-KV introduces a Cascade Cache hierarchy, managing the token lifecycle across dense, sparse, and condensed tiers. Crucially, we propose a Logits Calibration mechanism that normalizes diverse importance metrics into a unified probability space, enabling consistent Top-$p$ budgeting across heterogeneous heads. To bridge the efficiency gap, we offer a system-level solution, which rewrites fragmented, dynamic indices into contiguous physical layouts compatible with high-performance inference engine. Extensive experiments on math-reasoning benchmarks (AIME, U-Math) verify that HARD-KV achieves up to 2$\times$ throughput improvement over static baselines while maintaining high-fidelity generation in 10k+ token scenarios. Code is available at https://github.com/SuDIS-ZJU/HARDInfer.

cs.LG↗

arXiv2026

Scale-Free Response with Directional Amplification in Critical Non-Hermitian Systems

The non-Hermitian skin effect can lead to directional amplification of response, with the associated end-to-end Green's function generally exhibiting size dependence. Any deviation in length or local disorder can drastically alter the amplification factor, rendering the response fragile in practical implementations. In this work, we identify a new type of scale-free, topological, and directionally amplified response in a Hatano-Nelson model under perturbed open boundary conditions. The scale-free response can be attributed to the first order boundary effect and characterized by a winding number defined on a continuous generalization of the finite-size Brillouin zone-a concept introduced in this work. Such scale-free behavior endows the end-to-end Green's function with significant robustness and making it promising for practical applications.

quant-ph↗

arXiv2026

Efficient Inference for Large Vision-Language Models: Bottlenecks, Techniques, and Prospects

Large Vision-Language Models (LVLMs) enable sophisticated reasoning over images and videos, yet their inference is hindered by a systemic efficiency barrier known as visual token dominance. This overhead is driven by a multi-regime interplay between high-resolution feature extraction, quadratic attention scaling, and memory bandwidth constraints. We present a systematic taxonomy of efficiency techniques structured around the inference lifecycle, consisting of encoding, prefilling, and decoding. Unlike prior reviews focused on isolated optimizations, we analyze the end-to-end pipeline to reveal how upstream decisions dictate downstream bottlenecks, covering compute-bound visual encoding, the intensive prefilling of massive contexts, and the ''visual memory wall'' in bandwidth-bound decoding. By decoupling the efficiency landscape into the axes of shaping information density, managing long-context attention, and overcoming memory limits, this work provides a structured analysis of how isolated optimizations compose to navigate the trade-off between visual fidelity and system efficiency. The survey concludes by outlining four future frontiers supported by pilot empirical insights, including hybrid compression based on functional unit sensitivity, modality-aware decoding with relaxed verification, progressive state management for streaming continuity, and stage-disaggregated serving through hardware-algorithm co-design. Our literature repository is at https://github.com/SuDIS-ZJU/Efficient-LVLMs-Inference.

cs.CL↗

arXiv2026

HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference

Multimodal Large Language Models (MLLMs) have advanced unified reasoning over text, images, and videos, but their inference is hindered by the rapid growth of key-value (KV) caches. Each visual input expands into thousands of tokens, causing caches to scale linearly with context length and remain resident in GPU memory throughout decoding, which leads to prohibitive memory overhead and latency even on high-end GPUs. A common solution is to compress caches under a fixed allocated budget at different granularities: token-level uniformly discards less important tokens, layer-level varies retention across layers, and head-level redistributes budgets across heads. Yet these approaches stop at allocation and overlook the heterogeneous behaviors of attention heads that require distinct compression strategies. We propose HybridKV, a hybrid KV cache compression framework that integrates complementary strategies in three stages: heads are first classified into static or dynamic types using text-centric attention; then a top-down budget allocation scheme hierarchically assigns KV budgets; finally, static heads are compressed by text-prior pruning and dynamic heads by chunk-wise retrieval. Experiments on 11 multimodal benchmarks with Qwen2.5-VL-7B show that HybridKV reduces KV cache memory by up to $7.9\times$ and achieves $1.52\times$ faster decoding, with almost no performance drop or even higher relative to the full-cache MLLM.

