arXiv subjects

Yulun Wu

Publications and source records attributed to Yulun Wu.

At least 37 records · Page 2Linked to original sources

arXiv2026

Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning

Inspired by the remarkable reasoning capabilities of Deepseek-R1 in complex textual tasks, many works attempt to incentivize similar capabilities in Multimodal Large Language Models (MLLMs) by directly applying reinforcement learning (RL). However, they still struggle to activate complex reasoning. In this paper, rather than examining multimodal RL in isolation, we delve into current training pipelines and identify three crucial phenomena: 1) Effective cold start initialization is critical for enhancing MLLM reasoning. Intriguingly, we find that initializing with carefully selected text data alone can lead to performance surpassing many recent multimodal reasoning models, even before multimodal RL. 2) Standard GRPO applied to multimodal RL suffers from gradient stagnation, which degrades training stability and performance. 3) Subsequent text-only RL training, following the multimodal RL phase, further enhances multimodal reasoning. This staged training approach effectively balances perceptual grounding and cognitive reasoning development. By incorporating the above insights and addressing multimodal RL issues, we introduce ReVisual-R1, achieving a new state-of-the-art among open-source 7B MLLMs on challenging benchmarks including MathVerse, MathVision, WeMath, LogicVista, DynaMath, and challenging AIME2024 and AIME2025.

cs.LG↗

arXiv2026

PersonaLedger: Generating Realistic Financial Transactions with Persona Conditioned LLMs and Rule Grounded Feedback

Strict privacy regulations limit access to real transaction data, slowing open research in financial AI. Synthetic data can bridge this gap, but existing generators do not jointly achieve behavioral diversity and logical groundedness. Rule-driven simulators rely on hand-crafted workflows and shallow stochasticity, which miss the richness of human behavior. Learning-based generators such as GANs capture correlations yet often violate hard financial constraints and still require training on private data. We introduce PersonaLedger, a generation engine that uses a large language model conditioned on rich user personas to produce diverse transaction streams, coupled with an expert configurable programmatic engine that maintains correctness. The LLM and engine interact in a closed loop: after each event, the engine updates the user state, enforces financial rules, and returns a context aware "nextprompt" that guides the LLM toward feasible next actions. With this engine, we create a public dataset of 30 million transactions from 23,000 users and a benchmark suite with two tasks, illiquidity classification and identity theft segmentation. PersonaLedger offers a realistic, privacy preserving resource that supports rigorous evaluation of forecasting and anomaly detection models. PersonaLedger offers the community a rich, realistic, and privacy preserving resource -- complete with code, rules, and generation logs -- to accelerate innovation in financial AI and enable rigorous, reproducible evaluation.

cs.LG↗

arXiv2025

Evaluating Parameter Efficient Methods for RLVR

We systematically evaluate Parameter-Efficient Fine-Tuning (PEFT) methods under the paradigm of Reinforcement Learning with Verifiable Rewards (RLVR). RLVR incentivizes language models to enhance their reasoning capabilities through verifiable feedback; however, while methods like LoRA are commonly used, the optimal PEFT architecture for RLVR remains unidentified. In this work, we conduct the first comprehensive evaluation of over 12 PEFT methodologies across the DeepSeek-R1-Distill families on mathematical reasoning benchmarks. Our empirical results challenge the default adoption of standard LoRA with three main findings. First, we demonstrate that structural variants, such as DoRA, AdaLoRA, and MiSS, consistently outperform LoRA. Second, we uncover a spectral collapse phenomenon in SVD-informed initialization strategies (\textit{e.g.,} PiSSA, MiLoRA), attributing their failure to a fundamental misalignment between principal-component updates and RL optimization. Furthermore, our ablations reveal that extreme parameter reduction (\textit{e.g.,} VeRA, Rank-1) severely bottlenecks reasoning capacity. We further conduct ablation studies and scaling experiments to validate our findings. This work provides a definitive guide for advocating for more exploration for parameter-efficient RL methods.

cs.LG↗

arXiv2025

Inverse Design of Integrated Terahertz Vortex Beam Emitters with Staged-Annealing Topology Optimization

