arXiv subjects

Siyuan Chen

Publications and source records attributed to Siyuan Chen.

At least 91 records · Page 5Linked to original sources

arXiv2025

Quasars can Signpost Supermassive Black Hole Binaries

Supermassive black holes (SMBHs) are found in the centers of massive galaxies, and galaxy mergers should eventually lead to SMBH mergers. Quasar activity has long been associated with galaxy mergers, so here we investigate if supermassive black hole binaries (SMBHBs) are preferentially found in quasars. Our multimessenger investigation folds together a gravitational wave background signal from NANOGrav, a sample of periodic AGN candidates from the Catalina Real-Time Transient Survey, and a quasar mass function, to estimate an upper limit on the fraction of quasars which could host a SMBHB. We find at 95\% confidence that quasars are at most five times as likely to host a SMBHB as a random galaxy. Pulsar timing arrays may therefore be more likely to find SMBHBs by prioritizing quasars over a random selection of galaxies in their searches.

astro-ph.HE↗

arXiv2025

SLOs-Serve: Optimized Serving of Multi-SLO LLMs

This paper introduces SLOs-Serve, a system designed for serving multi-stage large language model (LLM) requests with application- and stage-specific service level objectives (SLOs). The key idea behind SLOs-Serve is to customize the allocation of tokens to meet these SLO requirements. SLOs-Serve uses a multi-SLO dynamic programming-based algorithm to continuously optimize token allocations under SLO constraints by exploring the full design space of chunked prefill and (optional) speculative decoding. Leveraging this resource planning algorithm, SLOs-Serve effectively supports multi-SLOs and multi-replica serving with dynamic request routing while being resilient to bursty arrivals. Our evaluation across 6 LLM application scenarios (including summarization, coding, chatbot, tool calling, and reasoning) demonstrates that SLOs-Serve improves per-GPU serving capacity by 2.2x on average compared to prior state-of-the-art systems.

cs.DC↗

arXiv2025

Learnable Group Transform: Enhancing Genotype-to-Phenotype Prediction for Rice Breeding with Small, Structured Datasets

Genotype-to-Phenotype (G2P) prediction plays a pivotal role in crop breeding, enabling the identification of superior genotypes based on genomic data. Rice (Oryza sativa), one of the most important staple crops, faces challenges in improving yield and resilience due to the complex genetic architecture of agronomic traits and the limited sample size in breeding datasets. Current G2P prediction methods, such as GWAS and linear models, often fail to capture complex non-linear relationships between genotypes and phenotypes, leading to suboptimal prediction accuracy. Additionally, population stratification and overfitting are significant obstacles when models are applied to small datasets with diverse genetic backgrounds. This study introduces the Learnable Group Transform (LGT) method, which aims to overcome these challenges by combining the advantages of traditional linear models with advanced machine learning techniques. LGT utilizes a group-based transformation of genotype data to capture spatial relationships and genetic structures across diverse rice populations, offering flexibility to generalize even with limited data. Through extensive experiments on the Rice529 dataset, a panel of 529 rice accessions, LGT demonstrated substantial improvements in prediction accuracy for multiple agronomic traits, including yield and plant height, compared to state-of-the-art baselines such as linear models and recent deep learning approaches. Notably, LGT achieved an R^2 improvement of up to 15\% for yield prediction, significantly reducing error and demonstrating its ability to extract meaningful signals from high-dimensional, noisy genomic data. These results highlight the potential of LGT as a powerful tool for genomic prediction in rice breeding, offering a promising solution for accelerating the identification of high-yielding and resilient rice varieties.

q-bio.GN↗

arXiv2025

Design of a CsI(Tl) Calorimeter for Muonium-to-Antimuonium Conversion Experiment

The Muonium-to-Antimuonium Conversion Experiment (MACE) is proposed to search for charged lepton flavor violation and increase the sensitivity by more than two orders of magnitude compared to the MACS experiment at PSI in 1999. A clear signature of this conversion is the positron produced from antimuonium decay. This paper presents a near-$4π$-coverage calorimeter designed for MACE, which can provide an energy resolution of 10.8% at 511 keV, and a signal efficiency of 78.3% for annihilation $γ$-ray events. Detailed Monte-Carlo simulations using MACE offline software based on Geant4 are performed for geometry optimization, coincidence system design, background estimation, and benchmark detector validation.

