arXiv subjects

Jianhao Zhang

Publications and source records attributed to Jianhao Zhang.

At least 19 recordsLinked to original sources

arXiv2026

Multi-Agent Closed-Loop Reasoning for Organic Structure Elucidation from Multimodal Spectra

Following the molecular discovery and synthesis revolutions, scalable automated structure elucidation from routine spectroscopic data remains an outstanding challenge. Despite decades of computational efforts, no existing system achieved reliable reasoning over unseen spectra. Here, we propose MACROS, a multi-agent system automating structure elucidation by emulating expert iterative hypothesis-testing. Trained on 100M simulated and 1.6M experimental spectra-molecule pairs, it natively supports arbitrary combinations of routine spectroscopic techniques. It achieves unprecedented zero-shot generalization to diverse real-world samples, correctly identifying synthetic compounds, natural products and metabolites above 500 Da with 1D NMR. Remarkably, MACROS spontaneously recovers textbook spectroscopic correlations from unassigned data and exhibits emergent chemical intuition such as a ring-first parsing preference, learning fundamental chemical principles rather than memorizing database patterns. MACROS augments chemists via collaboration to deliver sixfold faster, 40% more accurate elucidation. MACROS establishes a scalable foundation for fully automated structure elucidation, and catalyzes accelerated molecular discovery toward autonomous laboratories.

physics.chem-ph↗

arXiv2026

Two to Tango: Coupled Task-Reference Selection for Safe LLM Fine-tuning

Fine-tuning safety aligned large language models (LLMs) on downstream data improves adaptation but may erode learned safety behavior. Existing methods use fixed safety examples, global constraints, or one-sided task filtering. Our diagnostics show task updates expose different safety constraints, motivating joint selection of relevant references and compatible task samples. We propose DualSelect, a coupled framework for task and reference selection that refreshes task conditioned safety references before filtering whole task samples compatible with the induced reference direction. Under a minimax view, DualSelect selects safety references with high preservation loss and task conflict, together with compatible task samples, through entropy-regularized scoring surrogates, lazy reference refresh, and gradient correction. On 1B-8B LLMs, DualSelect preserves safety without losing task utility; using the REDORCA judge, it improves Safety Avg. over the strongest baseline by at least 5.10 points and remains highest in Safety Avg. across judges with moderate overhead. This view extends to retention focused continual learning.

cs.LG↗

arXiv2026

Boundary-dominated optomechanics in silicon metamaterial membranes

Stimulated Brillouin scattering in integrated photonic waveguides enables coherent coupling between optical photons and gigahertz acoustic phonons, providing a powerful mechanism for on-chip microwave photonics and opto-acoustic signal processing. Despite theoretical predictions of ultra-strong Brillouin interactions arising from enhanced light-sound coupling at device boundaries, most state-of-the-art integrated demonstrations remain governed by bulk photoelastic effects. This limitation stems from trade-offs between optical loss, interaction with waveguide boundaries and accessible phonon frequencies associated with the use of transverse-electric optical modes coupled to horizontally breathing mechanical modes. Here we demonstrate a new approach based on transverse-magnetic optical modes coupled to vertically breathing mechanical modes in suspended silicon membranes engineered with subwavelength metamaterial claddings. In this geometry, the interaction is dominated by the moving-boundary effect occurring at smooth top and bottom interfaces, while the phonon frequency is set primarily by the membrane thickness rather than its width. We observe forward Brillouin interactions at a record frequency of 12 GHz with a gain of 7200 W$^{-1}$ m$^{-1}$ and a mechanical quality factor of 620, yielding the highest Brillouin gain-to-quality-factor ratio reported in silicon waveguides. The devices exhibit net Brillouin amplification in millimeter-scale waveguides with pump powers below 15 mW, establishing a scalable platform for high-frequency integrated opto-acoustic signal processing.

physics.optics↗

arXiv2026

Directional and contra-directional coupling in Huygens' metawaveguide microring resonators

