arXiv subjects

Xian Gao

Publications and source records attributed to Xian Gao.

At least 19 recordsLinked to original sources

arXiv2026

MEMO: Multimodal Evidence Memory Organization for Long-Horizon LLM Agents

Long-running LLM agents rely on external memory to store and reuse information beyond a single context window, yet there is a fundamental tension between the continuous accumulation of interaction trajectories and the limited context capacity. The key challenge in agent memory is therefore not only to retrieve relevant records, but also to select necessary evidence under a given budget and organize it in an appropriate modality. Existing memory readout methods mainly use textual or visual forms. Text preserves high fidelity, but its linear token representation makes contents with different importance compete for the limited context at nearly uniform unit cost. Visual readout renders text into document-like images, which can use two-dimensional layouts to expose structure and emphasize key information, but it may lose fine-grained details during rendering and compression. To address this issue, we propose MEMO, a multimodal evidence memory organization method for LLM agents. MEMO first uses a trained evidence extractor to select relevant memory blocks and form evidence units with source information and presentation requirements. A trained query-conditioned memory manager assigns each unit to a textual, visual, or dual-channel carrier and selects a layout that matches the evidence structure. A deterministic memory construction module then generates the textual package and visual pages. The memory manager is trained with feedback from an offline reader that measures the utility of the guided memory plan, so that retention and presentation decisions align with downstream usage. We evaluate MEMO on four benchmarks, HotpotQA, 2WikiMultiHopQA, LoCoMo, and ALFWorld, with multiple reader backends. The results show that MEMO presents memory more efficiently with fewer memory tokens, improves downstream task performance, and builds more effective working memory under constrained budgets.

cs.CL↗

arXiv2026

Edit2TikZ: A Comprehensive and Challenging Benchmark for Scientific Figure Editing with TikZ

Although multimodal large language models (MLLMs) have shown substantial potential in visual understanding and graphic code generation, editing scientific figures through code presents a greater challenge: a model must jointly recover visual structure, ground the requested change, generate compilable code, and preserve all unrelated content. While existing TikZ benchmarks mainly focus on figure reconstruction and generation, few systematically evaluate instruction-guided scientific figure editing with compilable code. We introduce Edit2TikZ, a comprehensive benchmark for scientific figure editing tasks, featuring 1,548 diverse and high-quality samples. Edit2TikZ combines real-world and controlled synthetic edit cases, supports both textual and visual localization request, and contains multi-step editing, each with step-level annotations. We further construct a human-aligned evaluation framework to measure whether a requested edit is completed while irrelevant content is preserved. Utilizing Edit2TikZ, we evaluate 14 mainstream MLLMs and find that current systems remain unreliable: on average, proprietary models achieve a compilation success rate of merely 75% and remain limited in both figure restoration and edit correctness, while compact models below 9B struggle further with instruction following and complete figure generation. Therefore, we build a mixed training set TikZEditMix and adopt reconstruction-then-editing curriculum learning for compact models. On Qwen3.5-4B, this training improves the compilation success rate from 45.35% to 83.40% and yields an average improvement of 18.7 points across our proposed evaluation metrics. The code and data will be released at https://github.com/Solunny/Edit2TikZ.

cs.CV↗

arXiv2026

Parity-violating spatially covariant gravity at total derivative order $d=5$

We extend the polynomial construction of parity-violating spatially covariant gravity (SCG) to total derivative order $d=5$, where $d=d_{\mathrm{t}}+d_{\mathrm{s}}$ counts the total number of temporal and spatial derivatives. After organizing the monomials by $(d_{\mathrm{t}},d_{\mathrm{s}})$ and reducing them using integrations by parts, tensor symmetries, three-dimensional curvature identities, the Schouten identity, and the Cayley-Hamilton relation, we obtain a $59$-element basis: $2$, $40$, and $17$ monomials in the sectors $(0,5)$, $(2,3)$, and $(4,1)$, respectively. A direct inspection separates $35$ monomials containing neither the lapse velocity $\mathcal{L}_{\bm{u}}\ln N$ nor $\mathcal{L}_{\bm{u}}K_{ij}$ from $11$ lapse-velocity monomials and $13$ monomials containing higher normal derivatives of the spatial metric. The latter two sectors require a dedicated degeneracy analysis. For tensor perturbations about a spatially flat cosmological background, the acceleration-free part of the manifestly first-order-in-time sector contains $16$ basis elements, whose quadratic action depends on only four combinations of coefficients. These combinations generate helicity-odd corrections proportional to $k/a$ and $(k/a)^3$ in the kinetic and gradient functions. We derive two relations that enforce luminal phase velocity for both circular polarizations while still allowing helicity-dependent kinetic normalization and damping.