physics.ins-det↗

arXiv2025

Genotype-to-Phenotype Prediction in Rice with High-Dimensional Nonlinear Features

Genotype-to-Phenotype prediction can promote advances in modern genomic research and crop improvement, guiding precision breeding and genomic selection. However, high-dimensional nonlinear features often hinder the accuracy of genotype-to-phenotype prediction by increasing computational complexity. The challenge also limits the predictive accuracy of traditional approaches. Therefore, effective solutions are needed to improve the accuracy of genotype-to-phenotype prediction. In our paper, we propose MLFformer. MLFformer is a Transformer-based architecture that incorporates the Fast Attention mechanism and a multilayer perceptron module to handle high-dimensional nonlinear features. In MLFformer, the Fast Attention mechanism is utilized to handle computational complexity and enhance processing efficiency. In addition, the MLP structure further captures high-dimensional nonlinear features. Through experiments, the results show that MLFformer reduces the average MAPE by 7.73% compared to the vanilla Transformer. In univariate and multivariate prediction scenarios, MLFformer achieves the best predictive performance among all compared models.

q-bio.GN↗

arXiv2025

RTBAS: Defending LLM Agents Against Prompt Injection and Privacy Leakage

Tool-Based Agent Systems (TBAS) allow Language Models (LMs) to use external tools for tasks beyond their standalone capabilities, such as searching websites, booking flights, or making financial transactions. However, these tools greatly increase the risks of prompt injection attacks, where malicious content hijacks the LM agent to leak confidential data or trigger harmful actions. Existing defenses (OpenAI GPTs) require user confirmation before every tool call, placing onerous burdens on users. We introduce Robust TBAS (RTBAS), which automatically detects and executes tool calls that preserve integrity and confidentiality, requiring user confirmation only when these safeguards cannot be ensured. RTBAS adapts Information Flow Control to the unique challenges presented by TBAS. We present two novel dependency screeners, using LM-as-a-judge and attention-based saliency, to overcome these challenges. Experimental results on the AgentDojo Prompt Injection benchmark show RTBAS prevents all targeted attacks with only a 2% loss of task utility when under attack, and further tests confirm its ability to obtain near-oracle performance on detecting both subtle and direct privacy leaks.

cs.CR↗

arXiv2025

Chinese Pulsar Timing Array upper limits on microhertz gravitational waves from supermassive black-hole binaries using PSR J1713+0747 FAST data

We derive the gravitational-wave (GW) strain upper limits from resolvable supermassive black-hole binaries using the data from the Five-hundred-meter Aperture Spherical radio Telescope (FAST), in the context of the Chinese Pulsar Timing Array project. We focus on circular orbits in the $μ$Hz GW frequency band between $10^{-7}$ and $3\times10^{-6}$ Hz. This frequency band is higher than the traditional pulsar timing array band and is less explored. We used the data of the millisecond pulsar PSR J1713+5307 observed between August 2019 and April 2021. A dense observation campaign was carried out in September 2020 to allow for the $μ$Hz band coverage. Our sky-average continuous source upper limit at the 95% confidence level at 1$μ$Hz is 1.26$\times10^{-12}$, while the same limit in the direction of the pulsar is 4.77$\times10^{-13}$.

astro-ph.HE↗

arXiv2025

Practical offloading for fine-tuning LLM on commodity GPU via learned sparse projectors