cs.AI↗

arXiv2025

Finite-size Effects on The Edge Loss Probability in Non-Hermitian Quantum Walks

A dynamical bulk-edge relation in quantum walks has been theoretically proposed and experimentally observed, in which a power-law dependence of the bulk loss probability is associated with a pronounced peak of loss probability at the edge. This behavior has been proven to arise from imaginary gap closing and the non-Hermitian skin effect in the infinite limit without boundary effects. However, in a finite-size chain, we find that boundary scattering can suppress this edge burst. Meanwhile, imaginary gap opening, together with the non-Hermitian skin effect, can also induce a large loss probability at the edge. Our results provide insights into finite-size quantum dynamics.

quant-ph↗

arXiv2025

EgoLCD: Egocentric Video Generation with Long Context Diffusion

Generating long, coherent egocentric videos is difficult, as hand-object interactions and procedural tasks require reliable long-term memory. Existing autoregressive models suffer from content drift, where object identity and scene semantics degrade over time. To address this challenge, we introduce EgoLCD, an end-to-end framework for egocentric long-context video generation that treats long video synthesis as a problem of efficient and stable memory management. EgoLCD combines a Long-Term Sparse KV Cache for stable global context with an attention-based short-term memory, extended by LoRA for local adaptation. A Memory Regulation Loss enforces consistent memory usage, and Structured Narrative Prompting provides explicit temporal guidance. Extensive experiments on the EgoVid-5M benchmark demonstrate that EgoLCD achieves state-of-the-art performance in both perceptual quality and temporal consistency, effectively mitigating generative forgetting and representing a significant step toward building scalable world models for embodied AI. Code: https://github.com/AIGeeksGroup/EgoLCD. Website: https://aigeeksgroup.github.io/EgoLCD.

cs.CV↗

arXiv2025

Long-Context Attention Benchmark: From Kernel Efficiency to Distributed Context Parallelism

Transformer-based large language models (LLMs) have achieved remarkable success, yet their standard attention mechanism incurs quadratic computation and memory costs with respect to sequence length, posing a major bottleneck for long-context training. Prior work tackles this challenge along two directions: (1) kernel-level optimizations, which accelerate dense and sparse attention operators; and (2) module-level strategies, often referred to as distributed attention or context parallel training, which scale attention across multiple devices. However, systematic evaluation still remains limited: operator-level comparisons are often incomplete, while context parallel strategies are typically framework-specific, with unclear performance analysis across contexts. To address these gaps, we propose a unified benchmark that integrates representative attention kernels and context parallel mechanisms with a modular and extensible interface for evaluation. The benchmark evaluates methods along two critical dimensions: (1) attention mask patterns, which strongly affect efficiency, scalability, and usability, and (2) sequence length and distributed scale, which determine performance under extreme long-context training. Through comprehensive experiments on the cluster of up to 96 GPUs, our benchmark enables reproducible comparisons, highlights method-specific trade-offs, and provides practical guidance for designing and deploying attention mechanisms in long-context LLM training.

cs.LG↗

arXiv2025

Critical Non-Hermitian Edge Modes

We unveil a unique critical phenomenon of topological edge modes in non-Hermitian systems, dubbed the critical non-Hermitian edge modes (CNHEM). Specifically, in the thermodynamic limit, the eigenvectors of edge modes jump discontinuously under infinitesimal on-site staggered perturbations. The CNHEM arises from the competition between the introduced on-site staggered potentials and size-dependent non-reciprocal coupling between edge modes, and are closely connected to the exceptional point (EP). As the system size increases, the coupling between edge modes decreases while the non-reciprocity is enhanced, causing the eigenvectors to gradually collapse toward the EP. However, when the on-site potentials dominate, this weakened coupling assists the eigenvectors to stay away from the EP. Such a critical phenomenon is absent in Hermitian systems, where the coupling between edge modes is reciprocal.

