arXiv subjects

Zhiqi Huang

Publications and source records attributed to Zhiqi Huang.

At least 55 records · Page 3Linked to original sources

arXiv2025

G1: Bootstrapping Perception and Reasoning Abilities of Vision-Language Model via Reinforcement Learning

Vision-Language Models (VLMs) excel in many direct multimodal tasks but struggle to translate this prowess into effective decision-making within interactive, visually rich environments like games. This ``knowing-doing'' gap significantly limits their potential as autonomous agents, as leading VLMs often performing badly in simple games. To address this, we introduce VLM-Gym, a curated reinforcement learning (RL) environment featuring diverse visual games with unified interfaces and adjustable, compositional difficulty, specifically designed for scalable multi-game parallel training. Leveraging VLM-Gym, we train G0 models using pure RL-driven self-evolution, which demonstrate emergent perception and reasoning patterns. To further mitigate challenges arising from game diversity, we develop G1 models. G1 incorporates a perception-enhanced cold start prior to RL fine-tuning. Our resulting G1 models consistently surpass their teacher across all games and outperform leading proprietary models like Claude-3.7-Sonnet-Thinking. Systematic analysis reveals an intriguing finding: perception and reasoning abilities mutually bootstrap each other throughout the RL training process. Source code including VLM-Gym and RL training are released at https://github.com/chenllliang/G1 to foster future research in advancing VLMs as capable interactive agents.

cs.CV↗

arXiv2025

Key drivers of the preference for dynamic dark energy

Joint analysis of the baryon acoustic oscillations (BAO) measurement by the Dark Energy Spectroscopic Instrument (DESI) first data release, Type Ia supernovae (SNe) of the Dark Energy Survey Year 5 (DES5YR) release and cosmic microwave background (CMB) data favors a quintom-like dynamic dark energy model over the standard Lambda cold dark matter ($Λ$CDM) model at $3.9σ$ level (Adame et al. 2024). We confirm the previous finding in the literature that the preference for dynamic dark energy does not rely on the detailed modeling of CMB physics and remains at a similar significance level ($3.2σ$) when the full CMB likelihood is replaced by a CMB acoustic-oscillation angle ($θ_\star$) prior and a baryon abundance ($Ω_bh^2$) prior. The computationally efficient $θ_\star$ and $Ω_bh^2$ priors allow us to take a frequentist approach by comparing DES5YR SNe and DESI BAO with a large number ($\gtrsim 10^4$) of Planck-constrained $Λ$CDM simulations. We find that $\geq 3.2σ$ preference for dynamic dark energy is very rare (occurrence rate = $0.28\%$) in simulations. When we combine DESI BAO with SN simulations or combine DES5YR SNe with BAO simulations, the occurrence rate of $\geq 3.2σ$ preference for dynamic dark energy increases to $1.2\%$ and $4.8\%$, respectively. These results indicate an internal inconsistency, i.e., a significant tension between DESI BAO + DES5YR SNe and Planck-constrained $Λ$CDM models in both Bayesian and frequentist points of view. Although both DESI BAO and DES5YR SNe contribute to the preference for dynamic dark energy, the contribution from DES5YR SNe is more significant. In the frequentist point of view, even DES5YR SNe alone is in tension with Planck-constrained $Λ$CDM models, though in Bayesian point of view this tension is prior dependent and inconclusive.

astro-ph.CO↗

arXiv2025

MoBA: Mixture of Block Attention for Long-Context LLMs

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

cs.LG↗

arXiv2024

Quantifying the tension between cosmological models and JWST red candidate massive galaxies

We develop a Python tool to estimate the tail distribution of the number of dark matter halos beyond a mass threshold and in a given volume in a light-cone. The code is based on the extended Press-Schechter model and is computationally efficient, typically taking a few seconds on a personal laptop for a given set of cosmological parameters. The high efficiency of the code allows a quick estimation of the tension between cosmological models and the red candidate massive galaxies released by the James Webb Space Telescope, as well as scanning the theory space with the Markov Chain Monte Carlo method. As an example application, we use the tool to study the cosmological implication of the candidate galaxies presented in Labbé et al. (2023). The standard $Λ$ cold dark matter ($Λ$CDM) model is well consistent with the data if the star formation efficiency can reach $\sim 0.3$ at high redshift. For a low star formation efficiency $ε\sim 0.1$, $Λ$CDM model is disfavored at $\sim 2σ$-$3σ$ confidence level.