Fine-tuning large language models (LLMs) requires significant memory, often exceeding the capacity of a single GPU. A common solution to this memory challenge is offloading compute and data from the GPU to the CPU. However, this approach is hampered by the limited bandwidth of commodity hardware, which constrains communication between the CPU and GPU, and by slower matrix multiplications on the CPU. In this paper, we present an offloading framework, LSP-Offload, that enables near-native speed LLM fine-tuning on commodity hardware through learned sparse projectors. Our data-driven approach involves learning efficient sparse compressors that minimize communication with minimal precision loss. Additionally, we introduce a novel layer-wise communication schedule to maximize parallelism between communication and computation. As a result, our framework can fine-tune a 1.3 billion parameter model on a 4GB laptop GPU and a 6.7 billion parameter model on a 24GB NVIDIA RTX 4090 GPU. Compared to state-of-the-art offloading frameworks, our approach reduces end-to-end fine-tuning time by 33.1%-62.5% when converging to the same accuracy. We open source our framework at https://github.com/gulang2019/LSP-Offload.

cs.DC↗

arXiv2025

Strongly correlated states of transition metal spin defects: the case of an iron impurity in aluminum nitride

We investigate the electronic properties of an exemplar transition metal impurity in an insulator, with the goal of accurately describing strongly correlated, defect states. We consider iron in aluminum nitride, a material of interest for hybrid quantum technologies, and we carry out calculations with quantum embedding methods -- density matrix embedding theory (DMET) and quantum defect embedding theory (QDET) and with spin-flip time-dependent density functional theory (TDDFT). We show that both DMET and QDET accurately describe the ground state and low-lying excited states of the defect, and that TDDFT yields photoluminescence spectra in agreement with experiments. In addition, we provide a detailed discussion of the convergence of our results as a function of the active space used in the embedding methods, thus defining a protocol to obtain converged data, directly comparable with experiments.

cond-mat.str-el↗

arXiv2025

Reinforcement Learning Constrained Beam Search for Parameter Optimization of Paper Drying Under Flexible Constraints

Existing approaches to enforcing design constraints in Reinforcement Learning (RL) applications often rely on training-time penalties in the reward function or training/inference-time invalid action masking, but these methods either cannot be modified after training, or are limited in the types of constraints that can be implemented. To address this limitation, we propose Reinforcement Learning Constrained Beam Search (RLCBS) for inference-time refinement in combinatorial optimization problems. This method respects flexible, inference-time constraints that support exclusion of invalid actions and forced inclusion of desired actions, and employs beam search to maximize sequence probability for more sensible constraint incorporation. RLCBS is extensible to RL-based planning and optimization problems that do not require real-time solution, and we apply the method to optimize process parameters for a novel modular testbed for paper drying. An RL agent is trained to minimize energy consumption across varying machine speed levels by generating optimal dryer module and air supply temperature configurations. Our results demonstrate that RLCBS outperforms NSGA-II under complex design constraints on drying module configurations at inference-time, while providing a 2.58-fold or higher speed improvement.

cs.LG↗

arXiv2025

Balanced Residual Distillation Learning for 3D Point Cloud Class-Incremental Semantic Segmentation

Class-incremental learning (CIL) enables continuous learning of new classes while mitigating catastrophic forgetting of old ones. For the performance breakthrough of CIL, it is essential yet challenging to effectively refine past knowledge from the base model and balance it with new learning. However, such a challenge has not been considered in current research. This work proposes a balanced residual distillation learning framework (BRDL) to address this gap and advance CIL performance. BRDL introduces a residual distillation strategy to dynamically refine past knowledge by expanding the network structure and a balanced pseudo-label learning strategy to mitigate class bias and balance learning between old and new classes. We apply the proposed BRDL to a challenging 3D point cloud semantic segmentation task where the data is unordered and unstructured. Extensive experimental results demonstrate that BRDL sets a new benchmark with an outstanding balance capability in class-biased scenarios.

cs.CV↗

arXiv2024

SG-Splatting: Accelerating 3D Gaussian Splatting with Spherical Gaussians

3D Gaussian Splatting is emerging as a state-of-the-art technique in novel view synthesis, recognized for its impressive balance between visual quality, speed, and rendering efficiency. However, reliance on third-degree spherical harmonics for color representation introduces significant storage demands and computational overhead, resulting in a large memory footprint and slower rendering speed. We introduce SG-Splatting with Spherical Gaussians based color representation, a novel approach to enhance rendering speed and quality in novel view synthesis. Our method first represents view-dependent color using Spherical Gaussians, instead of three degree spherical harmonics, which largely reduces the number of parameters used for color representation, and significantly accelerates the rendering process. We then develop an efficient strategy for organizing multiple Spherical Gaussians, optimizing their arrangement to achieve a balanced and accurate scene representation. To further improve rendering quality, we propose a mixed representation that combines Spherical Gaussians with low-degree spherical harmonics, capturing both high- and low-frequency color information effectively. SG-Splatting also has plug-and-play capability, allowing it to be easily integrated into existing systems. This approach improves computational efficiency and overall visual fidelity, making it a practical solution for real-time applications.

