arXiv subjects

Qian Qian

Publications and source records attributed to Qian Qian.

14 recordsLinked to original sources

arXiv2026

MedFabric: Gold Evidence Hides the Difficulty of Word-Level Medical Fabrication Detection

Large language models fabricate in medicine, producing fluent statements that are factually wrong, so reliable fabrication detection is a prerequisite for clinical deployment. Reported progress on this task is inflated by two evaluation artifacts: an authorship-style shortcut, where human-written ground truths are paired with LLM-written hallucinations so detectors key on writing style rather than facts, and the provision of gold evidence at test time. A benchmark that tests factual reasoning must therefore remove the style shortcut, ground every fabrication in a real retrievable passage, and score detectors across the range of evidence quality faced in deployment. We build MedFabric to these requirements, a benchmark of 646 word-level medical fabrications, each paired with a ground truth that shares its LLM authorship and near-identical surface form (median ROUGE-L 0.95). On MedFabric the task is unsolved: expert clinicians reach only 53.3% macro F1 and no detector family clears about 60% without gold evidence. Our central finding is that detection is governed by evidence correctness rather than fabrication subtlety, since a strong LLM scores 91% with the gold passage but falls to 35%, below its own no-evidence baseline, under a wrong one, a pattern that holds on two benchmarks at two model scales. The failure is actionable: a retrieval-confidence gate that abstains on low-confidence evidence raises macro F1 from 61% to 74%, and we release MedFabric, all code, and every baseline.

cs.CL↗

arXiv2026

Pair-loaded electron-only magnetic reconnection using laser-driven capacitor coils

We propose and simulate a laboratory platform to study the effects of positrons in magnetic reconnection using laser-driven capacitor coils. Using particle-in-cell simulations, we show that externally injected MeV electron-positron pairs are trapped in the coil current sheet, significantly modifying the reconnection dynamics and particle acceleration. These pairs increase the reconnection rate by a factor of approximately 8, which Ohm's law decomposition reveals to be driven by the divergence of the generalized pressure tensor. Based on their high energy and magnetization, the pairs also substantially broaden the diffusion region. Particle tracking simulations in realistic coil magnetic fields further demonstrate that injected pairs can remain confined for several picoseconds, providing conditions for sustained interaction with the reconnection region. These results establish a near-term pathway to laboratory studies of positron-influenced reconnection, bridging high-energy-density experiments with pair-dominated astrophysical environments.

physics.plasm-ph↗

arXiv2026

Fast and Effective On-policy Distillation from Reasoning Prefixes

On-policy distillation (OPD), which samples trajectories from the student model and supervises them with a teacher at the token level, avoids relying solely on verifiable terminal rewards and can yield better generalization than off-policy distillation. However, OPD requires expensive on-the-fly sampling of the student policy during training, which substantially increases training cost, especially for long responses. Our initial analysis shows that, during OPD, training signals are often concentrated in the prefix of each output, and that even a short teacher-generated prefix can significantly help the student produce the correct answer. Motivated by these observations, we propose a simple yet effective modification of OPD: we apply the distillation objective only to prefixes of student-generated outputs and terminate each sampling early during distillation. Experiments on a suite of AI-for-Math and out-of-domain benchmarks show that on-policy prefix distillation matches the performance of full OPD while reducing training FLOP by 2x-47x.

cs.LG↗

arXiv2026

Health-SCORE: Towards Scalable Rubrics for Improving Health-LLMs

Rubrics are essential for evaluating open-ended LLM responses, especially in safety-critical domains such as healthcare. However, creating high-quality and domain-specific rubrics typically requires significant human expertise time and development cost, making rubric-based evaluation and training difficult to scale. In this work, we introduce Health-SCORE, a generalizable and scalable rubric-based training and evaluation framework that substantially reduces rubric development costs without sacrificing performance. We show that Health-SCORE provides two practical benefits beyond standalone evaluation: it can be used as a structured reward signal to guide reinforcement learning with safety-aware supervision, and it can be incorporated directly into prompts to improve response quality through in-context learning. Across open-ended healthcare tasks, Health-SCORE achieves evaluation quality comparable to human-created rubrics while significantly lowering development effort, making rubric-based evaluation and training more scalable.

cs.AI↗

arXiv2025

Photon Phase-Space Dynamics in a Plasma Wakefield Accelerator

Frequency up-shifting of laser light in a beam-driven plasma wakefield has the potential to provide high-intensity sources of short wavelength radiation. Simulations have demonstrated that a laser pulse can undergo large frequency shifts, limited only by the drive beam energy, when the plasma density is tailored to match the accelerating phase of the wake to the group velocity of the pulse. Here, we study the dynamical evolution of photons in the phase-space vicinity of the plasma wake-phase matching condition. Numerical calculations using a photon kinetic model are validated by direct comparison with full electromagnetic particle-in-cell simulations. These calculations form the basis of a linear theory of the photon dynamics which reveals several important results, including scalings for the properties of the witness pulse and a self-similar solution for the photon phase-space dynamics. One prediction of the theory is that the pulse can be compressed indefinitely with no lower bound on the duration. This predication suggests that photon acceleration can provide a novel source of sub-femtosecond, short wavelength radiation.