astro-ph.CO↗

arXiv2024

Pair Counting without Binning -- A New Approach to Correlation Functions in Clustering Statistics

This paper presents a novel perspective on correlation functions in the clustering analysis of the large-scale structure of the universe. We first recognise that pair counting in bins of radial separation is equivalent to evaluating counts-in-cells (CIC), which can be modelled using a filtered density field with a binning-window function. This insight leads to an in situ expression for the two-point correlation function (2PCF). Essentially, the core idea underlying our method is to introduce a window function to define the binning scheme, enabling pair-counting without binning. This approach develops a concept of generalised 2PCF, which extends beyond conventional discrete pair counting by accommodating non-sharp-edged window functions. To extend this framework to N-point correlation functions (NPCF) using current optimal edge-corrected estimators, we developed a binning scheme independent of the specific parameterisation of polyhedral configurations. In particular, we demonstrate a fast algorithm for the three-point correlation function (3PCF), where triplet counting is accomplished by assigning either a spherical tophat or a Gaussian filter to each vertex of triangles. Additionally, we derive analytical expressions for the 3PCF using a multipole expansion in Legendre polynomials, accounting for filtered field (binning) corrections. Numerical tests using several suites of N-body simulation samples show that our approach aligns remarkably well with the theoretical predictions. Our method provides an exact solution for quantifying binning effects in practical measurements and offers a high-speed algorithm, enabling high-order clustering analysis in extremely large datasets from ongoing and upcoming surveys such as Euclid, LSST, and DESI.

astro-ph.CO↗

arXiv2024

A PAge-like Unified Dark Fluid Model

The unified dark fluid model unifies dark matter and dark energy into a single component, providing an alternative and more concise framework for interpreting cosmological observations. We introduce a PAge-like Unified Dark Fluid (PUDF) model based on the PAge approximation (Huang 2020), which is parameterized by the age of the universe and an $η$ parameter indicating the deviation from Einstein-De Sitter Universe. The PUDF model shares many similar features of the standard Lambda cold dark matter ($Λ$CDM) model and can effectively describe the large-scale structure formation and late-time cosmic acceleration. We constrain the PUDF model with the Planck 2018 cosmic microwave background anisotropies, baryon acoustic oscillation measurements including those from the most recent DESI 2024, the Pantheon+ sample of Type Ia supernovae, and the Cosmic Chronometers compilation. Although the PUDF performs well in fitting all the cosmological datasets, the joint analysis of the data still favors the $Λ$CDM model over the PUDF model, according to the Bayesian evidence of model comparison.

astro-ph.CO↗

arXiv2024

Curvature perturbations from kinetic preheating after $α$-attractor inflation

Preheating at the end of inflation is a violent nonlinear process that efficiently transfers the energy of the inflaton to a second field, the preheat field. When the preheat field is light during inflation and its background value modulates the preheating process, the superhorizon isocurvature perturbations of the preheat field may be converted to curvature perturbations that leave an imprint on the cosmic microwave background and the large-scale structure of the universe. We use high-precision lattice simulations to study kinetic preheating after $α$-attractor inflation, a case where the effective mass of the preheat field is naturally suppressed during inflation. By comparing the expansion e-folds between different Hubble patches, we find that the conversion from isocurvature perturbations to curvature perturbations is very inefficient and can hardly be detected by cosmological observations.

astro-ph.CO↗

arXiv2024

Rhythmic Foley: A Framework For Seamless Audio-Visual Alignment In Video-to-Audio Synthesis

