arXiv subjects

Zilu Zhou

Publications and source records attributed to Zilu Zhou.

6 recordsLinked to original sources

arXiv2026

FlexiGrad: Adaptive Gradient Modulation for Hierarchical Fine-Grained Classification

Many fine-grained recognition tasks contain hierarchical labels such as order, family and species. Although this supervision should be beneficial, jointly optimising all levels often leads to unstable training because coarse and fine classifiers impose inconsistent gradients on the shared backbone. This hierarchical gradient conflict prevents the model from learning a coherent coarse-to-fine representation. In this paper, we propose FlexiGrad, a simple and parameter-free method that regulates gradient interactions during backpropagation. FlexiGrad removes only the harmful conflicting component when tasks disagree and reinforces the shared direction when they partially agree through a smooth hierarchy-aware weighting function. This produces stable optimisation and preserves both global structure and fine-grained discriminative cues. FlexiGrad integrates into existing architectures without modification while improves multi-granularity accuracy on CUB-200-2011, FGVC-Aircraft and Stanford Cars. The code will be available at PRIS-CV/FlexiGrad.

cs.CV↗

arXiv2026

Seeing as Experts Do: A Knowledge-Augmented Agent for Open-Set Fine-Grained Visual Understanding

Fine-grained visual understanding is shifting from static classification to knowledge-augmented reasoning, where models must justify as well as recognise. Existing approaches remain limited by closed-set taxonomies and single-label prediction, leading to significant degradation under open-set or context-dependent conditions. We present the Knowledge-Augmented Fine-Grained Reasoning Agent (KFRA), a unified framework that transforms fine-grained perception into evidence-driven reasoning. KFRA operates through a three-stage closed reasoning loop that emulates expert analysis. It first performs open-vocabulary detection and web-scale retrieval to generate category hypotheses. It then conducts discriminative regions localisation by aligning textual knowledge with visual evidence through a global-to-local focusing mechanism. Finally, it integrates all multimodal evidence within a large multimodal model to perform interpretable reasoning. Unlike existing agents that treat retrieval and reasoning as independent processes, KFRA establishes a retrieval-grounding coupling that converts retrieved knowledge into spatially grounded evidence for verification. This design enables factual, interpretable, and task-agnostic reasoning across diverse fine-grained scenarios. To evaluate this capability, we construct FGExpertBench, a benchmark designed to assess reasoning depth and cross-task generalisation across six knowledge dimensions. Extensive experiments demonstrate that KFRA consistently surpasses both standalone large multimodal models and current agent frameworks, achieving up to 19 percent improvement in reasoning accuracy and delivering evidence-grounded interpretability in open-set fine-grained visual understanding.

cs.CV↗

arXiv2026

ABCMB: A Python+JAX Package for the Cosmic Microwave Background Power Spectrum

We present ABCMB, a differentiable Einstein-Boltzmann solver for the cosmic microwave background (CMB). ABCMB is a complete code capturing important effects to linear order in $\Lambda{\rm CDM}$ cosmology. It computes the CMB power spectrum and includes effects like lensing, polarization, massive neutrinos, and a state-of-the-art treatment of BBN and recombination. ABCMB has sub-percent-level agreement with CLASS and can be run on a GPU with competitive, and sometimes even faster, run times. It is refactored compared to previous codes and takes advantage of object-oriented programming to improve extensibility, meaning new physics can be added to it without the need for modifying source files. ABCMB provides accurate and stable gradients to the user, making Fisher analyses straightforward, and enabling the use of efficient gradient-based sampling methods.