physics.plasm-ph↗

arXiv2025

Widespread remote introgression in the grass genomes

Genetic transfers are pervasive across both prokaryotes and eukaryotes, encompassing canonical genomic introgression between species or genera and horizontal gene transfer (HGT) across kingdoms. However, DNA transfer between phylogenetically distant species, here defined as remote introgression (RI), has remained poorly explored in evolutionary genomics. In this study, we present RIFinder, a novel phylogeny-based method for RI event detection, and apply it to a comprehensive dataset of 122 grass genomes. Our analysis identifies 622 RI events originating from 543 distinct homologous genes, revealing distinct characteristics among grass subfamilies. Specifically, the subfamily Pooideae exhibits the highest number of introgressed genes while Bambusoideae contains the lowest. Comparisons among accepted genes, their donor copies and native homologs demonstrate that introgressed genes undergo post-transfer localized adaptation, with significant functional enrichment in stress-response pathways. Notably, we identify a large Triticeae-derived segment in a Chloridoideae species Cleistogenes songorica, which is potentially associated with its exceptional drought tolerance. Furthermore, we provide compelling evidence that RI has contributed to the origin and diversification of biosynthetic gene clusters of gramine, a defensive alkaloid chemical, across grass species. Collectively, our study establishes a robust method for RI detection and highlights its critical role in adaptive evolution.

q-bio.PE↗

arXiv2025

Reaching extreme fields in laser-electron beam collisions with XUV laser light

Plasma-based particle accelerators promise to extend the revolutionary work performed with conventional particle accelerators to studies with smaller footprints, lower costs, and higher energies. Here, we propose a new approach to access an unexplored regime of strong-field quantum electrodynamics by plasma wakefield acceleration of both charged particles and photons. Instead of using increasingly powerful accelerators and lasers, we show that photon acceleration of optical pulses into the extreme ultraviolet allows multi-GeV electrons to reach quantum nonlinearity parameters $χ_e \gg 10$ with a high probability due to the reduced radiative losses. A significant fraction of photons produced in high-$χ_e$ regions will propagate to detectors without generating pairs because of the reduction in the quantum rates. The photon spectra obtained may be used to characterize the predicted breakdown of strong-field quantum electrodynamics theory as it enters the fully non-perturbative regime.

physics.plasm-ph↗

arXiv2023

Parametric study of the polarization dependence of nonlinear Breit-Wheeler pair creation process using two laser pulses

With the rapid development of high-power petawatt class lasers worldwide, exploring physics in the strong field QED regime will become one of the frontiers for laser-plasma interactions research. Particle-in-cell codes, including quantum emission processes, are powerful tools for predicting and analyzing future experiments where the physics of relativistic plasma is strongly affected by strong-field QED processes. The spin/polarization dependence of these quantum processes has been of recent interest. In this article, we perform a parametric study of the interaction of two laser pulses with an ultrarelativistic electron beam. The first pulse is optimized to generate high-energy photons by nonlinear Compton scattering and efficiently decelerate the electron beam through quantum radiation reaction. The second pulse is optimized to generate electron-positron pairs by nonlinear Breit-Wheeler decay of the photons with the maximum polarization dependence. This may be experimentally realized as a verification of the strong field QED framework, including the spin/polarization rates.

hep-ph↗

arXiv2021

Beyond optimization -- supervised learning applications in relativistic laser-plasma experiments

We explore the applications of machine learning techniques in relativistic laser-plasma experiments beyond optimization purposes. We predict the beam charge of electrons produced in a laser wakefield accelerator given the laser wavefront change caused by a deformable mirror. Machine learning enables feature analysis beyond merely searching for an optimal beam charge, showing that specific aberrations in the laser wavefront are favored in generating higher beam charges. Supervised learning models allow characterizing the measured data quality as well as recognizing irreproducible data and potential outliers. We also include virtual measurement errors in the experimental data to examine the model robustness under these conditions. This work demonstrates how machine learning methods can benefit data analysis and physics interpretation in a highly nonlinear problem of relativistic laser-plasma interaction.

physics.plasm-ph↗

arXiv2019

On the connections between algorithmic regularization and penalization for convex losses

In this work we establish the equivalence of algorithmic regularization and explicit convex penalization for generic convex losses. We introduce a geometric condition for the optimization path of a convex function, and show that if such a condition is satisfied, the optimization path of an iterative algorithm on the unregularized optimization problem can be represented as the solution path of a corresponding penalized problem.