Our research introduces an innovative framework for video-to-audio synthesis, which solves the problems of audio-video desynchronization and semantic loss in the audio. By incorporating a semantic alignment adapter and a temporal synchronization adapter, our method significantly improves semantic integrity and the precision of beat point synchronization, particularly in fast-paced action sequences. Utilizing a contrastive audio-visual pre-trained encoder, our model is trained with video and high-quality audio data, improving the quality of the generated audio. This dual-adapter approach empowers users with enhanced control over audio semantics and beat effects, allowing the adjustment of the controller to achieve better results. Extensive experiments substantiate the effectiveness of our framework in achieving seamless audio-visual alignment.

cs.SD↗

arXiv2024

FD2Talk: Towards Generalized Talking Head Generation with Facial Decoupled Diffusion Model

Talking head generation is a significant research topic that still faces numerous challenges. Previous works often adopt generative adversarial networks or regression models, which are plagued by generation quality and average facial shape problem. Although diffusion models show impressive generative ability, their exploration in talking head generation remains unsatisfactory. This is because they either solely use the diffusion model to obtain an intermediate representation and then employ another pre-trained renderer, or they overlook the feature decoupling of complex facial details, such as expressions, head poses and appearance textures. Therefore, we propose a Facial Decoupled Diffusion model for Talking head generation called FD2Talk, which fully leverages the advantages of diffusion models and decouples the complex facial details through multi-stages. Specifically, we separate facial details into motion and appearance. In the initial phase, we design the Diffusion Transformer to accurately predict motion coefficients from raw audio. These motions are highly decoupled from appearance, making them easier for the network to learn compared to high-dimensional RGB images. Subsequently, in the second phase, we encode the reference image to capture appearance textures. The predicted facial and head motions and encoded appearance then serve as the conditions for the Diffusion UNet, guiding the frame generation. Benefiting from decoupling facial details and fully leveraging diffusion models, extensive experiments substantiate that our approach excels in enhancing image quality and generating more accurate and diverse results compared to previous state-of-the-art methods.

cs.CV↗

arXiv2024

Forecasting the BAO Measurements of the CSST galaxy and AGN Spectroscopic Surveys

The spectroscopic survey of China's Space Survey Telescope (CSST) is expected to obtain a huge number of slitless spectra, including more than one hundred million galaxy spectra and millions of active galactic nuclei (AGN) spectra. By making use of these spectra, we can measure the Baryon Acoustic Oscillation (BAO) signals over large redshift ranges with excellent precisions. In this work, we predict the CSST measurements of the post-reconstruction galaxy power spectra at $0<z<1.2$ and pre-reconstruction AGN power spectra at $0<z<4$, and derive the BAO signals at different redshift bins by constraining the BAO scaling parameters using the Markov Chain Monte Carlo method. Our result shows that the CSST spectroscopic survey can provide accurate BAO measurements with precisions higher than 1\% and 3\% for the galaxy and AGN surveys, respectively. By comparing with current measurements in the same range at low redshifts, this can improve the precisions by a factor of $2\sim3$, and similar precisions can be obtained in the pessimistic case. We also investigate the constraints on the cosmological parameters using the measured BAO data by the CSST, and obtain stringent constraint results for the energy density of dark matter, Hubble constant, and equation of state of dark energy.

astro-ph.CO↗

arXiv2024

Mask-ControlNet: Higher-Quality Image Generation with An Additional Mask Prompt

Text-to-image generation has witnessed great progress, especially with the recent advancements in diffusion models. Since texts cannot provide detailed conditions like object appearance, reference images are usually leveraged for the control of objects in the generated images. However, existing methods still suffer limited accuracy when the relationship between the foreground and background is complicated. To address this issue, we develop a framework termed Mask-ControlNet by introducing an additional mask prompt. Specifically, we first employ large vision models to obtain masks to segment the objects of interest in the reference image. Then, the object images are employed as additional prompts to facilitate the diffusion model to better understand the relationship between foreground and background regions during image generation. Experiments show that the mask prompts enhance the controllability of the diffusion model to maintain higher fidelity to the reference image while achieving better image quality. Comparison with previous text-to-image generation methods demonstrates our method's superior quantitative and qualitative performance on the benchmark datasets.