Integrated photonics is increasingly demanded in applications such as large-scale data centers, intelligent sensing, and next-generation wireless communications, where compact, multifunctional, and energy-efficient components are essential. Inverse-designed photonics, empowered by optimization and learning algorithms, have emerged as a powerful paradigm for realizing compact and multifunctional integrated photonic components. In this work, we develop a staged-annealing topological optimization (SATO) framework tailored for the design of integrated terahertz (THz) beam-shaping devices. Employing this inverse-designed framework, we experimentally demonstrate a class of compact THz vortex beam emitters on an all-silicon on-chip platform. These devices efficiently convert the in-plane fundamental transverse electric (TE) waveguide mode into free-space vortex beams with mode purity up to 87% and energy conversion efficiency up to 74% across the target wavelength range (680 μm to 720 μm). The inverse-designed emitters exhibit ultracompact footprints (lateral size < 4λ) and a free-standing configuration, enabling the generation of dual-directional vortex beams carrying opposite topological charges. The proposed SATO framework provides a generalizable and fabrication-compatible approach for THz photonic device engineering, offering a scalable pathway toward complex structured beam manipulation in next-generation wireless communication systems and on-chip integrated THz photonic systems.

physics.optics↗

arXiv2025

P1: Mastering Physics Olympiads with Reinforcement Learning

Recent progress in large language models (LLMs) has moved the frontier from puzzle-solving to science-grade reasoning-the kind needed to tackle problems whose answers must stand against nature, not merely fit a rubric. Physics is the sharpest test of this shift, which binds symbols to reality in a fundamental way, serving as the cornerstone of most modern technologies. In this work, we manage to advance physics research by developing large language models with exceptional physics reasoning capabilities, especially excel at solving Olympiad-level physics problems. We introduce P1, a family of open-source physics reasoning models trained entirely through reinforcement learning (RL). Among them, P1-235B-A22B is the first open-source model with Gold-medal performance at the latest International Physics Olympiad (IPhO 2025), and wins 12 gold medals out of 13 international/regional physics competitions in 2024/2025. P1-30B-A3B also surpasses almost all other open-source models on IPhO 2025, getting a silver medal. Further equipped with an agentic framework PhysicsMinions, P1-235B-A22B+PhysicsMinions achieves overall No.1 on IPhO 2025, and obtains the highest average score over the 13 physics competitions. Besides physics, P1 models also present great performance on other reasoning tasks like math and coding, showing the great generalibility of P1 series.

cs.LG↗

arXiv2025

AI Progress Should Be Measured by Capability-Per-Resource, Not Scale Alone: A Framework for Gradient-Guided Resource Allocation in LLMs

This position paper challenges the "scaling fundamentalism" dominating AI research, where unbounded growth in model size and computation has led to unsustainable environmental impacts and widening resource inequality. We argue that LLM development should be fundamentally reoriented toward capability-per-resource rather than capability alone. We present a theoretical framework demonstrating that resource-allocation decisions guided by gradient influence patterns can dramatically improve efficiency throughout the AI lifecycle. Our analysis shows that in transformer-based models, where a small fraction of parameters exert outsized influence (following heavy-tailed distributions), three critical insights emerge: (1) updating only high-influence parameters strictly outperforms full-parameter tuning on a performance-per-resource basis; (2) simple gradient norms provide computationally efficient proxies for identifying these high-influence components; and (3) coordinated parameter and data selection yields multiplicative efficiency gains, potentially reducing resource requirements by orders of magnitude. Building on these theoretical foundations, we propose a two stage paradigm marginal-return pretraining for foundation developers and influence guided adaptation for downstream users bridged by gradient blueprints, metadata describing which parameters matter most for various tasks. This capability-per-resource perspective transforms what were once considered pragmatic hardware workarounds into theoretically optimal strategies, democratizing access to cutting-edge AI capabilities while significantly reducing environmental impact. By embedding resource consciousness into how we develop, adapt, and evaluate models, we can reshape AI progress toward a more sustainable and equitable future.

cs.LG↗

arXiv2025

Provably Robust Adaptation for Language-Empowered Foundation Models

Language-empowered foundation models (LeFMs), such as CLIP and GraphCLIP, have transformed multimodal learning by aligning visual (or graph) features with textual representations, enabling powerful downstream capabilities like few-shot learning. However, the reliance on small, task-specific support datasets collected in open environments exposes these models to poisoning attacks, where adversaries manipulate the support samples to degrade performance. Existing defenses rely on empirical strategies, which lack formal guarantees and remain vulnerable to unseen and adaptive attacks. Certified robustness offers provable guarantees but has been largely unexplored for few-shot classifiers based on LeFMs. This study seeks to fill these critical gaps by proposing the first provably robust few-shot classifier that is tailored for LeFMs. We term our model Language-empowered Few-shot Certification (\textbf{LeFCert}). It integrates both textual and feature embeddings with an adaptive blending mechanism. To achieve provable robustness, we propose a twofold trimmed mean prototype and derive provable upper and lower bounds for classification scores, enabling certification under worst-case poisoning scenarios. To further enhance the performance, we extend LeFCert with two variants by considering a more realistic and tighter attack budget: LeFCert-L incorporates randomized smoothing to provide Lipschitz continuity and derive robustness under dual budget constraints, and LeFCert-C provides collective certification for scenarios where attackers distribute a shared poisoning budget across multiple samples. Experiments demonstrate that LeFCert achieves state-of-the-art performance, significantly improving both clean and certified accuracy compared to existing baselines. Despite its advanced robustness mechanisms, LeFCert is computationally efficient, making it practical for real-world applications.

cs.LG↗

arXiv2025

HiPhO: How Far Are (M)LLMs from Humans in the Latest High School Physics Olympiad Benchmark?

