arXiv subjects

Yitian Chen

Publications and source records attributed to Yitian Chen.

At least 19 recordsLinked to original sources

arXiv2026

Omni-Flow: A Unified Workflow Orchestration and Distributed KV Cache Sharing Framework for Multimodal Inference

As large language model (LLM) inference evolves from text-only to multimodal paradigms, inference systems face three challenges: (1) flexible orchestration of multimodal workflows, where heterogeneous computing units exhibit complex dependencies and concurrent control; (2) efficient transmission of massive intermediate data across processes and nodes, with tensors flowing at high speed among heterogeneous roles; and (3) efficient sharing of KV caches and model weights across roles to eliminate redundant GPU memory. Existing solutions deploy LLMs and diffusion models independently, lacking a system-level abstraction for multimodal pipelines; this scatters orchestration logic, tightly couples transmission paths to specific models, and incurs high cost to integrate new models. To address these challenges, we present Omni-Flow, a distributed scheduling framework for multimodal inference through a three-layer abstraction. The Control Flow layer defines workflows via a Python DSL, orchestrating heterogeneous units into a unified dataflow graph that supports static DAGs and dynamic routing, with built-in service discovery and diverse load-balancing strategies. The Data Flow layer provides a distributed KV cache abstraction beyond prefill/decode separation, unifying allocation and enabling direct cross-role transmission across a three-tier paged storage hierarchy (GPU/CPU/SSD) over zero-copy, low-latency channels. The Compute Flow layer supports complex multimodal prefix matching for KV reuse across multi-turn dialogues, and takes over KV cache and sampling logic via a unified SGLang interface, letting diffusion models directly reuse the LLM forward path under unified parallel semantics. We demonstrate that Omni-Flow supports diverse heterogeneous scenarios with a consistent programming model, including omni-modal dialogue (LongCat-Next) and complex image generation pipelines (HunyuanImage-3).

cs.DC↗

arXiv2026

LongCat-Next: Lexicalizing Modalities as Discrete Tokens

The prevailing Next-Token Prediction (NTP) paradigm has driven the success of large language models through discrete autoregressive modeling. However, contemporary multimodal systems remain language-centric, often treating non-linguistic modalities as external attachments, leading to fragmented architectures and suboptimal integration. To transcend this limitation, we introduce Discrete Native Autoregressive (DiNA), a unified framework that represents multimodal information within a shared discrete space, enabling a consistent and principled autoregressive modeling across modalities. A key innovation is the Discrete Native Any-resolution Visual Transformer (dNaViT), which performs tokenization and de-tokenization at arbitrary resolutions, transforming continuous visual signals into hierarchical discrete tokens. Building on this foundation, we develop LongCat-Next, a native multimodal model that processes text, vision, and audio under a single autoregressive objective with minimal modality-specific design. As an industrial-strength foundation model, it excels at seeing, painting, and talking within a single framework, achieving strong performance across a wide range of multimodal benchmarks. In particular, LongCat-Next addresses the long-standing performance ceiling of discrete vision modeling on understanding tasks and provides a unified approach to effectively reconcile the conflict between understanding and generation. As an attempt toward native multimodality, we open-source the LongCat-Next and its tokenizers, hoping to foster further research and development in the community. GitHub: https://github.com/meituan-longcat/LongCat-Next

cs.CV↗

arXiv2026

RSA-Bench: Benchmarking Audio Large Models in Real-World Acoustic Scenarios

While Audio Large Models (ALMs) have achieved remarkable proficiency, their robustness remains brittle in real-world deployment. Existing evaluations largely rely on synthetic Gaussian noise or simplistic single-source interference, failing to capture the intricate, multi-layered acoustic dynamics -- or ``Acoustic Ecology'' -- that characterize authentic physical environments. To bridge this ecological gap, we introduce \textbf{RSA-Bench}, a comprehensive robustness benchmark designed to stress-test ALLMs through high-fidelity auditory scene simulations. Unlike traditional methods, we construct evaluation samples by naturally superimposing diverse environmental soundscapes -- spanning \textit{Pasture}, \textit{Extreme Weather}, \textit{Classroom}, and \textit{Outdoors} -- onto clean speech signals across a spectrum of interference intensities. By evaluating models on six core tasks ranging from fundamental perception to complex reasoning, our study unveils three macro-level insights: \textbf{(I) The Perception-Cognition Gap:} Models maintain relative resilience in low-level recognition but suffer a \textbf{functional collapse} in high-order reasoning tasks under stress; \textbf{(II) Scenario Sensitivity:} ``Vocal-like'' interference (e.g., background laughter) proves significantly more destructive than mechanical noise, challenging the model's auditory attention mechanisms; and \textbf{(III) The Denoising Paradox:} Standard speech enhancement often exacerbates performance degradation, as ALLMs prove highly sensitive to the semantic distortions introduced by denoising artifacts.