cond-mat.mes-hall↗

arXiv2025

Non-Hermitian Aharonov-Bohm Cage in Bosonic Bogoliubov-de Gennes Systems

The non-Hermitian Aharonov-Bohm (AB) cage is a unique localization phenomenon that confines all possible excitations. This confinement leads to fully flat spectra in momentum space, which are typically accompanied with the degeneracy with various types. Classifying the degeneracy type is crucial for studying the dynamical properties of the non-Hermitian AB cage, but the methods for such classification and their physical connections remain not very clear. Here, we construct a non-Hermitian AB cage in a bosonic Bogoliubov-de Gennes (BdG) system with various types of degenerate flat bands (DFBs). Using the transfer matrix, we demonstrate the localization mechanism for the formation of AB cage and derive the minimal polynomial in mathematics for classifying the degeneracy types of DFBs, thus providing comprehensive understanding of the correspondence among the degeneracy type of DFBs, the minimal polynomial, and the transfer matrix. With such correspondence, we propose a scheme to realize highly degenerate flat bands.

quant-ph↗

arXiv2024

Inverse Design of Winding Tuple for Non-Hermitian Topological Edge Modes

The interplay between topological localization and non-Hermiticity localization in non-Hermitian crystal systems results in a diversity of shapes of topological edge modes (EMs), offering opportunities to manipulate these modes for potential topological applications. The characterization of the domain of EMs and the engineering of these EMs require detailed information about their wave functions, which conventional calculation of topological invariants cannot provide. In this Letter, by recognizing EMs as specified solutions of eigenequation, we derive their wave functions in an extended non-Hermitian Su-Schrieffer-Heeger model. We then inversely construct a winding tuple $\left \{ w_{\scriptscriptstyle GBZ},w_{\scriptscriptstyle BZ}\right \} $ that characterizes the existence of EMs and their spatial distribution. Moreover, we define a novel spectral winding number equivalent to $w_{\scriptscriptstyle BZ}$, which is determined by the product of energies of different bands. The inverse design of topological invariants allows us to categorize the localized nature of EMs even in systems lacking sublattice symmetry, which can facilitate the manipulation and utilization of EMs in the development of novel quantum materials and devices.

cond-mat.mes-hall↗

arXiv2024

Abnormal Frequency Response Determined by Saddle Points in Non-Hermitian Crystal Systems

In non-Hermitian crystal systems under open boundary condition (OBC), it is generally believed that the OBC modes with frequencies containing positive imaginary parts, when excited by external driving, will experience exponential growth in population, thereby leading to instability. However, our work challenges this conventional understanding. In such a system, we find an anomalous response that grows exponentially with the frequency aligned with those of saddle points. The frequencies of these saddle points on the complex plane are below the maximum imaginary part of OBC spectrum, but they can lie within or beyond the OBC spectrum. We derive general formulas of excitation-response relationships and find that this anomalous response can occur because the excitation of OBC modes eventually evolve toward these saddle points at long times. Only when the frequencies of all these saddle points are below the real axis do the non-Hermitian crystal systems remain stable under periodic excitation. Thus our results also provide new insights on the stability criterion of non-Hermitian crystal systems.

cond-mat.other↗

arXiv2023

Non-Hermitian Topological Magnonics

Dissipation in mechanics, optics, acoustics, and electronic circuits is nowadays recognized to be not always detrimental but can be exploited to achieve non-Hermitian topological phases or properties with functionalities for potential device applications. As elementary excitations of ordered magnetic moments that exist in various magnetic materials, magnons are the information carriers in magnonic devices with low-energy consumption for reprogrammable logic, non-reciprocal communication, and non-volatile memory functionalities. Non-Hermitian topological magnonics deals with the engineering of dissipation and/or gain for non-Hermitian topological phases or properties in magnets that are not achievable in the conventional Hermitian scenario, with associated functionalities cross-fertilized with their electronic, acoustic, optic, and mechanic counterparts, such as giant enhancement of magnonic frequency combs, magnon amplification, (quantum) sensing of the magnetic field with unprecedented sensitivity, magnon accumulation, and perfect absorption of microwaves. In this review article, we address the unified approach in constructing magnonic non-Hermitian Hamiltonian, introduce the basic non-Hermitian topological physics, and provide a comprehensive overview of the recent theoretical and experimental progress towards achieving distinct non-Hermitian topological phases or properties in magnonic devices, including exceptional points, exceptional nodal phases, non-Hermitian magnonic SSH model, and non-Hermitian skin effect. We emphasize the non-Hermitian Hamiltonian approach based on the Lindbladian or self-energy of the magnonic subsystem but address the physics beyond it as well, such as the crucial quantum jump effect in the quantum regime and non-Markovian dynamics. We provide a perspective for future opportunities and challenges before concluding this article.