Recently, the physical capabilities of (M)LLMs have garnered increasing attention. However, existing benchmarks for physics suffer from two major gaps: they neither provide systematic and up-to-date coverage of real-world physics competitions such as physics Olympiads, nor enable direct performance comparison with humans. To bridge these gaps, we present HiPhO, the first benchmark dedicated to high school physics Olympiads with human-aligned evaluation. Specifically, HiPhO highlights three key innovations. (1) Comprehensive Data: It compiles 13 latest Olympiad exams from 2024-2025, spanning both international and regional competitions, and covering mixed modalities that encompass problems spanning text-only to diagram-based. (2) Professional Evaluation: We adopt official marking schemes to perform fine-grained grading at both the answer and step level, fully aligned with human examiners to ensure high-quality and domain-specific evaluation. (3) Comparison with Human Contestants: We assign gold, silver, and bronze medals to models based on official medal thresholds, thereby enabling direct comparison between (M)LLMs and human contestants. Our large-scale evaluation of 30 state-of-the-art (M)LLMs shows that: across 13 exams, open-source MLLMs mostly remain at or below the bronze level; open-source LLMs show promising progress with multiple golds; closed-source reasoning MLLMs can achieve 6 to 12 gold medals; and most models still have a significant gap from full marks. These results highlight the performance gap between open-source models and top students, the strong reasoning abilities of closed-source models, and the remaining room for improvement. HiPhO, a human-aligned Olympiad benchmark for multimodal physical reasoning, is open-source at https://github.com/SciYu/HiPhO with a public leaderboard at https://phyarena.github.io/.

cs.AI↗

arXiv2025

Trace Sampling 2.0: Code Knowledge Enhanced Span-level Sampling for Distributed Tracing

Distributed tracing is an essential diagnostic tool in microservice systems, but the sheer volume of traces places a significant burden on backend storage. A common approach to mitigating this issue is trace sampling, which selectively retains traces based on specific criteria, often preserving only anomalous ones. However, this method frequently discards valuable information, including normal traces that are essential for comparative analysis. To address this limitation, we introduce Trace Sampling 2.0, which operates at the span level while maintaining trace structure consistency. This approach allows for the retention of all traces while significantly reducing storage overhead. Based on this concept, we design and implement Autoscope, a span-level sampling method that leverages static analysis to extract execution logic, ensuring that critical spans are preserved without compromising structural integrity. We evaluated Autoscope on two open-source microservices. Our results show that it reduces trace size by 81.2% while maintaining 98.1% faulty span coverage, outperforming existing trace-level sampling methods. Furthermore, we demonstrate its effectiveness in root cause analysis, achieving an average improvement of 8.3%. These findings indicate that Autoscope can significantly enhance observability and storage efficiency in microservices, offering a robust solution for performance monitoring.

cs.SE↗

arXiv2025

HingeNet: A Harmonic-Aware Fine-Tuning Approach for Beat Tracking

Fine-tuning pre-trained foundation models has made significant progress in music information retrieval. However, applying these models to beat tracking tasks remains unexplored as the limited annotated data renders conventional fine-tuning methods ineffective. To address this challenge, we propose HingeNet, a novel and general parameter-efficient fine-tuning method specifically designed for beat tracking tasks. HingeNet is a lightweight and separable network, visually resembling a hinge, designed to tightly interface with pre-trained foundation models by using their intermediate feature representations as input. This unique architecture grants HingeNet broad generalizability, enabling effective integration with various pre-trained foundation models. Furthermore, considering the significance of harmonics in beat tracking, we introduce harmonic-aware mechanism during the fine-tuning process to better capture and emphasize the harmonic structures in musical signals. Experiments on benchmark datasets demonstrate that HingeNet achieves state-of-the-art performance in beat and downbeat tracking