cs.CV↗

arXiv2024

Cosmological constraints on neutrino masses in light of JWST red and massive candidate galaxies

The overabundance of the red and massive candidate galaxies observed by the James Webb Space Telescope (JWST) implies efficient structure formation or large star formation efficiency at high redshift $z\sim 10$. In the scenario of a low or moderate star formation efficiency, because massive neutrinos tend to suppress the growth of structure of the universe, the JWST observation tightens the upper bound of the neutrino masses. Assuming $Λ$ cold dark matter cosmology and a star formation efficiency $ \in [0.05, 0.3]$ (flat prior), we perform joint analyses of Planck+JWST and Planck+BAO+JWST, and obtain improved constraints $\sum m_ν< 0.196\,\mathrm{eV}$ and $\sum m_ν< 0.111\,\mathrm{eV}$ at 95% confidence level, respectively. Based on the above assumptions, the inverted mass ordering, which implies $\sum m_ν\geq 0.1\mathrm{eV}$, is excluded by Planck+BAO+JWST at 92.7% confidence level.

astro-ph.CO↗

arXiv2023

Toward a stellar population catalog in the Kilo Degree Survey: the impact of stellar recipes on stellar masses and star formation rates

The Kilo Degree Survey (KiDS) is currently the only sky survey providing optical ($ugri$) plus near-infrared (NIR, $ZYHJK_S$) seeing matched photometry over an area larger than 1000 $\rm deg^2$. This is obtained by incorporating the NIR data from the VISTA Kilo Degree Infrared Galaxy (VIKING) survey, covering the same KiDS footprint. As such, the KiDS multi-wavelength photometry represents a unique dataset to test the ability of stellar population models to return robust photometric stellar mass ($M_*$) and star-formation rate (SFR) estimates. Here we use a spectroscopic sample of galaxies for which we possess $u g r i Z Y J H K_s$ ``gaussianized'' magnitudes from KiDS data release 4. We fit the spectral energy distribution from the 9-band photometry using: 1) three different popular libraries of stellar {population} templates, 2) single burst, simple and delayed exponential star-formation history models, and 3) a wide range of priors on age and metallicity. As template fitting codes we use two popular softwares: LePhare and CIGALE. We investigate the variance of the stellar masses and the star-formation rates from the different combinations of templates, star formation recipes and codes to assess the stability of these estimates and define some ``robust'' median quantities to be included in the upcoming KiDS data releases. As a science validation test, we derive the mass function, the star formation rate function, and the SFR-$M_*$ relation for a low-redshift ($z<0.5$) sample of galaxies, that result in excellent agreement with previous literature data. The final catalog, containing $\sim290\,000$ galaxies with redshift $0.01<z<0.9$, is made publicly available.

astro-ph.GA↗

arXiv2023

Revisiting progenitor-age dependence of type Ia supernova luminosity standardization process

Much of the research in supernova cosmology is based on an assumption that the peak luminosity of type Ia supernovae (SNe Ia), after a standardization process, is independent of the galactic environment. A series of recent studies suggested that there is a significant correlation between the standardized luminosity and the progenitor age of SNe Ia. The correlation found in the most recent work by Lee et al. is strong enough to explain the extra dimming of distant SNe Ia and therefore casts doubts on the direct evidence of cosmic acceleration. The present work incorporates the uncertainties of progenitor ages, which were ignored in Lee et al., into a fully Bayesian inference framework. We find a weaker dependence of supernova standardized luminosity on the progenitor age, but the detection of correlation remains significant (3.5$σ$). Assuming that such correlation can be extended to high redshift and applying it to the Pantheon SN Ia data set, we confirm that when the Hubble residual does not include intrinsic scatter, the age-bias could be the primary cause of the observed extra dimming of distant SNe Ia. Furthermore, we use the PAge formalism, which is a good approximation to many dark energy and modified gravity models, to do a model comparison. We find that if intrinsic scatter is included in the Hubble residual, the Lambda cold dark matter model remains a good fit. However, in a scenario without intrinsic scatter, the Lambda cold dark matter model faces a challenge.

astro-ph.CO↗

arXiv2023

Are Large Language Models Really Robust to Word-Level Perturbations?