cs.SD↗

arXiv2026

OPT-Engine: Benchmarking the Limits of LLMs in Optimization Modeling via Complexity Scaling

We investigate the capabilities and scalability of Large Language Models (LLMs) in optimization modeling, a domain requiring structured reasoning and precise formulation. To this end, we introduce OPT-ENGINE, an extensible benchmark framework with quantifiable and controllable complexity. OPT-ENGINE spans ten canonical Operations Research problems, systematically scaling from Linear Programming to Mixed-Integer Programming, providing a structured environment to probe the limits of automated problem formulation and solving. Utilizing OPT-Engine, we address three pivotal research questions. First, we examine whether Pure-Text Reasoning (PTR) via classical Chain-of-Thought can efficiently tackle optimization tasks, finding that PTR suffers from a critical robustness gap as task complexity increases. Second, we examine whether integrating external computational tools can mitigate PTR's arithmetic weaknesses and improve performance. Our results indicate that while such tools help with local calculations, they still fail to adhere to global optimization constraints. Finally, we pinpoint that for the current SOTA paradigm, Solver-integrated Reasoning (SIR), the automated formulation of constraints represents the primary bottleneck. These findings clarify the limitations of current paradigms and provide a structured roadmap for developing next-generation LLMs for optimization modeling. We release our code and data to facilitate future research (https://github.com/Cardinal-Operations/OPTEngine).

cs.CL↗

arXiv2026

HearSay Benchmark: Do Audio LLMs Leak What They Hear?

While Audio Large Language Models (ALLMs) have achieved remarkable progress in understanding and generation, their potential privacy implications remain largely unexplored. This paper takes the first step to investigate whether ALLMs inadvertently leak user privacy solely through acoustic voiceprints and introduces $\textit{HearSay}$, a comprehensive benchmark constructed from over 22,000 real-world audio clips. To ensure data quality, the benchmark is meticulously curated through a rigorous pipeline involving automated profiling and human verification, guaranteeing that all privacy labels are grounded in factual records. Extensive experiments on $\textit{HearSay}$ yield three critical findings: $\textbf{Significant Privacy Leakage}$: ALLMs inherently extract private attributes from voiceprints, reaching 92.89% accuracy on gender and effectively profiling social attributes. $\textbf{Insufficient Safety Mechanisms}$: Alarmingly, existing safeguards are severely inadequate; most models fail to refuse privacy-intruding requests, exhibiting near-zero refusal rates for physiological traits. $\textbf{Reasoning Amplifies Risk}$: Chain-of-Thought (CoT) reasoning exacerbates privacy risks in capable models by uncovering deeper acoustic correlations. These findings expose critical vulnerabilities in ALLMs, underscoring the urgent need for targeted privacy alignment. The codes and dataset are available at https://github.com/JinWang79/HearSay_Benchmark

cs.CL↗

arXiv2025

Towards Mitigating Excessive Forgetting in LLM Unlearning via Entanglement-Guidance with Proxy Constraint

Large language models (LLMs) are trained on massive datasets that may include private or copyrighted content. Due to growing privacy and ownership concerns, data owners may request the removal of their data from trained models. Machine unlearning provides a practical solution by removing the influence of specific data without full retraining. However, most existing methods still suffer from over-unlearning due to the lack of a principled mechanism to regulate the forgetting boundary, leading to unnecessary utility degradation and heightened privacy and robustness risks. In this work, we propose EGUP (Entanglement-Guided Unlearning with Proxy Constraint), a novel framework that leverages entanglement and proxy constraint to guide the unlearning process while mitigating over-unlearning. Within each iteration, EGUP employs inter-sample entanglement to adaptively reweight the unlearning strength, assigning greater unlearning efforts to forget samples that are semantically closer to retained knowledge. Across iterations, EGUP leverages intra-sample entanglement to track the representation shift of each forget sample and dynamically adjust its unlearning effort. In addition, we incorporate a proxy constraint that approximates the model's expected outputs after unlearning, forming a reference boundary that softly regularizes the unlearning process. EGUP is compatible with existing gradient-based objectives and serves as a plug-and-play enhancement. We evaluate EGUP on the TOFU and MUSE benchmarks, demonstrating consistent improvements in the unlearning-utility trade-off across multiple LLMs. Moreover, EGUP achieves performance close to the retrained model while remaining scalable and robust.