cs.SD↗

arXiv2025

BeatFM: Improving Beat Tracking with Pre-trained Music Foundation Model

Beat tracking is a widely researched topic in music information retrieval. However, current beat tracking methods face challenges due to the scarcity of labeled data, which limits their ability to generalize across diverse musical styles and accurately capture complex rhythmic structures. To overcome these challenges, we propose a novel beat tracking paradigm BeatFM, which introduces a pre-trained music foundation model and leverages its rich semantic knowledge to improve beat tracking performance. Pre-training on diverse music datasets endows music foundation models with a robust understanding of music, thereby effectively addressing these challenges. To further adapt it for beat tracking, we design a plug-and-play multi-dimensional semantic aggregation module, which is composed of three parallel sub-modules, each focusing on semantic aggregation in the temporal, frequency, and channel domains, respectively. Extensive experiments demonstrate that our method achieves state-of-the-art performance in beat and downbeat tracking across multiple benchmark datasets.

cs.SD↗

arXiv2025

Inverse-Designed On-Chip Terahertz Three-Channel Mode and Wavelength Division Demultiplexer

High-performance multimode/multiwavelength (de)multiplexer is one of the most pivotal photonic devices for advanced on-chip interconnect systems. Traditional on-chip photonic (de)multiplexing requires large device footprint for maintaining high efficiency, large operation bandwidth, and small insertion losses. Here a hybrid inverse design method is therefore proposed to combine genetic algorithms (GA) and topology optimization for developing an ultracompact (lateral size< 2λ) terahertz (THz) mode-/wavelength-division demultiplexer (MDM-WDM). The method leverages the global search capability of GA in continuous parameter spaces and the local topology optimization strategy driven by the adjoint method, effectively improving design convergence efficiency and global performance robustness. Experimental results demonstrate that the device simultaneously achieves stable three-channel MDM and WDM with insertion loss (IL) of less than 3 dB and inter-channel crosstalk (CT) can be -22 dB. The output achieves spatial separation of the orthogonal TE10, TE20, and TE30 modes that correspond to three central wavelengths λ1~690 μm, λ2~700 μm, and λ3~710 μm, respectively, verifying the device's precise control over target modes and wavelengths. This work provides an efficient optimization approach for developing broadband, multichannel, and highly integrated THz multiplexing devices, offering new pathways for constructing next-generation integrated photonic interconnects and signal processing systems.

physics.optics↗

arXiv2025

Unsupervised Multi-channel Speech Dereverberation via Diffusion

We consider the problem of multi-channel single-speaker blind dereverberation, where multi-channel mixtures are used to recover the clean anechoic speech. To solve this problem, we propose USD-DPS, {U}nsupervised {S}peech {D}ereverberation via {D}iffusion {P}osterior {S}ampling. USD-DPS uses an unconditional clean speech diffusion model as a strong prior to solve the problem by posterior sampling. At each diffusion sampling step, we estimate all microphone channels' room impulse responses (RIRs), which are further used to enforce a multi-channel mixture consistency constraint for diffusion guidance. For multi-channel RIR estimation, we estimate reference-channel RIR by optimizing RIR parameters of a sub-band RIR signal model, with the Adam optimizer. We estimate non-reference channels' RIRs analytically using forward convolutive prediction (FCP). We found that this combination provides a good balance between sampling efficiency and RIR prior modeling, which shows superior performance among unsupervised dereverberation approaches. An audio demo page is provided in https://usddps.github.io/USDDPS_demo/.

cs.SD↗

arXiv2025

Label-free evaluation of lung and heart transplant biopsies using tissue autofluorescence-based virtual staining

Organ transplantation serves as the primary therapeutic strategy for end-stage organ failures. However, allograft rejection is a common complication of organ transplantation. Histological assessment is essential for the timely detection and diagnosis of transplant rejection and remains the gold standard. Nevertheless, the traditional histochemical staining process is time-consuming, costly, and labor-intensive. Here, we present a panel of virtual staining neural networks for lung and heart transplant biopsies, which digitally convert autofluorescence microscopic images of label-free tissue sections into their brightfield histologically stained counterparts, bypassing the traditional histochemical staining process. Specifically, we virtually generated Hematoxylin and Eosin (H&E), Masson's Trichrome (MT), and Elastic Verhoeff-Van Gieson (EVG) stains for label-free transplant lung tissue, along with H&E and MT stains for label-free transplant heart tissue. Subsequent blind evaluations conducted by three board-certified pathologists have confirmed that the virtual staining networks consistently produce high-quality histology images with high color uniformity, closely resembling their well-stained histochemical counterparts across various tissue features. The use of virtually stained images for the evaluation of transplant biopsies achieved comparable diagnostic outcomes to those obtained via traditional histochemical staining, with a concordance rate of 82.4% for lung samples and 91.7% for heart samples. Moreover, virtual staining models create multiple stains from the same autofluorescence input, eliminating structural mismatches observed between adjacent sections stained in the traditional workflow, while also saving tissue, expert time, and staining costs.