Huygens' metawaveguides represent a transformative concept in photonic device engineering, enabling unprecedented control over light propagation. This study presents, for the first time, integrated Huygens'-based microring resonators and directional and contra-directional couplers, specifically designed for operation at the S- and C-band telecommunication wavelengths. By leveraging the unique properties of resonant Huygens' waveguides, we demonstrate efficient evanescent directional coupling with high-Q resonators, characterized by negative group index and near-zero dispersion, which are critical for enhancing performance in compact, high-performance add-drop filters. The research further explores the implications of these novel structures on group index and group velocity dispersion, providing insights into their potential applications in nonlinear optics and quantum information technologies. Notably, the introduction of a hybrid subwavelength grating-Huygens' contra-directional coupler facilitates backward coupling between resonant and non-resonant metawaveguides, achieving a broad spectral rejection bandwidth. Our findings advance the integration of resonant metamaterials into scalable photonic platforms, laying the groundwork for innovative applications in optical communications, quantum photonics and sensing systems.

physics.optics↗

arXiv2025

Stimulated Forward Brillouin Scattering in Subwavelength Silicon Membranes

Brillouin scattering enables efficient and coherent conversion between optical photons and gigahertz-frequency phonons. Integrated circuits that harness this nonlinear interaction have immense potential for signal processing, quantum transduction, and sensing applications. However, achieving strong overlap and tight confinement of optical and mechanical modes in silicon nanophotonic waveguides remains a significant challenge. Here, we propose and demonstrate a novel strategy that enables independent control of optical and mechanical modes in periodically segmented silicon waveguides. Our approach combines two distinct periodic lattices: one with a period shorter than half of the optical wavelength, providing light guiding by metamaterial-induced index contrast, and another that creates a complete phononic bandgap confining acoustic modes. This dual-lattice strategy opens new degrees of freedom to optimize optomechanical confinement and coupling simultaneously. Based on this approach, we experimentally demonstrate remarkably high Brillouin gain of $G_\mathrm{B}=2673$ W$^{-1}$m$^{-1}$, resulting in a Stokes gain of 3 dB and an anti-Stokes loss of 4 dB with 6.4 MHz mechanical linewidth. These results illustrate the potential of subwavelength silicon metamaterials for engineering on-chip optomechanical interactions.

physics.optics↗

arXiv2025

Skywork-R1V3 Technical Report

We introduce Skywork-R1V3, an advanced, open-source vision-language model (VLM) that pioneers a new approach to visual reasoning. Its key innovation lies in effectively transferring reasoning skills from text-only Large Language Models (LLMs) to visual tasks. The strong performance of Skywork-R1V3 primarily stems from our elaborate post-training RL framework, which effectively activates and enhances the model's reasoning ability, without the need for additional continue pre-training. Through this framework, we further uncover the fundamental role of the connector module in achieving robust cross-modal alignment for multimodal reasoning models. In addition, we introduce a unique indicator of reasoning capability, the entropy of critical reasoning tokens, which has proven highly effective for checkpoint selection during RL training. Skywork-R1V3 achieves state-of-the-art results on MMMU, significantly improving from 64.3% to 76.0%. This performance matches entry-level human capabilities. Remarkably, our RL-powered post-training approach enables even the 38B parameter model to rival top closed-source VLMs. The implementation successfully transfers mathematical reasoning to other subject-related reasoning tasks. We also include an analysis of curriculum learning and reinforcement finetuning strategies, along with a broader discussion on multimodal reasoning. Skywork-R1V3 represents a significant leap in multimodal reasoning, showcasing RL as a powerful engine for advancing open-source VLM capabilities.

