arXiv subjects

Haoxiang Jiang

Publications and source records attributed to Haoxiang Jiang.

8 recordsLinked to original sources

arXiv2026

RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains

Pointwise reward modeling offers critical signals for LLM post-training, yet struggles with absolute scoring in subjective, non-verifiable settings. Rubric-based methods address this by decomposing evaluation into explicit criteria, but existing approaches typically depend on frontier LLMs and suffer from ties caused by hard Boolean aggregation. We present RUBRIC-ARROW, an alternating framework that jointly trains a rubric generator and a rubric-conditioned judge, with its RL stage using only pairwise preference data. Our method couples a probability-based scoring rule that reduces ties with phase-specific preference-based rewards and an alternating GRPO scheme that together train the pointwise evaluator. Extensive experiments show that RUBRIC-ARROW achieves competitive reward-modeling accuracy and yields consistent gains for downstream policy post-training.

cs.LG↗

arXiv2025

RoseRAG: Robust Retrieval-augmented Generation with Small-scale LLMs via Margin-aware Preference Optimization

Large language models (LLMs) have achieved impressive performance but face high computational costs and latency, limiting their deployment in resource-constrained settings. In contrast, small-scale LLMs (SLMs) are more efficient yet struggle to capture evolving real-world knowledge. Retrieval-augmented generation (RAG) helps by integrating external knowledge, but imperfect retrieval can introduce distracting noise that misleads SLMs. We propose RoseRAG, a robust RAG framework for SLMs via Margin-aware Preference Optimization. RoseRAG employs multi-turn prompting for detailed reasoning, rejection sampling for high-quality explanations, and contrastive preference selection to refine responses by maximizing the likelihood gap between preferred and non-preferred outputs. By integrating these components into a margin-aware optimization process, RoseRAG robustly enhances the accuracy and reliability of SLMs for RAG applications. Extensive experiments on three open-domain question answering benchmarks indicate that our innovative RoseRAG surpasses state-of-the-art baselines significantly.

cs.CL↗

arXiv2024

PEANuT: Parameter-Efficient Adaptation with Weight-aware Neural Tweakers

Fine-tuning large pre-trained foundation models often yields excellent downstream performance but is prohibitively expensive when updating all parameters. Parameter-efficient fine-tuning (PEFT) methods such as LoRA alleviate this by introducing lightweight update modules, yet they commonly rely on weight-agnostic linear approximations, limiting their expressiveness. In this work, we propose PEANuT, a novel PEFT framework that introduces weight-aware neural tweakers, compact neural modules that generate task-adaptive updates conditioned on frozen pre-trained weights. PEANuT provides a flexible yet efficient way to capture complex update patterns without full model tuning. We theoretically show that PEANuT achieves equivalent or greater expressivity than existing linear PEFT methods with comparable or fewer parameters. Extensive experiments across four benchmarks with over twenty datasets demonstrate that PEANuT consistently outperforms strong baselines in both NLP and vision tasks, while maintaining low computational overhead.

cs.LG↗

arXiv2023

Enhanced coherent light-matter interaction and room-temperature quantum yield of plasmonic resonances engineered by a chiral exceptional point

Strong dissipation of plasmonic resonances is detrimental to quantum manipulation. To enhance the quantum coherence, we propose to tailor the local density of states (LDOS) of plasmonic resonances by integrating with a photonic cavity operating at a chiral exceptional point (CEP), where the phase of light field can offer a new degree of freedom to flexibly manipulate the quantum states. A quantized few-mode theory is employed to reveal that the LDOS of the proposed hybrid cavity can evolve into sub-Lorentzian lineshape, with order-of-magnitude linewidth narrowing and additionally a maximum of eightfold enhancement compared to the usual plasmonic-photonic cavity without CEP. This results in the enhanced coherent light-matter interaction accompanied by the reduced dissipation of polaritonic states. Furthermore, a scattering theory based on eigenmode decomposition is present to elucidate two mechanisms responsible for the significant improvement of quantum yield at CEP, the reduction of plasmonic absorption by the Fano interference and the enhancement of cavity radiation through the superscattering. Importantly, we find the latter allows achieving a near-unity quantum yield at room temperature; in return, high quantum yield is beneficial to experimentally verify the enhanced LDOS at CEP by measuring the fluorescence lifetime of a quantum emitter. Therefore, our work demonstrates that the plasmonic resonances in CEP-engineered environment can serve as a promising platform for exploring the quantum states control by virtue of the non-Hermiticity of open optical resonators and building the high-performance quantum devices for sensing, spectroscopy, quantum information processing and quantum computing.