physics.med-ph↗

arXiv2025

Zero-shot Meta-learning for Tabular Prediction Tasks with Adversarially Pre-trained Transformer

We present an Adversarially Pre-trained Transformer (APT) that is able to perform zero-shot meta-learning on tabular prediction tasks without pre-training on any real-world dataset, extending on the recent development of Prior-Data Fitted Networks (PFNs) and TabPFN. Specifically, APT is pre-trained with adversarial synthetic data agents, who continue to shift their underlying data generating distribution and deliberately challenge the model with different synthetic datasets. In addition, we propose a mixture block architecture that is able to handle classification tasks with arbitrary number of classes, addressing the class size limitation -- a crucial weakness of prior deep tabular zero-shot learners. In experiments, we show that our framework matches state-of-the-art performance on small classification tasks without filtering on dataset characteristics such as number of classes and number of missing values, while maintaining an average runtime under one second. On common benchmark dataset suites in both classification and regression, we show that adversarial pre-training was able to enhance TabPFN's performance. In our analysis, we demonstrate that the adversarial synthetic data agents were able to generate a more diverse collection of data compared to the ordinary random generator in TabPFN. In addition, we demonstrate that our mixture block neural design has improved generalizability and greatly accelerated pre-training.

cs.LG↗

arXiv2025

Longitudinal Targeted Minimum Loss-based Estimation with Temporal-Difference Heterogeneous Transformer

We propose Deep Longitudinal Targeted Minimum Loss-based Estimation (Deep LTMLE), a novel approach to estimate the counterfactual mean of outcome under dynamic treatment policies in longitudinal problem settings. Our approach utilizes a transformer architecture with heterogeneous type embedding trained using temporal-difference learning. After obtaining an initial estimate using the transformer, following the targeted minimum loss-based likelihood estimation (TMLE) framework, we statistically corrected for the bias commonly associated with machine learning algorithms. Furthermore, our method also facilitates statistical inference by enabling the provision of 95% confidence intervals grounded in asymptotic statistical theory. Simulation results demonstrate our method's superior performance over existing approaches, particularly in complex, long time-horizon scenarios. It remains effective in small-sample, short-duration contexts, matching the performance of asymptotically efficient estimators. To demonstrate our method in practice, we applied our method to estimate counterfactual mean outcomes for standard versus intensive blood pressure management strategies in a real-world cardiovascular epidemiology cohort study.

stat.ML↗

arXiv2025

OSVAuto: automatic proofs about functional specifications in OS verification

We present OSVAuto for automatic proofs about functional specifications that commonly arise when verifying operating system kernels. The algorithm behind OSVAuto is designed to support natively those data types that commonly occur in OS verification, including sequences, maps, structures and enumerations. Propositions about these data are encoded into a form that is suitable for SMT solving. For quantifier instantiation, we propose an extension of recent work for automatic proofs about sequences. We evaluate the algorithm on proof obligations adapted from existing verification of the uC-OS/II kernel in Coq, demonstrating that a large number of proof obligations can be solved automatically, significantly reducing the proof effort on the functional side.

cs.SC↗

arXiv2025

LLMPrism: Black-box Performance Diagnosis for Production LLM Training Platforms

Large Language Models (LLMs) have brought about revolutionary changes in diverse fields, rendering LLM training of utmost importance for modern enterprises. To meet this demand, multi-tenant large-scale LLM training platforms have been built to offer LLM training services. Nevertheless, due to the complexity and synchronous nature of LLM training process, performance issues occur frequently and can result in substantial resource wastage. The limited visibility from the perspective of platform providers impedes existing profiling methods and poses challenges to the monitoring and diagnosis of the performance of LLM training jobs. For the first time, this paper proposes the utilization of underlying network flow data to reconstruct the training timelines of jobs based on the distinct characteristics in the LLM training procedure. We design LLMPrism, the first black-box performance diagnosis system for LLM training platforms. By progressively recognizing LLM training jobs, identifying their parallelism strategies, and reconstructing the training timelines, LLMPrism achieves non-intrusive, lightweight, and continuous monitoring of LLM training systems. Leveraging this monitoring capability, it further effectively diagnoses potential performance issues. Since Oct. 2024, LLMPrism has been deployed on our large-scale production Platform-X, in which the evaluations and deployment experiences demonstrate that LLMPrism can achieve accurate timeline reconstruction with an error within 0.3% and effectively diagnose various performance issues.

cs.SE↗