cs.CL↗

arXiv2025

Skywork R1V2: Multimodal Hybrid Reinforcement Learning for Reasoning

We present Skywork R1V2, a next-generation multimodal reasoning model and a major leap forward from its predecessor, Skywork R1V. At its core, R1V2 introduces a hybrid reinforcement learning paradigm that jointly leverages the Mixed Preference Optimization (MPO) and the Group Relative Policy Optimization (GRPO), which harmonizes reward-model guidance with rule-based strategies, thereby addressing the long-standing challenge of balancing sophisticated reasoning capabilities with broad generalization. To further enhance training efficiency, we propose the Selective Sample Buffer (SSB) mechanism, which effectively addresses the vanishing advantages dilemma inherent in GRPO by prioritizing high-value samples throughout the optimization process. Notably, we observe that excessive reinforcement signals can induce visual hallucinations--a phenomenon we systematically monitor and mitigate through calibrated reward thresholds throughout the training process. Empirical results affirm the exceptional capability of R1V2, with benchmark-leading performances such as 62.6 on OlympiadBench, 78.9 on AIME2024, 63.6 on LiveCodeBench, and 73.6 on MMMU. These results underscore R1V2's superiority over existing open-source models and demonstrate significant progress in closing the performance gap with premier proprietary systems, including Gemini 2.5 and OpenAI-o4-mini. The Skywork R1V2 model weights have been publicly released to promote openness and reproducibility https://huggingface.co/Skywork/Skywork-R1V2-38B.

cs.CV↗

arXiv2024

Perfectly vertical silicon metamaterial grating couplers with large segmentation periods up to 650 nm

Perfectly vertical grating couplers leveraging metamaterials can achieve both high coupling efficiency and minimal back reflection. The fabricability of these designs, with segmentations in both the longitudinal and transverse dimensions, hinges on the minimum feature size offered by cutting-edge fabrication technologies. In this work we present both numerical and experimental evidence that high performance devices can be obtained while using large transverse segmentation periods of up to 650 nm, thereby increasing the critical feature sizes. For single-step etched couplers produced on the 220 nm silicon-on-insulator platform, we demonstrate coupling efficiencies of nearly 50% in the C-band and remarkably low back reflections of -22 dB at zero-degree incidence angle. Notably, the duty cycles used in our optimized designs deviate significantly from those predicted by traditional effective medium models, even for small periods. Our findings promise to expand the range of optical properties achievable in metamaterials and offer fresh insights into the fine-tuning of nanophotonic devices.

physics.optics↗

arXiv2024

Skywork-MoE: A Deep Dive into Training Techniques for Mixture-of-Experts Language Models

In this technical report, we introduce the training methodologies implemented in the development of Skywork-MoE, a high-performance mixture-of-experts (MoE) large language model (LLM) with 146 billion parameters and 16 experts. It is initialized from the pre-existing dense checkpoints of our Skywork-13B model. We explore the comparative effectiveness of upcycling versus training from scratch initializations. Our findings suggest that the choice between these two approaches should consider both the performance of the existing dense checkpoints and the MoE training budget. We highlight two innovative techniques: gating logit normalization, which improves expert diversification, and adaptive auxiliary loss coefficients, allowing for layer-specific adjustment of auxiliary loss coefficients. Our experimental results validate the effectiveness of these methods. Leveraging these techniques and insights, we trained our upcycled Skywork-MoE on a condensed subset of our SkyPile corpus. The evaluation results demonstrate that our model delivers strong performance across a wide range of benchmarks.

cs.CL↗

arXiv2023

Coop: Memory is not a Commodity

Tensor rematerialization allows the training of deep neural networks (DNNs) under limited memory budgets by checkpointing the models and recomputing the evicted tensors as needed. However, the existing tensor rematerialization techniques overlook the memory system in deep learning frameworks and implicitly assume that free memory blocks at different addresses are identical. Under this flawed assumption, discontiguous tensors are evicted, among which some are not used to allocate the new tensor. This leads to severe memory fragmentation and increases the cost of potential rematerializations. To address this issue, we propose to evict tensors within a sliding window to ensure all evictions are contiguous and are immediately used. Furthermore, we proposed cheap tensor partitioning and recomputable in-place to further reduce the rematerialization cost by optimizing the tensor allocation. We named our method Coop as it is a co-optimization of tensor allocation and tensor rematerialization. We evaluated Coop on eight representative DNNs. The experimental results demonstrate that Coop achieves up to $2\times$ memory saving and hugely reduces compute overhead, search latency, and memory fragmentation compared to the state-of-the-art baselines.