cs.LG↗

arXiv2025

Auto-Formulating Dynamic Programming Problems with Large Language Models

Dynamic programming (DP) is a fundamental method in operations research, but formulating DP models has traditionally required expert knowledge of both the problem context and DP techniques. Large Language Models (LLMs) offer the potential to automate this process. However, DP problems pose unique challenges due to their inherently stochastic transitions and the limited availability of training data. These factors make it difficult to directly apply existing LLM-based models or frameworks developed for other optimization problems, such as linear or integer programming. We introduce DP-Bench, the first benchmark covering a wide range of textbook-level DP problems to enable systematic evaluation. We present Dynamic Programming Language Model (DPLM), a 7B-parameter specialized model that achieves performance comparable to state-of-the-art LLMs like OpenAI's o1 and DeepSeek-R1, and surpasses them on hard problems. Central to DPLM's effectiveness is DualReflect, our novel synthetic data generation pipeline, designed to scale up training data from a limited set of initial examples. DualReflect combines forward generation for diversity and backward generation for reliability. Our results reveal a key insight: backward generation is favored in low-data regimes for its strong correctness guarantees, while forward generation, though lacking such guarantees, becomes increasingly valuable at scale for introducing diverse formulations. This trade-off highlights the complementary strengths of both approaches and the importance of combining them.

cs.AI↗

arXiv2025

The SXS Collaboration's third catalog of binary black hole simulations

We present a major update to the Simulating eXtreme Spacetimes (SXS) Collaboration's catalog of binary black hole simulations. Using highly efficient spectral methods implemented in the Spectral Einstein Code (SpEC), we have nearly doubled the total number of binary configurations from 2,018 to 3,756. The catalog now densely covers the parameter space with precessing simulations up to mass ratio $q=8$ and dimensionless spins up to $|\vec{\chi}|\le0.8$ with near-zero eccentricity. The catalog also includes some simulations at higher mass ratios with moderate spin and more than 250 eccentric simulations. We have also deprecated and rerun some simulations from our previous catalog (e.g., simulations run with a much older version of SpEC or that had anomalously high errors in the waveform). The median waveform difference (which is similar to the mismatch) between resolutions over the simulations in the catalog is $4\times10^{-4}$. The simulations have a median of 22 orbits, while the longest simulation has 148 orbits. We have corrected each waveform in the catalog to be in the binary's center-of-mass frame and exhibit gravitational-wave memory. We estimate the total CPU cost of all simulations in the catalog to be 480,000,000 core-hours. We find that using spectral methods for binary black hole simulations is over 1,000 times more efficient than much shorter finite-difference simulations of comparable accuracy. The full catalog is publicly available through the sxs Python package and at https://data.black-holes.org .

gr-qc↗

arXiv2025

Solver-Informed RL: Grounding Large Language Models for Authentic Optimization Modeling

Optimization modeling is fundamental to decision-making across diverse domains. Despite progress in automating optimization formulation from natural language descriptions, Large Language Models (LLMs) often struggle to generate formally correct and usable models against hallucinations, posing a challenge for reliable automation. Inspired by the success of Reinforcement Learning (RL) in enhancing Large Reasoning Models, we present Solver-Informed Reinforcement Learning (SIRL), a novel framework that significantly improves the authenticity of LLMs for optimization modeling using Reinforcement Learning with Verifiable Reward by leveraging external optimization solvers as verifiers. These verifiers automatically assess the executable code and the instance-level mathematical model represented by the associated LP file, yielding precise and comprehensive feedback signals -- including syntax, feasibility, and solution quality, serving as direct rewards for the RL process. This automated verification process, particularly from classic optimization solvers, also underpins our instance-enhanced self-consistency method to synthesize high-quality training data. Extensive experiments on diverse public benchmarks demonstrate that SIRL achieves state-of-the-art performance, substantially outperforming existing methods in generating accurate and executable optimization models. Our code is publicly available at https://github.com/Cardinal-Operations/SIRL.

