arXiv subjects

Huichao Li

Publications and source records attributed to Huichao Li.

10 recordsLinked to original sources

arXiv2026

Personalizing Large Language Model Agents with Small Policy Models

Large language model (LLM) agents can retrieve memory, call tools, ask clarifying questions, and vary response style, yet adapting these execution decisions to an individual user remains difficult. Fine-tuning a separate LLM is costly or impossible for proprietary systems, while prompts and memory primarily expose user information to the agent rather than adapt its execution decisions from feedback. We formulate personalization of a frozen agent as online learning of a per-user execution policy from scalar feedback observed only for the executed action. We propose FABLE (Factorized Adaptive Bandit Layer for Execution), a lightweight policy layer outside a potentially black-box host agent. FABLE factorizes memory, information-acquisition, and response decisions so feedback updates related choices; filters actions through an externally specified feasible set before exploration; and learns user-specific residual preferences relative to a fixed default-and-cost score via Bayesian contextual Thompson sampling. Under a linear residual-reward model, a calibrated variant inherits an expected-regret bound against the best feasible action. We also characterize preferences unidentifiable under persistent feasibility constraints and provide anytime-valid false-promotion control. Across personalized-reasoning, controlled-feedback, and executable tool-use evaluations, FABLE improves several preference-sensitive behaviors relative to rule-only control while remaining competitive on end-to-end task performance.

cs.AI↗

arXiv2026

Deep Multitask Learning for Mixed-Type Outcomes with Shared Sparsity

Most existing multitask learning approaches are limited by their reliance on task-specific loss functions tailored to the scale and type of each outcome. When outcomes differ across tasks, these losses are generally not directly comparable, which makes it difficult to formulate a unified objective and may limit information sharing across tasks. We propose a multitask transformation framework in which task-specific responses may differ through unknown monotone transformations. Motivated by high-dimensional biological applications in which the predictor dimension may diverge with the sample size while only a common subset of predictors is informative, we consider shared sparsity across tasks. Under this framework, we estimate the target functions and identify important predictors by optimizing a smoothed rank-based criterion with a group-Lasso penalty, implemented through a multitask deep neural network with a shared first layer. We establish the nonasymptotic excess-risk bounds, and variable-selection consistency for the proposed estimator. Simulation studies show that the proposed method achieves competitive prediction and variable-selection performance compared with competing approaches. Analyses of gene-expression studies with continuous, binary, and mixed outcomes further illustrate that the proposed method improves prediction and identifies biologically meaningful shared predictors.

stat.ML↗

arXiv2026

Modeling of Coronal Mass Ejection Originated from a Sheared Arcade of Realistic Active-Region Scale and Its Propagation in the Heliosphere: Methodology

Simulating coronal mass ejections (CMEs) from their origin in active regions (ARs) to their propagation to Earth remains challenging, particularly when aiming to resolve AR scales and employ realistic magnetic field strengths without compromising computational efficiency. Here we present a methodology for end-to-end CME modeling that addresses these challenges. Three nested magnetohydrodynamic simulations are coupled to jointly cover the heliosphere from solar surface to beyond $1.5$ au. A block-structured adaptive mesh refinement scheme is employed to achieve $\sim 700$ km resolution in the low corona, allowing AR scales to be resolved while maintaining the total grid count below $10^8$ across the entire computational domain. A semi-relativistic Boris correction combined with a relativistic mass-density factor is used to handle magnetic field strengths up to $10^3$ G without prohibitively small time steps. Using this model, we simulate the emergence of a bipolar AR into the corona, the initiation of a CME by shearing of the AR core field and the subsequent evolution. Our simulation captures its pre-eruption energy buildup, triggering by magnetic reconnection, rapid acceleration, and propagation to 1 au and beyond. The simulated CME exhibits a three-part structure in synthetic coronagraph images and a torus-shaped flux rope in the heliosphere, with synthetic in-situ observations showing shock formation, density compression, and a prolonged southward $B_z$ component at 1 au. The entire simulation requires about one day on a moderately sized cluster (e.g., $600$ processors), while the simulated CME takes three days to arrive at $1$ au, offering a lead time of two days if used for forecasting.

astro-ph.SR↗

arXiv2026

Hierarchical Contrastive Learning for Multimodal Data

Multimodal representation learning is commonly built on a shared-private decomposition, treating latent information as either common to all modalities or specific to one. This binary view is often inadequate: many factors are shared by only subsets of modalities, and ignoring such partial sharing can over-align unrelated signals and obscure complementary information. We propose Hierarchical Contrastive Learning (HCL), a framework that learns globally shared, partially shared, and modality-specific representations within a unified model. HCL combines a hierarchical latent-variable formulation with structural sparsity and a structure-aware contrastive objective that aligns only modalities that genuinely share a latent factor. Under uncorrelated latent variables, we prove identifiability of the hierarchical decomposition, establish recovery guarantees for the loading matrices, and derive parameter estimation and excess-risk bounds for downstream prediction. Simulations show accurate recovery of hierarchical structure and effective selection of task-relevant components. On multimodal electronic health records, HCL yields more informative representations and consistently improves predictive performance.