gr-qc↗

arXiv2026

On the Fragility of Data Attribution When Learning Is Distributed

Data attribution has become an important component of pricing, auditing, and governance in machine learning pipelines, yet most attribution methods implicitly assume that attribution values faithfully reflect participants' contributions. We show that this assumption can fail: a single participant in a standard distributed training workflow can substantially inflate its measured attribution value while preserving global utility. Our attribution-first attack uses latent optimization to inject small synthetic batches that preserve utility while exploiting non-IID label coverage and evaluator sensitivities. Across datasets, models, and multiple marginal-utility evaluators, the attack consistently increases the adversary's attribution value and reshapes the relative attribution structure among benign clients without degrading accuracy or triggering geometry-based defenses. These results show that attribution itself forms a new attack surface and motivate the development of attribution-robust and incentive-compatible scoring mechanisms.

cs.LG↗

arXiv2026

Democratizing Music Therapy: LLM-Based Automated EEG Analysis and Progress Tracking for Low-Cost Home Devices

Home-based music therapy devices require accessible and cost-effective solutions for users to understand and track their therapeutic progress. Traditional physiological signal analysis, particularly EEG interpretation, relies heavily on domain experts, creating barriers to scalability and home adoption. Meanwhile, few experts are capable of interpreting physiological signal data while also making targeted music recommendations. While large language models (LLMs) have shown promise in various domains, their application to automated physiological report generation for music therapy represents an unexplored task. We present a prototype system that leverages LLMs to bridge this gap -- transforming raw EEG and cardiovascular data into human-readable therapeutic reports and personalized music recommendations. Unlike prior work focusing on real-time physiological adaptation during listening, our approach emphasizes post-session analysis and interpretable reporting, enabling non-expert users to comprehend their psychophysiological states and track therapeutic outcomes over time. By integrating signal processing modules with LLM-based reasoning agents, the system provides a practical and low-cost solution for short-term progress monitoring in home music therapy contexts. This work demonstrates the feasibility of applying LLMs to a novel task -- democratizing access to physiology-driven music therapy through automated, interpretable reporting.

cs.HC↗

arXiv2026

BasketHAR: A Multimodal Dataset for Human Activity Recognition and Sport Analysis in Basketball Training Scenarios

Human Activity Recognition (HAR) involves the automatic identification of user activities and has gained significant research interest due to its broad applicability. Most HAR systems rely on supervised learning, which necessitates large, diverse, and well-annotated datasets. However, existing datasets predominantly focus on basic activities such as walking, standing, and stair navigation, limiting their utility in specialized contexts like sports performance analysis. To address this gap, we present BasketHAR, a novel multimodal HAR dataset tailored for basketball training, encompassing a diverse set of professional-level actions. BasketHAR includes comprehensive motion data from inertial measurement units (accelerometers and gyroscopes), angular velocity, magnetic field, heart rate, skin temperature, and synchronized video recordings. We also provide a baseline multimodal alignment method to benchmark performance. Experimental results underscore the dataset's complexity and suitability for advanced HAR tasks. Furthermore, we highlight its potential applications in the analysis of basketball training sessions and in the generation of specialized performance reports, representing a valuable resource for future research in HAR and sports analytics. The dataset are publicly accessible at https://huggingface.co/datasets/Xian-Gao/BasketHAR licensed under Apache License 2.0.

cs.CV↗

arXiv2026

Spatially covariant gravity with two degrees of freedom: A perturbative analysis up to cubic order

There has been considerable interest in constructing modified gravity theories that propagate only two degrees of freedom (DOFs), corresponding to the tensorial gravitational waves of general relativity. Within the framework of spatially covariant gravity (SCG), the conditions for obtaining 2-DOF theories can be derived from Hamiltonian constraint analysis, but it is generally difficult to translate those conditions into explicit SCG Lagrangians, especially when the Lagrangian depends nonlinearly on the extrinsic curvature. In this work, we adopt an alternative perturbative approach. We consider polynomial-type SCG Lagrangians up to $d=3$, where $d$ denotes the total number of derivatives in each monomial, and expand them around a cosmological background. By requiring the scalar mode to be eliminated up to cubic order in perturbations, we derive the corresponding conditions on the coefficient functions in the Lagrangian. We find five explicit Lagrangians that propagate only 2 DOFs up to cubic order in perturbations around a cosmological background. These theories therefore provide concrete candidate 2-DOF SCG models, at least at the perturbative level up to cubic order.