quant-ph↗

arXiv2023

Topologically protected subradiant cavity polaritons through linewidth narrowing enabled by dissipationless edge states

Cavity polaritons derived from the strong light-matter interaction at the quantum level provide a basis for efficient manipulation of quantum states via cavity field. Polaritons with narrow linewidth and long lifetime are appealing in applications such as quantum sensing and storage. Here, we propose a prototypical arrangement to implement a whispering-gallery-mode resonator with topological mirror moulded by one-dimensional atom array, which allows to boost the lifetime of cavity polaritons over an order of magnitude. This considerable enhancement attributes to the coupling of polaritonic states to dissipationless edge states protected by the topological bandgap of atom array that suppresses the leakage of cavity modes. When exceeding the width of Rabi splitting, topological bandgap can further reduce the dissipation from polaritonic states to bulk states of atom array, giving arise to subradiant cavity polaritons with extremely sharp linewidth. The resultant Rabi oscillation decays with a rate even below the free-space decay of a single quantum emitter. Inheriting from the topologically protected properties of edge states, the subradiance of cavity polaritons can be preserved in the disordered atom mirror with moderate perturbations involving the atomic frequency, interaction strengths and location. Our work opens up a new paradigm of topology-engineered quantum states with robust quantum coherence for future applications in quantum computing and network.

quant-ph↗

arXiv2022

Radii of spherical photon orbits around Kerr-Newman black holes

The spherical photon orbits around a black hole with constant radii are particular important in astrophysical observations of the black hole. In this paper, the equatorial and non-equatorial spherical photon orbits around Kerr-Newman black holes are studied. The radii of these orbits satisfy a sextic polynomial equation with three parameters, the rotation parameter $u$, charge parameter $w$ and effective inclination angle $v$. It is found that there are two polar orbits and one is inside and the other is outside the event horizon. For orbits in the equatorial plane around an extremal Kerr-Newman black hole , we find three positive solutions to the equation and provide analytical formulas for the radii of the three orbits. It is also found that a critical value of rotation parameter $u=\frac{1}{4}$ exists, above which only one retrograde orbit exists outside the event horizon and below which two orbits (one prograde and one retrograde) exist outside the event horizon. For orbits in the equatorial plane of a nonextremal Kerr-Newman black hole, there are four or two positive solutions. A critical curve in $(u,w)$-plane which separates the regions is also found. On this critical curve, the explicit formulas of three solutions are found. There always exist two equatorial photon orbits outside the event horizon in this case. For orbits between the above two special planes, there are four or two solutions to the general sextic equation. When the black hole is extremal, it is found that there is a critical inclination angle $v_{cr}$, below which only one orbit exists and above which two orbits exist outside the event horizon for a rapidly rotating black hole $u>\frac{1}{4}$. At this critical inclination angle, exact formula for the radii is also derived. Finally, a critical surface in parameter space of ($u,w,v$) is shown that separates regions with four or two solutions.

gr-qc↗

arXiv2019

Preferred States of Open Electronic Systems

System-environment interaction may introduce dynamic destruction of quantum coherence, resulting in a special representation named as pointer states. Here, pointer states of an open electronic system are studied. The decoherence effect is taken into account through two different models which are Büttiker's virtual probe model and electron-phonon interaction in the polaron picture. The pointer states of the system with different coupling strength are investigated. The pointer states are identified by tracking the eigenstates of the density matrix in real-time evolution. It is found that the pointer states can emerge for arbitrary coupling strength. And the pointer states deform to the eigenstates of the system in the strong coupling limit, which indicates the vanish of quantumness in the strong coupling limit.

cond-mat.mes-hall↗

arXiv2012

Ab-initio Mapping of Projected Local Density of States in Arbitrary Nanostructures: Application to Photonic Crystal Slabs and Cavities

Based upon projected local density of states (PLDOS) for photons, we develop a local coupling theory to simultaneously treat the weak and strong interaction between a quantum emitter and photons in arbitrary nanostructures. The PLDOS is mapped by an extremely flexible and efficient method. The recent experiment observation for the photonic crystal slabs is very well interpreted by our ab-initio PLDOS. More importantly, a bridge linking the PLDOS and cavity quantum electrodynamics is for the first time established to settle quality factor, g factor and vacuum Rabi splitting. Our work greatly enriches the knowledge about the interaction between light and matter in nanostructures.

physics.optics↗