The swift advancement in the scales and capabilities of Large Language Models (LLMs) positions them as promising tools for a variety of downstream tasks. In addition to the pursuit of better performance and the avoidance of violent feedback on a certain prompt, to ensure the responsibility of the LLM, much attention is drawn to the robustness of LLMs. However, existing evaluation methods mostly rely on traditional question answering datasets with predefined supervised labels, which do not align with the superior generation capabilities of contemporary LLMs. To address this issue, we propose a novel rational evaluation approach that leverages pre-trained reward models as diagnostic tools to evaluate the longer conversation generated from more challenging open questions by LLMs, which we refer to as the Reward Model for Reasonable Robustness Evaluation (TREvaL). Longer conversations manifest the comprehensive grasp of language models in terms of their proficiency in understanding questions, a capability not entirely encompassed by individual words or letters, which may exhibit oversimplification and inherent biases. Our extensive empirical experiments demonstrate that TREvaL provides an innovative method for evaluating the robustness of an LLM. Furthermore, our results demonstrate that LLMs frequently exhibit vulnerability to word-level perturbations that are commonplace in daily language usage. Notably, we are surprised to discover that robustness tends to decrease as fine-tuning (SFT and RLHF) is conducted. The code of TREval is available in https://github.com/Harry-mic/TREvaL.

cs.CL↗

arXiv2023

Soft Prompt Decoding for Multilingual Dense Retrieval

In this work, we explore a Multilingual Information Retrieval (MLIR) task, where the collection includes documents in multiple languages. We demonstrate that applying state-of-the-art approaches developed for cross-lingual information retrieval to MLIR tasks leads to sub-optimal performance. This is due to the heterogeneous and imbalanced nature of multilingual collections -- some languages are better represented in the collection and some benefit from large-scale training data. To address this issue, we present KD-SPD, a novel soft prompt decoding approach for MLIR that implicitly "translates" the representation of documents in different languages into the same embedding space. To address the challenges of data scarcity and imbalance, we introduce a knowledge distillation strategy. The teacher model is trained on rich English retrieval data, and by leveraging bi-text data, our distillation framework transfers its retrieval knowledge to the multilingual document encoder. Therefore, our approach does not require any multilingual retrieval training data. Extensive experiments on three MLIR datasets with a total of 15 languages demonstrate that KD-SPD significantly outperforms competitive baselines in all cases. We conduct extensive analyses to show that our method has less language bias and better zero-shot transfer ability towards new languages.

cs.IR↗

arXiv2023

Sum of digamma asymptotic error terms of an arithmetic series

We define an S function as the sum of the asymptotic error terms of digamma function of an arithmetic series, $S(a) \equiv \sum_{n=1}^\infty \left[\ln\frac{n}{a} - \frac{a}{2n}-ψ\left(\frac{n}{a}\right)\right]$, and show a few properties of it. Using the S function, we construct a real and positive $ϕ$ function. Riemann hypothesis holds if $\tildeϕ(k)$, the complex Fourier transform of $ϕ$, has only real zeros.

math.GM↗

arXiv2023

Cross-lingual Knowledge Transfer via Distillation for Multilingual Information Retrieval

In this paper, we introduce the approach behind our submission for the MIRACL challenge, a WSDM 2023 Cup competition that centers on ad-hoc retrieval across 18 diverse languages. Our solution contains two neural-based models. The first model is a bi-encoder re-ranker, on which we apply a cross-lingual distillation technique to transfer ranking knowledge from English to the target language space. The second model is a cross-encoder re-ranker trained on multilingual retrieval data generated using neural machine translation. We further fine-tune both models using MIRACL training data and ensemble multiple rank lists to obtain the final result. According to the MIRACL leaderboard, our approach ranks 8th for the Test-A set and 2nd for the Test-B set among the 16 known languages.

cs.IR↗