gr-qc↗

arXiv2026

DB SwinT: A Dual-Branch Swin Transformer Network for Road Extraction in Optical Remote Sensing Imagery

With the continuous improvement in the spatial resolution of optical remote sensing imagery, accurate road extraction has become increasingly important for applications such as urban planning, traffic monitoring, and disaster management. However, road extraction in complex urban and rural environments remains challenging, as roads are often occluded by trees, buildings, and other objects, leading to fragmented structures and reduced extraction accuracy. To address this problem, this paper proposes a Dual-Branch Swin Transformer network (DB SwinT) for road extraction. The proposed framework combines the long-range dependency modeling capability of the Swin Transformer with the multi-scale feature fusion strategy of U-Net, and employs a dual-branch encoder to learn complementary local and global representations. Specifically, the local branch focuses on recovering fine structural details in occluded areas, while the global branch captures broader semantic context to preserve the overall continuity of road networks. In addition, an Attentional Feature Fusion (AFF) module is introduced to adaptively fuse features from the two branches, further enhancing the representation of occluded road segments. Experimental results on the Massachusetts and DeepGlobe datasets show that DB SwinT achieves Intersection over Union (IoU) scores of 79.35\% and 74.84\%, respectively, demonstrating its effectiveness for road extraction from optical remote sensing imagery.

cs.CV↗

arXiv2026

On the degrees of freedom of spatially covariant vector field theory

We investigate a class of spatially covariant vector field theories on a flat background, where the Lagrangians are constructed as polynomials of first-order derivatives of the vector field. Because Lorentz and $\mathrm{U}(1)$ invariances are broken, such theories generally propagate three degrees of freedom (DOFs): two transverse modes and one longitudinal mode. We examine the conditions under which the additional longitudinal mode is eliminated so that only two DOFs remain. To this end, we perform a Hamiltonian constraint analysis and identify two necessary and sufficient degeneracy conditions that reduce the number of DOFs from three to two. We find three classes of solutions satisfying these degeneracy conditions, corresponding to distinct types of theories. Type-I theories possess one first-class and two second-class constraints, type-II theories have four second-class constraints, and type-III theories contain two first-class constraints. The Maxwell theory is recovered as a special case of the type-III theories, where Lorentz symmetry is restored.

hep-th↗

arXiv2026

Spatially covariant gravity with two degrees of freedom in the presence of an auxiliary scalar field: Hamiltonian analysis

A class of gravity theories respecting spatial covariance and in the presence of non-dynamical auxiliary scalar fields with only spatial derivatives is investigated. Generally, without higher temporal derivatives in the metric sector, there are 3 degrees of freedom (DOFs) propagating due to the breaking of general covariance. Through a Hamiltonian constraint analysis, we examine the conditions to eliminate the scalar DOF such that only 2 DOFs, which correspond the tensorial gravitational waves in a homogeneous and isotropic background, are propagating. We find that two conditions are needed, each of which can eliminate half degree of freedom. The second condition can be further classified into two cases according to its effect on the Dirac matrix. We also apply the formal conditions to a polynomial-type Lagrangian as a concrete example, in which all the monomials are spatially covariant scalars containing two derivatives. Our results are consistent with the previous analysis based on the perturbative method.

gr-qc↗

arXiv2026

Scalar-induced gravitational waves in spatially covariant gravity

We investigate scalar-induced gravitational waves (SIGWs) in the framework of spatially covariant gravity (SCG), a broad class of Lorentz-violating modified gravity theories respecting only spatial diffeomorphism invariance. Extending earlier SCG formulations, we compute the general kernel function for SIGWs on a flat Friedmann-Lemaître-Robertson-Walker background, focusing on polynomial-type SCG Lagrangians up to $d=3$, where $d$ denotes the total number of derivatives in each monomial. We derive explicit expressions for the kernel in the case of power-law time evolution of the coefficients, and restrict attention to the subset of SCG operators whose tensor modes propagate at the speed of light, thereby avoiding late-time divergences in the fractional energy density of SIGWs. Instead of the usual Newtonian gauge, the breaking of time reparametrization symmetry in SCG necessitates a unitary gauge analysis. We compute the energy density of SIGWs for representative parameter combinations, finding distinctive deviations from general relativity (GR), including scale-dependent modifications to both the amplitude and the spectral shape. Our results highlight the potential of stochastic GW background measurements to probe spatially covariant gravity and other Lorentz-violating extensions of GR.