cs.AI↗

arXiv2025

Two-Player Dynamic Potential LQ Games with Sequentially Revealed Costs

We investigate a novel finite-horizon linear-quadratic (LQ) feedback dynamic potential game with a priori unknown cost matrices played between two players. The cost matrices are revealed to the players sequentially, with the potential for future values to be previewed over a short time window. We propose an algorithm that enables the players to predict and track a feedback Nash equilibrium trajectory, and we measure the quality of their resulting decisions by introducing the concept of \emph{price of uncertainty}. We show that under the proposed algorithm, the price of uncertainty is bounded by horizon-invariant constants. The constants are the sum of three terms; the first and second terms decay exponentially as the preview window grows, and another depends on the magnitude of the differences between the cost matrices for each player. Through simulations, we illustrate that the resulting price of uncertainty initially decays at an exponential rate as the preview window lengthens, then remains constant for large time horizons.

math.OC↗

arXiv2024

LLM-Powered Ensemble Learning for Paper Source Tracing: A GPU-Free Approach

We participated in the KDD CUP 2024 paper source tracing competition and achieved the 3rd place. This competition tasked participants with identifying the reference sources (i.e., ref-sources, as referred to by the organizers of the competition) of given academic papers. Unlike most teams that addressed this challenge by fine-tuning pre-trained neural language models such as BERT or ChatGLM, our primary approach utilized closed-source large language models (LLMs). With recent advancements in LLM technology, closed-source LLMs have demonstrated the capability to tackle complex reasoning tasks in zero-shot or few-shot scenarios. Consequently, in the absence of GPUs, we employed closed-source LLMs to directly generate predicted reference sources from the provided papers. We further refined these predictions through ensemble learning. Notably, our method was the only one among the award-winning approaches that did not require the use of GPUs for model training. Code available at https://github.com/Cklwanfifa/KDDCUP2024-PST.

cs.LG↗

arXiv2024

Improved frequency spectra of gravitational waves with memory in a binary-black-hole simulation

Numerical relativists can now produce gravitational waveforms with memory effects routinely and accurately. The gravitational-wave memory effect contains very low-frequency components, including a persistent offset. The presence of these components violates basic assumptions about time-shift behavior underpinning standard data-analysis techniques in gravitational-wave astronomy. This poses a challenge to the analysis of waveform spectra: How to preserve the low-frequency characteristics when transforming a time-domain waveform to the frequency domain. To tackle this challenge, we revisit the preprocessing procedures applied to the waveforms that contain memory effects. We find inconsistency between the zero-frequency limit of displacement memory and the low- frequency spectrum of the same memory preprocessed using the common scheme in literature. To resolve the inconsistency, we propose a new robust preprocessing scheme that produces the spectra of memory waveforms more faithfully. Using this new scheme, we inspect several characteristics of the spectrum of a memory waveform. In particular, we find a discernible beating pattern formed by the dominant oscillatory mode and the displacement memory. This pattern is absent in the spectrum of a waveform without memory. The difference between the memory and no-memory waveforms is too small to be observed by current-generation detectors in a single binary-black-hole event. Detecting the memory in a single event is likely to occur in the era of next-generation detectors.

gr-qc↗

arXiv2024

Imprints of Changing Mass and Spin on Black Hole Ringdown

We numerically investigate the imprints of gravitational radiation-reaction driven changes to a black hole's mass and spin on the corresponding ringdown waveform. We do so by comparing the dynamics of a perturbed black hole evolved with the full (nonlinear) versus linearized Einstein equations. As expected, we find that the quasinormal mode amplitudes extracted from nonlinear evolution deviate from their linear counterparts at third order in initial perturbation amplitude. For perturbations leading to a change in the black hole mass and spin of $\sim 5\%$, which is reasonable for a remnant formed in an astrophysical merger, we find that nonlinear distortions to the complex amplitudes of some quasinormal modes can be as large as $\sim 50\%$ at the peak of the waveform. Furthermore, the change in the mass and spin results in a drift in the quasinormal mode frequencies, which for large amplitude perturbations causes the nonlinear waveform to rapidly dephase with respect to its linear counterpart. %These two nonlinear effects together create a large distortion in both the amplitude and phase of the ringdown gravitational waveform. Surprisingly, despite these nonlinear effects creating significant deviations in the nonlinear waveform, we show that a linear quasinormal mode model still performs quite well from close to the peak amplitude onwards.