astro-ph.CO↗

arXiv2024

Searching for Dark Matter Interactions with ACT, SPT and DES

Models of a dark radiation sector with a mass threshold (WZDR+) have proved to be an appealing alternative to $Λ$CDM. These models provide simple comparison models, grounded in well-understood particle physics and with limited additional parameters. In addition, they have shown relevance in easing existing cosmological tensions, specifically the $H_0$ tension and the $S_8$ tension. Recently, measurements of CMB lensing by the ACT collaboration have provided strong additional information on clustering at late times. Within $Λ$CDM, these results yield a high value of $S_8$ at odds with weak-lensing measurements. In this work, we study this in the context of WZDR+, and find a much wider range of allowed values of $S_8$, and in particular much better agreement between data sets and an overall improvement of fit versus $Λ$CDM. We expand our analyses to include a wide set of data, including the ACT-DR6 lensing data, as well as primary CMB information from ACT-DR4 and SPT-3G, scale-dependent power spectra from DES and measurements of $H_0$ from SH0ES. We find that there is little to no tension in measurements of structure within the data sets, and the inferred value of $S_8$ is generally lower than that in $Λ$CDM. We find that the inclusion of DES generally favors a higher $H_0$, but there is some direct tension between the high-$\ell$ multipole data and this result. Future data should clarify whether this is a statistical artifact, or a true incompatibility of these datasets within this model.

hep-ph↗

arXiv2022

Constraining the Star Formation Rate using Joint CIB Continuum and [CII] Intensity Mapping

Line intensity mapping (LIM) experiments probing the nearby universe can expect a considerable amount of cosmic infrared background (CIB) contiuum emission coming from near and far-infrared galaxies. For the purpose of using the LIM data to constrain the star formation rate (SFR), we argue that the CIB continuum - traditionally treated as contamination - can be combined with the LIM signal to enhance the SFR constraints achievable. We first present a power spectrum model that is capable of joining continuum and line emissions that assume the same prior SFR model. We subsequently analyze the effectiveness of the joint model in the context of the EXperiment for Cryogenic Large-Aperture Intensity Mapping (EXCLAIM), which utilizes the [CII] molecular line to study the SFR. We numerically compute the theoretical power spectra according to our model and the EXCLAIM survey specifics, and perform Fisher analysis to obtain SFR parameter constraints. We find that although the joint model has no considerable advantage over LIM alone assuming the current survey level of EXCLAIM, its effects become significant when we consider more optimistic values of survey resolution and angular span that are expected of future LIM experiments. By manipulating the Fisher formalism, we show that the CIB is not only an additional SFR sensitive signal, but also serves to break the SFR parameter degeneracy that naturally emerges from the [CII] Fisher matrix. For this reason, addition of the CIB will allow improvements in the survey parameters to be better reflected in the SFR constraints, and can be effectively utilized by future LIM experiments.

astro-ph.CO↗

arXiv2022

Extragalactic Science with the Experiment for Cryogenic Large-aperture Intensity Mapping

The EXperiment for Cryogenic Large-Aperture Intensity Mapping (EXCLAIM) is a balloon-borne cryogenic telescope that will survey the spectrum of diffuse emission from both the Milky Way and the cosmic web to probe star formation, the interstellar medium, and galaxy evolution across cosmic time. EXCLAIM's primary extragalactic science survey maps 305 deg$^2$ along the celestial equator with an R=512 spectrometer over the frequency range ν=420-540 GHz, targeting emission of the [CII] line over redshifts 2.5<z<3.5 and several CO lines for z<1. Cross-correlation with galaxy redshift catalogs isolates line emission from the large-scale structure at target redshifts. In this paper, we forecast the sensitivity for both the two-point and conditional one-point cross-correlation. We predict that EXCLAIM will detect both the [CII]-QSO cross-power spectrum and the conditional voxel intensity distribution (CVID) at various redshifts under a broad range of [CII] intensity models, allowing it to differentiate among these models in the literature. These forecasts for the power spectra include the effects of line interlopers and continuum foreground contamination. We then convert the joint [CII] constraints from both the cross-power spectrum and the CVID into constraints on the [CII] halo luminosity-mass relation $L_\mathrm{[CII]}(M)$ model parameters and the star formation rate density (SFRD) from [CII] emission. We also develop sensitivity estimates for CO, showing the ability to differentiate between models.

astro-ph.CO↗