cs.LG↗

arXiv2023

Skywork: A More Open Bilingual Foundation Model

In this technical report, we present Skywork-13B, a family of large language models (LLMs) trained on a corpus of over 3.2 trillion tokens drawn from both English and Chinese texts. This bilingual foundation model is the most extensively trained and openly published LLMs of comparable size to date. We introduce a two-stage training methodology using a segmented corpus, targeting general purpose training and then domain-specific enhancement training, respectively. We show that our model not only excels on popular benchmarks, but also achieves \emph{state of the art} performance in Chinese language modeling on diverse domains. Furthermore, we propose a novel leakage detection method, demonstrating that test data contamination is a pressing issue warranting further investigation by the LLM community. To spur future research, we release Skywork-13B along with checkpoints obtained during intermediate stages of the training process. We are also releasing part of our SkyPile corpus, a collection of over 150 billion tokens of web text, which is the largest high quality open Chinese pre-training corpus to date. We hope Skywork-13B and our open corpus will serve as a valuable open-source resource to democratize access to high-quality LLMs.

cs.CL↗

arXiv2023

KOIOS: Top-k Semantic Overlap Set Search

We study the top-k set similarity search problem using semantic overlap. While vanilla overlap requires exact matches between set elements, semantic overlap allows elements that are syntactically different but semantically related to increase the overlap. The semantic overlap is the maximum matching score of a bipartite graph, where an edge weight between two set elements is defined by a user-defined similarity function, e.g., cosine similarity between embeddings. Common techniques like token indexes fail for semantic search since similar elements may be unrelated at the character level. Further, verifying candidates is expensive (cubic versus linear for syntactic overlap), calling for highly selective filters. We propose KOIOS, the first exact and efficient algorithm for semantic overlap search. KOIOS leverages sophisticated filters to minimize the number of required graph-matching calculations. Our experiments show that for medium to large sets less than 5% of the candidate sets need verification, and more than half of those sets are further pruned without requiring the expensive graph matching. We show the efficiency of our algorithm on four real datasets and demonstrate the improved result quality of semantic over vanilla set similarity search.

cs.DB↗

arXiv2023

Genetic optimization of Brillouin scattering gain in subwavelength-structured silicon membrane waveguides

On-chip Brillouin optomechanics has great potential for applications in communications, sensing, and quantum technologies. Tight confinement of near-infrared photons and gigahertz phonons in integrated waveguides remains a key challenge to achieving strong on-chip Brillouin gain. Here, we propose a new strategy to harness Brillouin gain in silicon waveguides, based on the combination of genetic algorithm optimization and periodic subwavelength structuration to engineer photonic and phononic modes simultaneously. The proposed geometry is composed of a waveguide core and a lattice of anchoring arms with a subwavelength period requiring a single etch step. The waveguide geometry is optimized to maximize the Brillouin gain using a multi-physics genetic algorithm. Our simulation results predict a remarkable Brillouin gain exceeding 3300 1/(W m), for a mechanical frequency near 15 GHz.

physics.optics↗

arXiv2022

Broadband behavior of quadratic metalenses with a wide field of view

Metalenses are attracting a large interest for the implementation of complex optical functionalities in planar and compact devices. However, chromatic and off-axis aberrations remain standing challenges. Here, we experimentally investigate the broadband behavior of metalenses based on quadratic phase profiles. We show that these metalenses do not only guarantee an arbitrarily large field of view but are also inherently tolerant to longitudinal and transverse chromatic aberrations. As such, we demonstrate a single-layer, silicon metalens with a field of view of 86° and a bandwidth up to 140 nm operating at both 1300 nm and 1550 nm telecommunication wavelength bands.