cond-mat.mes-hall↗

arXiv2023

Efficient Spin Seebeck and Spin Nernst Effects of Magnons in Altermagnets

We report two non-degenerate magnon modes with opposite spins or chiralities in collinearly antiferromagnetic insulators driven by symmetry-governed anisotropic exchange couplings. The consequent giant spin splitting contributes to spin Seebeck and spin Nernst effects generating longitudinal and transverse spin currents when the temperature gradient applies along and away from the main crystal axis, without requiring any external magnetic field and spin-orbit coupling. Based on first-principle calculations, we predict feasible material candidates holding robust altermagnetic spin configurations and room-temperature structural stability to efficiently transport spin. The spin Seebeck conductivity is comparable to the records of antiferromagnets that require the magnetic field, and the spin Nernst conductivity is two orders in magnitude larger than that in antiferromagnetic monolayers that need Berry curvature.

cond-mat.mes-hall↗

arXiv2022

Radiation-free and non-Hermitian topology inertial defect states of on-chip magnons

Radiative damping is a strong dissipation source for the quantum emitters hybridized with propagating photons, electrons, or phonons, which is not easily avoidable for on-chip magnonic emitters as well that can radiate via the surface acoustic waves of the substrate. Here we demonstrate in an array of on-chip nano-magnets coupled in a long range via exchanging the surface acoustic waves that a point defect in the array, which can be introduced by the local magnon frequency shift by a local biased magnetic field or the absence of a magnetic wire, strongly localizes the magnons, in contrast to the well spreading Bloch-like collective magnon modes in such an array setting. The radiation of the magnon defect states is exponentially suppressed by the distance of the defect to the array edges. Moreover, this defect state is strikingly inertial to the non-Hermitian topology that localizes all the extended states at one boundary. Such configuration robust magnon defect states towards radiation-free limit may be suitable for high-fidelity magnon quantum information storage in the future on-chip magnonic devices.

cond-mat.mes-hall↗

arXiv2022

Giant Microwave Sensitivity of Magnetic Array by Long-Range Chiral Interaction Driven Skin Effect

Non-Hermitian skin effect was observed in one-dimensional systems with short-range chiral interaction. Long-range chiral interaction mediated by traveling waves also favors the accumulation of energy, but has not yet showed non-Hermitian topology. Here we find that the strong interference brought by the wave propagation is detrimental for accumulation. By suppression of interference via the damping of traveling waves, we predict the non-Hermitian skin effect of magnetic excitation in a periodic array of magnetic nanowires that are coupled chirally via spin waves of thin magnetic films. The local excitation of a wire at one edge by weak microwaves of magnitude $\sim μ{\rm T}$ leads to a considerable spin-wave amplitude at the other edge, i.e. a remarkable functionality useful for sensitive, non-local, and non-reciprocal detection of microwaves.

cond-mat.mes-hall↗

arXiv2019

Stress-sign-tunable Poisson's Ratio in Monolayer Blue Phosphorus Oxide

Negative Poisson's ratio (NPR) materials have attracted tremendous interest due to their unusual physical properties and potential applications. Certain two-dimensional (2D) monolayer materials have also been found to exhibit NPR and the corresponding deformation mechanism varies. In this study, we found, based on first-principles calculations, that the Poisson's ratio (PR) sign of monolayer Blue Phosphorus Oxide (BPO) can be tuned by strain: the PR is positive under uniaxial strain <= -1% but becomes negative under > 0. The deformation mechanism for BPO under strain depends on the mutual competition between the P-P attraction and P-O repulsion effect, and these two factors induce two different deformation pathways (one with positive PR, and the other with NPR). Moreover, with increasing of strain, both the decreased strength of P-P attraction and the increased strength of P-O repulsion effect modulate the PR of BPO from positive to negative.

cond-mat.mtrl-sci↗