math.OC↗

arXiv2019

GR-MHD disk winds and jets from black holes and resistive accretion disks

We perform GR-MHD simulations of outflow launching from thin accretion disks. As in the non-relativistic case, resistivity is essential for the mass loading of the disk wind. We implemented resistivity in the ideal GR-MHD code HARM3D, extending previous works (Qian et al. 2017, 2018) for larger physical grids, higher spatial resolution, and longer simulation time. We consider an initially thin, resistive disk orbiting the black hole, threaded by a large-scale magnetic flux. As the system evolves, outflows are launched from the black hole magnetosphere and the disk surface. We mainly focus on disk outflows, investigating their MHD structure and energy output in comparison with the Poynting-dominated black hole jet. The disk wind encloses two components -- a fast component dominated by the toroidal magnetic field and a slower component dominated by the poloidal field. The disk wind transitions from sub to super-Alfvénic speed, reaching velocities $\simeq 0.1c$. We provide parameter studies varying spin parameter and resistivity level, and measure the respective mass and energy fluxes. A higher spin strengthens the $B_ϕ$-dominated disk wind along the inner jet. We disentangle a critical resistivity level that leads to a maximum matter and energy output for both, resulting from the interplay between re-connection and diffusion, which in combination govern the magnetic flux and the mass loading. For counter-rotating black holes the outflow structure shows a magnetic field reversal. We estimate the opacity of the inner-most accretion stream and the outflow structure around it. This stream may be critically opaque for a lensed signal, while the axial jet funnel remains optically thin.

astro-ph.HE↗

arXiv2019

The Implicit Bias of AdaGrad on Separable Data

We study the implicit bias of AdaGrad on separable linear classification problems. We show that AdaGrad converges to a direction that can be characterized as the solution of a quadratic optimization problem with the same feasible set as the hard SVM problem. We also give a discussion about how different choices of the hyperparameters of AdaGrad might impact this direction. This provides a deeper understanding of why adaptive methods do not seem to have the generalization ability as good as gradient descent does in practice.

stat.ML↗

arXiv2018

Jet launching in resistive GR-MHD black hole - accretion disk systems

We investigate the launching mechanism of relativistic jets from black hole sources, in particular the strong winds from the surrounding accretion disk. Numerical investigations of the disk wind launching - the simulation of the accretion-ejection transition - have so far almost only been done for non-relativistic systems. From these simulations we know that resistivity, or magnetic diffusivity, plays an important role for the launching process. Here, we extend this treatment to general relativistic magnetohydrodynamics (GR-MHD) applying the resistive GR-MHD code rHARM. Our model setup considers a thin accretion disk threaded by a large-scale open magnetic field. We run a series of simulations with different Kerr parameter, field strength and diffusivity level. Indeed we find strong disk winds with, however, mildly relativistic speed, the latter most probably due to our limited computational domain. Further, we find that magnetic diffusivity lowers the efficiency of accretion and ejection, as it weakens the efficiency of the magnetic lever arm of the disk wind. As major driving force of the disk wind we disentangle the toroidal magnetic field pressure gradient, however,magneto-centrifugal driving may also contribute. Black hole rotation in our simulations suppresses the accretion rate due to an enhanced toroidal magnetic field pressure that seems to be induced by frame-dragging. Comparing the energy fluxes from the Blandford-Znajek-driven central spine and the surrounding disk wind, we find that the total electromagnetic energy flux is dominated by the total matter energy flux of the disk wind (by a factor 20). The kinetic energy flux of the matter outflow is comparatively small and comparable to the Blandford-Znajek electromagnetic energy flux.

astro-ph.HE↗

arXiv2016

rHARM: Accretion and Ejection in Resistive GR-MHD

Turbulent magnetic diffusivity plays an important role for accretion disks and the launching of disk winds. We have implemented magnetic diffusivity, respective resistivity in the general relativistic MHD code HARM. This paper describes the theoretical background of our implementation, its numerical realization, our numerical tests and preliminary applications. The test simulations of the new code rHARM are compared with an analytic solution of the diffusion equation and a classical shock tube problem. We have further investigated the evolution of the magneto-rotational instability (MRI) in tori around black holes for a range of magnetic diffusivities. We find indication for a critical magnetic diffusivity (for our setup) beyond which no MRI develops in the linear regime and for which accretion of torus material to the black hole is delayed. Preliminary simulations of magnetically diffusive thin accretion disks around Schwarzschild black holes that are threaded by a large-scale poloidal magnetic field show the launching of disk winds with mass fluxes of about 50% of the accretion rate. The disk magnetic diffusivity allows for efficient disk accretion that replenishes the mass reservoir of the inner disk area and thus allows for long-term simulations of wind launching for more than 5000 time units.

astro-ph.HE↗