stat.ML↗

arXiv2024

On the Northward Shift of the Heliospheric Current Sheet at the End of Solar Cycle 24

Since solar cycle 16, the { heliospheric} current sheet (HCS) has been found to be shifted southward during the late declining to minimum phase. However, this trend is broken at the end of solar cycle 24. In this paper, we analyze the shift of the HCS by using information obtained from coronal model and insitu data provide by the near-Earth OMNI database and the Parker Solar Probe (PSP). Coronal potential field source surface (PFSS) modeling results show that the northward shift is established at the beginning of 2018 and remains stable for about two years. Interplanetary magnetic field data obtained from and within 1 au also support the northward shift, as the southern polarity T appears more frequently than the northern polarity A between 2018-2020. Both model results and insitu observation obtained by PSP imply that the HCS shift is established in the corona, and then propagates into the heliosphere. The quadrupole term still has a significant influence on the formation of the HCS shift.

astro-ph.SR↗

arXiv2021

Is Solar Minimum 24/25 Another Unusual One?

The solar minimum 23/24 is considered to be unusual because it exhibits features that differ notably from those commonly seen in pervious minima. In this letter, we analyze the solar polar magnetic field, the potential-field solution of the solar corona, and the in-situ solar wind measurements to see whether the recent solar minimum 24/25 is another unusual one. While the dipolar configuration that are commonly seen during minimum 22/23 and earlier minima persist for about half a year after the absolute minimum of solar cycle 24, the corona has a morphology more complex than a simple dipole before the absolute minimum. The fast solar wind streams are less dominant than minimum 23/24. The IMF strength, density and mass flux that are historically low in the minimum 23/24 are regained during minimum 24/25, but still do not reach the minimum 22/23 level. From the analysis of this Letter, it seems that the minimum 24/25 is only partially unusual, and the recovery of the commonly minimum features may result from the enhancement of the polar field.

astro-ph.SR↗

arXiv2013

Stabilization of Quantum Spin Hall Effect by Designed Removal of Time-Reversal Symmetry of Edge States

The quantum spin Hall (QSH) effect is known to be unstable to perturbations violating time-reversal symmetry. We show that creating a narrow ferromagnetic (FM) region near the edge of a QSH sample can push one of the counterpropagating edge states to the inner boundary of the FM region, and leave the other at the outer boundary, without changing their spin polarizations and propagation directions. Since the two edge states are spatially separated into different "lanes", the QSH effect becomes robust against symmetry-breaking perturbations.

cond-mat.mes-hall↗

arXiv2013

Topological phase transitions with and without energy gap closing

Topological phase transitions in a three-dimensional (3D) topological insulator (TI) with an exchange field of strength $g$ are studied by calculating spin Chern numbers $C^\pm(k_z)$ with momentum $k_z$ as a parameter. When $|g|$ exceeds a critical value $g_c$, a transition of the 3D TI into a Weyl semimetal occurs, where two Weyl points appear as critical points separating $k_z$ regions with different first Chern numbers. For $|g|<g_c$, $C^\pm(k_z)$ undergo a transition from $\pm 1$ to 0 with increasing $|k_z|$ to a critical value $k_z^{\tiny C}$. Correspondingly, surface states exist for $|k_z| < k_z^{\tiny C}$, and vanish for $|k_z| \ge k_z^{\tiny C}$. The transition at $|k_z| = k_z^{\tiny C}$ is acompanied by closing of spin spectrum gap rather than energy gap.

cond-mat.mes-hall↗

arXiv2012

A topological look at the quantum spin Hall state

We propose a topological understanding of the quantum spin Hall state without considering any symmetries, and it follows from the gauge invariance that either the energy gap or the spin spectrum gap needs to close on the system edges, the former scenario generally resulting in counterpropagating gapless edge states. Based upon the Kane-Mele model with a uniform exchange field and a sublattice staggered confining potential near the sample boundaries, we demonstrate the existence of such gapless edge states and their robust properties in the presence of impurities. These gapless edge states are protected by the band topology alone, rather than any symmetries.

cond-mat.mes-hall↗

arXiv2011

Quantum Hall Effect in Thin Films of Three-Dimensional Topological Insulators

We show that a thin film of a three-dimensional topological insulator (3DTI) with an exchange field is a realization of the famous Haldane model for quantum Hall effect (QHE) without Landau levels. The exchange field plays the role of staggered fluxes on the honeycomb lattice, and the hybridization gap of the surface states is equivalent to alternating on-site energies on the AB sublattices. A peculiar phase diagram for the QHE is predicted in 3DTI thin films under an applied magnetic field, which is quite different from that either in traditional QHE systems or in graphene.

cond-mat.mes-hall↗