cs.CV↗

arXiv2024

Solving the Poisson Yang-Baxter equation via deformation-to-quasiclassical-limits

A fundamental construction of Poisson algebras is to derive them as the quasiclassical limits (QCLs) of associative algebra deformations of commutative associative algebras. This paper lifts this process to the level of classical Yang-Baxter type equations. Solutions of the Poisson Yang-Baxter equation (PYBE) in Poisson algebras is obtained by scalarly deforming solutions of the associative Yang-Baxter equation (AYBE) in commutative associative algebras. Inspired by the characterization of solutions of various classical Yang-Baxter type equations by $\mathcal{O}$-operators, we introduce the notions of deformations of (bi)module algebras and scalar deformations of the corresponding $\mathcal{O}$-operators, from which the QCLs give $\mathcal{O}$-operators for Poisson algebras, which in turn provide solutions of the PYBE. Furthermore, as the QCLs of tridendriform algebra deformations of commutative tridendriform algebras, post-Poisson algebras produce deformations-QCLs of $\mathcal{O}$-operators for Poisson algebras, thus offering explicit solutions of the PYBE.

math.QA↗

arXiv2024

Deep Neural Emulation of the Supermassive Black-hole Binary Population

While supermassive black-hole (SMBH)-binaries are not the only viable source for the low-frequency gravitational wave background (GWB) signal evidenced by the most recent pulsar timing array (PTA) data sets, they are expected to be the most likely. Thus, connecting the measured PTA GWB spectrum and the underlying physics governing the demographics and dynamics of SMBH-binaries is extremely important. Previously, Gaussian processes (GPs) and dense neural networks have been used to make such a connection by being built as conditional emulators; their input is some selected evolution or environmental SMBH-binary parameters and their output is the emulated mean and standard deviation of the GWB strain ensemble distribution over many Universes. In this paper, we use a normalizing flow (NF) emulator that is trained on the entirety of the GWB strain ensemble distribution, rather than only mean and standard deviation. As a result, we can predict strain distributions that mirror underlying simulations very closely while also capturing frequency covariances in the strain distributions as well as statistical complexities such as tails, non-Gaussianities, and multimodalities that are otherwise not learnable by existing techniques. In particular, we feature various comparisons between the NF-based emulator and the GP approach used extensively in past efforts. Our analyses conclude that the NF-based emulator not only outperforms GPs in the ease and computational cost of training but also outperforms in the fidelity of the emulated GWB strain ensemble distributions.

astro-ph.IM↗

arXiv2024

Distinguishing the Demographics of Compact Binaries with Merger Entropy Index

The coalescence of binary black holes and neutron stars increases the entropy in the universe. The release of entropy from the inspiral stage to the merger depends primarily on the mass and spin vectors of the compact binary. In this study, we report a novel application of entropy to study the demographics of the compact binaries reported by the LIGO-Virgo-KAGRA (LVK) Collaboration. We compute an astrophysical distribution of the Merger Entropy Index ($\mathcal{I}_\mathrm{BBH}$) - a mass-independent measure of the efficiency of entropy transfer for black hole binaries - for all the events reported in the LVK Gravitational-Wave Transient Catalogs. We derive $\mathcal{I}_\mathrm{BBH}$ for six astrophysically motivated population models describing dynamical and isolated formation channels. We find that $\mathcal{I}_\mathrm{BBH}$ offers a new criterion to probe the formation channels of LVK events with compact objects in the upper $(\gtrsim 60~M_\odot)$ and lower ($\lesssim 5~M_\odot$) mass-gaps. For GW190521, an event with both objects in the upper mass gap, $\mathcal{I}_\mathrm{BBH}$ distribution strongly favors second-generation mergers. For GW230529, a new event with the primary object in the lower mass gap, we note that $\mathcal{I}_\mathrm{BBH}$ mildly favors it with neutron star - black holes events. Our work provides a new framework to study the underlying demographics of compact binaries in the data-rich era of gravitational-wave astronomy.