physics.optics↗

arXiv2021

Sparse Logistic Tensor Decomposition for Binary Data

Tensor data are increasingly available in many application domains. We develop several tensor decomposition methods for binary tensor data. Different from classical tensor decompositions for continuous-valued data with squared error loss, we formulate logistic tensor decompositions for binary data with a Bernoulli likelihood. To enhance the interpretability of estimated factors and improve their stability further, we propose sparse formulations of logistic tensor decomposition by considering $\ell_{1}$-norm and $\ell_{0}$-norm regularized likelihood. To handle the resulting optimization problems, we develop computational algorithms which combine the strengths of tensor power method and majorization-minimization (MM) algorithm. Through simulation studies, we demonstrate the utility of our methods in analysis of binary tensor data. To illustrate the effectiveness of the proposed methods, we analyze a dataset concerning nations and their political relations and perform co-clustering of estimated factors to find associations between the nations and political relations.

stat.AP↗

arXiv2021

Generating 10-GHz phonons in nanostructured silicon membrane optomechanical cavity

Flexible control of photons and phonons in silicon nanophotonic waveguides is a key feature for emerging applications in communications, sensing and quantum technologies. Strong phonon leakage towards the silica under-cladding hampers optomechanical interactions in silicon-on-insulator. This limitation has been circumvented by totally or partially removing the silica under-cladding to form pedestal or silicon membrane waveguides. Remarkable optomechanical interactions have been demonstrated in silicon using pedestal strips, membrane ribs, and photonic/phononic crystal membrane waveguides. Still, the mechanical frequencies are limited to the 1-5 GHz range. Here, we exploit the periodic nanostructuration in Si membrane gratings to shape GHz phononic modes and near-infrared photonic modes, achieving ultrahigh mechanical frequency (10 GHz) and strong photon-phonon overlap (61.5%) simultaneously. Based on this concept, we experimentally demonstrate a one-dimension optomechanical micro-resonator with a high mechanical frequency of 10 GHz and a quality factor of 1000. These results were obtained at room temperature and ambient conditions with an intracavity optical power below 1 mW, illustrating the efficient optical driving of the mechanical mode enabled by the proposed approach.

physics.optics↗

arXiv2021

Silicon-on-insulator optomechanical microresonator with tight photon and phonon confinement

The implementation of optomechanical devices in silicon-on-insulator (SOI), the canonical silicon photonics technology is seriously hampered by the strong phonon leakage into the silica under-cladding. This limitation has been partially circumvented by total or partial removal of the silica under-cladding to form Si membranes or pedestal waveguides. However, this approach complicates integration with standard silicon optoelectronics circuitry, limiting the versatility and application of the strategy. Here, we propose and demonstrate a new strategy to confine photons and phonons in SOI without removing the silica under-cladding. Inspired by end-fire antenna arrays, we implement a periodic nanostructuration of silicon that simultaneously enables cancelling phonon leakage by destructive interference and guiding of photons by metamaterial index confinement. Based on this concept, we implement SOI optomechanical micro-resonators yielding remarkable optomechanical coupling (go=49 kHz) between 0.66 GHz mechanical modes and near-infrared optical modes. The mechanical mode exhibits a measured quality factor of Qm = 730, the largest reported for SOI optomechanical resonators, without silica removal. This value compares favorably with state-of-the-art Si membrane waveguides recently used to demonstrate remarkable Brillouin interactions in silicon (Qm ~ 700). These results open a new path for developing optomechanics in SOI without the need for silica removal, allowing seamless co-integration with current Si optoelectronics circuits, with a great potential for applications in communications, sensing, metrology, and quantum technologies.

physics.optics↗

arXiv2021

An electro-optically tunable microring laser monolithically integrated on lithium niobate on insulator

We demonstrate monolithic integration of an electro-optically (EO) tunable microring laser on lithium niobate on insulator (LNOI) platform. The device is fabricated by photolithography assisted chemo-mechanical etching (PLACE), and the pump laser is evanescently coupled into the erbium (Er3+) doped LN microring laser using an undoped LN waveguide mounted above the microring. The quality factor of the LN microring resonator is measured as high as 1.54x10^5 at the wavelength of 1542 nm. Lasing action can be observed at a pump power threshold below 3.5 mW using a 980 nm continuous-wave pump laser. Finally, tuning of the laser wavelength is achieved by varying the electric voltage on the microelectrodes fabricated in the vicinity of microring waveguide, showing an EO coefficient of 0.33 pm/V.

physics.optics↗