gr-qc↗

arXiv2026

Circularly polarized gravitational waves from parity-violating scalar-tensor theory

We study both primordial gravitational waves (GWs) and scalar-induced gravitational waves (SIGWs) in a class of the parity-violating scalar-tensor (PVST) theory, of which the Lagrangian is the linear combination of seven ghost-free parity-violating scalar-tensor monomials dubbed the "Qi-Xiu" Lagrangians. At linear order, we obtain the quadratic action for tensor perturbations and show that parity-violating terms associated with L_1, L_2, L_5, L_6, and L_7 render the tensor propagation polarization dependent, leading to chiral primordial spectra and a nonvanishing degree of circular polarization. At second order, we derive the equation of motion for SIGWs and identify the explicit parity-violating source terms. In particular, L_3 and L_4 enter exclusively through the source term for SIGWs, allowing parity violation to arise even when the linear GWs' propagation remains effectively general-relativity-like. During the radiation-dominated era, we compute the fractional energy density of SIGWs for both monochromatic and log-normal curvature power spectra. We find that, around the peak frequency, SIGWs in PVST gravity exhibit characteristic deviations from those in general relativity, resulting in a nonzero degree of circular polarization.

gr-qc↗

arXiv2026

Beyond Uniform SVD:Dual-Level Optimization across Columns and Modules for LLM Compression

Low-rank decomposition, particularly Singular Value Decomposition (SVD), is a pivotal technique for mitigating the storage and computational demands of Large Language Models (LLMs). However, prevalent SVD-based approaches overlook the critical phenomenon that decomposition errors exhibit significant disparity across different components of the parameter matrix, often leading to suboptimal approximation. Furthermore, existing methods lack a direct metric to evaluate the importance of individual weight matrices. To address these limitations, we propose Duo-SVD (Dual-level Optimization SVD), a novel training-free framework that synergizes optimization at both the column and the module levels. First, Duo-SVD incorporates a Column-Preserving Strategy that explicitly retains columns exhibiting high decomposition errors, while applying low-rank approximation solely to those with lower errors. Second, at the module level, we employ a Module-Adaptive Allocation Strategy that formulates ratio allocation as a global constrained optimization problem based on perturbation-induced model deviation. Extensive experiments demonstrate that Duo-SVD consistently outperforms state-of-the-art SVD-based baselines and structured pruning methods, establishing it as a superior paradigm for efficient LLM compression.

cs.LG↗

arXiv2026

OnlineMate: An LLM-Based Multi-Agent Companion System for Cognitive Support in Online Learning

In online learning environments, students often lack personalized peer interactions, which are crucial for cognitive development and learning engagement. Although previous studies have employed large language models (LLMs) to simulate interactive learning environments, these interactions are limited to conversational exchanges, failing to adapt to learners' individualized cognitive and psychological states. As a result, students' engagement is low and they struggle to gain inspiration. To address this challenge, we propose OnlineMate, a multi-agent learning companion system driven by LLMs integrated with Theory of Mind (ToM). OnlineMate simulates peer-like roles, infers learners' psychological states such as misunderstandings and confusion during collaborative discussions, and dynamically adjusts interaction strategies to support higher-order thinking. Comprehensive evaluations, including simulation-based experiments, human assessments, and real classroom trials, demonstrate that OnlineMate significantly promotes deep learning and cognitive engagement by elevating students' average cognitive level while substantially improving emotional engagement scores.

cs.CY↗

arXiv2025

Degrees of freedom of a quadratic scalar-nonmetricity theory

We study the number of degrees of freedom (DOFs) in quadratic scalar-nonmetricity (QSN) theory, whose Lagrangian is the linear combination of five quadratic nonmetricity invariants with coefficients depending on a dynamical scalar field. Working in the coincident gauge, we perform the Arnowitt-Deser-Misner decomposition and classify QSN models into 13 cases according to the numbers of their primary constraints. For cases that are physically viable in the sense that both a consistent cosmological background and tensor gravitational waves exist, we count the number of DOFs based on two approaches. First, we investigate the linear cosmological perturbations around a Friedmann-Lema\^{\i}tre-Robertson-Walker background. Then we perform a Dirac-Bergmann Hamiltonian constraint analysis to count the number of DOFs at the nonperturbative level. We focus on three representative cases. In case II, both the perturbative and nonperturbative approaches yield the same result, which indicates that the theory propagates 10 degrees of freedom. In contrast, in cases V and VI, the Hamiltonian analysis yields 8 degrees of freedom, while only 6 and 5 modes are visible at linear order in perturbations, respectively. This indicates that additional modes are strongly coupled on cosmological backgrounds.