gr-qc↗

arXiv2024

ScrollTimes: Tracing the Provenance of Paintings as a Window into History

The study of cultural artifact provenance, tracing ownership and preservation, holds significant importance in archaeology and art history. Modern technology has advanced this field, yet challenges persist, including recognizing evidence from diverse sources, integrating sociocultural context, and enhancing interactive automation for comprehensive provenance analysis. In collaboration with art historians, we examined the handscroll, a traditional Chinese painting form that provides a rich source of historical data and a unique opportunity to explore history through cultural artifacts. We present a three-tiered methodology encompassing artifact, contextual, and provenance levels, designed to create a "Biography" for handscroll. Our approach incorporates the application of image processing techniques and language models to extract, validate, and augment elements within handscroll using various cultural heritage databases. To facilitate efficient analysis of non-contiguous extracted elements, we have developed a distinctive layout. Additionally, we introduce ScrollTimes, a visual analysis system tailored to support the three-tiered analysis of handscroll, allowing art historians to interactively create biographies tailored to their interests. Validated through case studies and expert interviews, our approach offers a window into history, fostering a holistic understanding of handscroll provenance and historical significance.

cs.HC↗

arXiv2024

Nonlinear Effects In Black Hole Ringdown From Scattering Experiments I: spin and initial data dependence of quadratic mode coupling

We investigate quadratic quasinormal mode coupling in black hole spacetime through numerical simulations of single perturbed black holes using both numerical relativity and second-order black hole perturbation theory. Focusing on the dominant $\ell=|m|=2$ quadrupolar modes, we find good agreement (within $\sim10\%$) between these approaches, with discrepancies attributed to truncation error and uncertainties from mode fitting. Our results align with earlier studies extracting the coupling coefficients from select binary black hole merger simulations, showing consistency for the same remnant spins. Notably, the coupling coefficient is insensitive to a diverse range of initial data, including configurations that led to a significant (up to $5\%$) increase in the remnant black hole mass. These findings present opportunities for testing the nonlinear dynamics of general relativity with ground-based gravitational wave observatories. Lastly, we provide evidence of a bifurcation in coupling coefficients between counter-rotating and co-rotating quasinormal modes as black hole spin increases.

gr-qc↗

arXiv2023

Extending black-hole remnant surrogate models to extreme mass ratios

Numerical-relativity surrogate models for both black-hole merger waveforms and remnants have emerged as important tools in gravitational-wave astronomy. While producing very accurate predictions, their applicability is limited to the region of the parameter space where numerical-relativity simulations are available and computationally feasible. Notably, this excludes extreme mass ratios. We present a machine-learning approach to extend the validity of existing and future numerical-relativity surrogate models toward the test-particle limit, targeting in particular the mass and spin of post-merger black-hole remnants. Our model is trained on both numerical-relativity simulations at comparable masses and analytical predictions at extreme mass ratios. We extend the gaussian-process-regression model NRSur7dq4Remnant, validate its performance via cross validation, and test its accuracy against additional numerical-relativity runs. Our fit, which we dub NRSur7dq4EmriRemnant, reaches an accuracy that is comparable to or higher than that of existing remnant models while providing robust predictions for arbitrary mass ratios.

gr-qc↗

arXiv2023

Regret Analysis of Online LQR Control via Trajectory Prediction and Tracking: Extended Version

In this paper, we propose and analyze a new method for online linear quadratic regulator (LQR) control with a priori unknown time-varying cost matrices. The cost matrices are revealed sequentially with the potential for future values to be previewed over a short window. Our novel method involves using the available cost matrices to predict the optimal trajectory, and a tracking controller to drive the system towards it. We adopted the notion of dynamic regret to measure the performance of this proposed online LQR control method, with our main result being that the (dynamic) regret of our method is upper bounded by a constant. Moreover, the regret upper bound decays exponentially with the preview window length, and is extendable to systems with disturbances. We show in simulations that our proposed method offers improved performance compared to other previously proposed online LQR methods.

math.OC↗

arXiv2023

GP-NAS-ensemble: a model for NAS Performance Prediction

It is of great significance to estimate the performance of a given model architecture without training in the application of Neural Architecture Search (NAS) as it may take a lot of time to evaluate the performance of an architecture. In this paper, a novel NAS framework called GP-NAS-ensemble is proposed to predict the performance of a neural network architecture with a small training dataset. We make several improvements on the GP-NAS model to make it share the advantage of ensemble learning methods. Our method ranks second in the CVPR2022 second lightweight NAS challenge performance prediction track.

cs.LG↗