gr-qc↗

arXiv2024

A Multi-Task Role-Playing Agent Capable of Imitating Character Linguistic Styles

The advent of large language models (LLMs) has significantly propelled the advancement of Role-Playing Agents (RPAs). However, current Role-Playing Agents predominantly focus on mimicking a character's fundamental attributes while neglecting the replication of linguistic style, and they are incapable of effectively replicating characters when performing tasks beyond multi-turn dialogues, which results in generated responses that lack authenticity. The reason current RPAs lack this capability is due to the nature of existing character datasets, which lack collections of character quotations and are limited to multi-turn dialogue tasks, constraining the RPA's performance across other task domains and failing to mimic a character's linguistic style. To address this gap, we developed a multi-task role-playing dataset named MRstyle, which encompasses a substantial number of real individuals along with their quotations and covers seven different tasks. On this basis, we develop StyleRPA, a Multi-Task Role-Playing Agent (MRPA) that significantly outperforms recent open-source LLMs and RPAs baselines on 7 tasks including Dialogue, Dictionary, Composition, Story Generation, Product Description, Music Commentary, and Open Question Answering. The code and data will be released.

cs.CL↗

arXiv2024

Heterogeneous Interaction Modeling With Reduced Accumulated Error for Multi-Agent Trajectory Prediction

Dynamical complex systems composed of interactive heterogeneous agents are prevalent in the world, including urban traffic systems and social networks. Modeling the interactions among agents is the key to understanding and predicting the dynamics of the complex system, e.g., predicting the trajectories of traffic participants in the city. Compared with interaction modeling in homogeneous systems such as pedestrians in a crowded scene, heterogeneous interaction modeling is less explored. Worse still, the error accumulation problem becomes more severe since the interactions are more complex. To tackle the two problems, this paper proposes heterogeneous interaction modeling with reduced accumulated error for multi-agent trajectory prediction. Based on the historical trajectories, our method infers the dynamic interaction graphs among agents, featured by directed interacting relations and interacting effects. A heterogeneous attention mechanism is defined on the interaction graphs for aggregating the influence from heterogeneous neighbors to the target agent. To alleviate the error accumulation problem, this paper analyzes the error sources from the spatial and temporal perspectives, and proposes to introduce the graph entropy and the mixup training strategy for reducing the two types of errors respectively. Our method is examined on three real-world datasets containing heterogeneous agents, and the experimental results validate the superiority of our method.

cs.MA↗

arXiv2024

A Hierarchical Framework with Spatio-Temporal Consistency Learning for Emergence Detection in Complex Adaptive Systems

Emergence, a global property of complex adaptive systems (CASs) constituted by interactive agents, is prevalent in real-world dynamic systems, e.g., network-level traffic congestions. Detecting its formation and evaporation helps to monitor the state of a system, allowing to issue a warning signal for harmful emergent phenomena. Since there is no centralized controller of CAS, detecting emergence based on each agent's local observation is desirable but challenging. Existing works are unable to capture emergence-related spatial patterns, and fail to model the nonlinear relationships among agents. This paper proposes a hierarchical framework with spatio-temporal consistency learning to solve these two problems by learning the system representation and agent representations, respectively. Spatio-temporal encoders composed of spatial and temporal transformers are designed to capture agents' nonlinear relationships and the system's complex evolution. Agents' and the system's representations are learned to preserve the spatio-temporal consistency by minimizing the spatial and temporal dissimilarities in a self-supervised manner in the latent space. Our method achieves more accurate detection than traditional methods and deep learning methods on three datasets with well-known yet hard-to-detect emergent behaviors. Notably, our hierarchical framework is generic in incorporating other deep learning methods for agent-level and system-level detection.

cs.MA↗