gr-qc↗

arXiv2025

ARA: Adaptive Rank Allocation for Efficient Large Language Model SVD Compression

In the field of large language model (LLM) compression, singular value decomposition (SVD) is a widely studied and adopted low-rank decomposition technique. Since SVD operates exclusively on linear modules, and these modules in LLMs are separated by nonlinear components, SVD can only be applied independently to each linear module. Under a global compression ratio constraint, determining the appropriate rank for different linear modules becomes a critical problem. Existing approaches, such as heuristic algorithms and mask-based training, have made progress in addressing this challenge. However, these methods still suffer from several limitations: heuristic algorithms explore the solution space within restricted regions, while mask-based training struggles to efficiently capture the relationship between singular value spectra and trainable parameters. More importantly, current methods overlook the key property that the gain function is non-smooth at a compression ratio of 1, which often leads the training process to suboptimal local minima. To address these issues, we propose an Adaptive Rank Allocation (ARA) method. Specifically, (1) ARA introduces a dedicated mask design that enables efficient mapping and updating between retained ranks and trainable parameters; and (2) it employs an additional loss function to guide parameter selection toward globally optimal solutions. Experimental results demonstrate that ARA achieves state-of-the-art performance. On the LLaMA2-7B model with a 80\% compression ratio, ARA reduces perplexity on WikiText2 from 8.38 to 6.42 and improves average zero-shot task accuracy by 9.72 percentage points compared with uniform compression. These results highlight the effectiveness of our method for rank allocation in SVD-based LLM compression.

cs.LG↗

arXiv2025

MMReview: A Multidisciplinary and Multimodal Benchmark for LLM-Based Peer Review Automation

With the rapid growth of academic publications, peer review has become an essential yet time-consuming responsibility within the research community. Large Language Models (LLMs) have increasingly been adopted to assist in the generation of review comments; however, current LLM-based review tasks lack a unified evaluation benchmark to rigorously assess the models' ability to produce comprehensive, accurate, and human-aligned assessments, particularly in scenarios involving multimodal content such as figures and tables. To address this gap, we propose \textbf{MMReview}, a comprehensive benchmark that spans multiple disciplines and modalities. MMReview includes multimodal content and expert-written review comments for 240 papers across 17 research domains within four major academic disciplines: Artificial Intelligence, Natural Sciences, Engineering Sciences, and Social Sciences. We design a total of 13 tasks grouped into four core categories, aimed at evaluating the performance of LLMs and Multimodal LLMs (MLLMs) in step-wise review generation, outcome formulation, alignment with human preferences, and robustness to adversarial input manipulation. Extensive experiments conducted on 16 open-source models and 5 advanced closed-source models demonstrate the thoroughness of the benchmark. We envision MMReview as a critical step toward establishing a standardized foundation for the development of automated peer review systems.

cs.CL↗

arXiv2025

GoAI: Enhancing AI Students' Learning Paths and Idea Generation via Graph of AI Ideas

With the rapid advancement of artificial intelligence technology, AI students are confronted with a significant "information-to-innovation" gap: they must navigate through the rapidly expanding body of literature, trace the development of a specific research field, and synthesize various techniques into feasible innovative concepts. An additional critical step for students is to identify the necessary prerequisite knowledge and learning paths. Although many approaches based on large language models (LLMs) can summarize the content of papers and trace the development of a field through citations, these methods often overlook the prerequisite knowledge involved in the papers and the rich semantic information embedded in the citation relationships between papers. Such information reveals how methods are interrelated, built upon, extended, or challenged. To address these limitations, we propose GoAI, a tool for constructing educational knowledge graphs from AI research papers that leverages these graphs to plan personalized learning paths and support creative ideation. The nodes in the knowledge graph we have built include papers and the prerequisite knowledge, such as concepts, skills, and tools, that they involve; the edges record the semantic information of citations. When a student queries a specific paper, a beam search-based path search method can trace the current development trends of the field from the queried paper and plan a learning path toward cutting-edge objectives. The integrated Idea Studio guides students to clarify problem statements, compare alternative designs, and provide formative feedback on novelty, clarity, feasibility, and alignment with learning objectives.

cs.AI↗