arXiv subjects

Xiaoyi Li

Publications and source records attributed to Xiaoyi Li.

9 recordsLinked to original sources

arXiv2026

SA-RA-JSCC: SNR-Adaptive and Semantic-Rate-Aware Joint Source-Channel Coding

In joint source-channel coding (JSCC)-based semantic communication systems, achieving stable and reliable image semantic transmission under channel constraints remains a key challenge. In most channel adaptation modules, the signal-to-noise ratio (SNR) is often injected into each layer of a channel-adaptation model in an independent and layer-wise manner, which undermines global coordination across layers. Therefore, consistent noise-robust representations may fail to be learned throughout the model. To address this problem, we propose SA-RA-JSCC, a novel channel-adaptive JSCC model. SA-RA-JSCC maps SNR into a unified semantic vector in the feature space and then applies a one-shot global reweighting to the encoded features, thereby enabling globally consistent and learnable channel adaptation. Moreover, in order to further enhance the anti-channel capability of semantic information, a semantic-rate-aware module is introduced, enabling the adaptive policy to respond simultaneously to fluctuations in channel quality and changes in semantic-rate constraints, thereby enhancing global network coordination and channel adaptivity. Extensive experiment results across multiple channels and datasets demonstrate that SA-RA-JSCC significantly outperforms existing semantic communication models in terms of reconstruction metrics such as PSNR and MS-SSIM, exhibiting stronger robustness across a broad range of SNR regimes.

cs.IT↗

arXiv2026

Do Post-Training Algorithms Actually Differ? A Controlled Study Across Model Scales Uncovers Scale-Dependent Ranking Inversions

Post-training alignment has produced dozens of competing algorithms -- DPO, SimPO, KTO, GRPO, and others -- yet practitioners lack controlled comparisons to guide algorithm selection. We present OXRL, a unified framework implementing 51 post-training algorithms with identical infrastructure, enabling the first large-scale apples-to-apples evaluation. Our study spans 8 algorithms across 4 model scales (0.5B--7B), 3 evaluation domains, and a 20-variant DPO taxonomy (100 runs at 1.5B, 5 seeds each), totaling $\sim$240 training runs on H100 GPUs. Three headline findings emerge. (1)~Algorithm rankings are unstable across scale: at 1.5B, online RL (SGRPO) tops all methods at 58.0\%~$\pm$0.57 on GSM8K; by 7B, the worst small-scale method (SimPO) becomes the best (85.8\%), a complete ranking inversion driven by model scale rather than LoRA regularization (confirmed via 2$\times$2 factorial). (2)~Loss function modifications yield negligible gains: none of 20 DPO variants significantly outperform vanilla DPO after Bonferroni correction; the sole significant outlier, SimPO, is worse ($-$11.5~pp, $p < 10^{-4}$). (3)~Algorithm leverage is task-specific: the 19.3~pp GSM8K spread collapses to 0.54~pp on MATH ($36\times$) and 0.47~pp on general-domain benchmarks ($41\times$), confirming that algorithm choice matters primarily within the training distribution. These findings yield a hierarchy of leverage for practitioners: model scale (${\sim}$50~pp) $\gg$ training paradigm (${\sim}$10~pp) $\gg$ online vs.\ offline (${\sim}$9~pp) $\gg$ loss function (${\sim}$1~pp). We release all code, configs, and evaluation data as a living community benchmark.

cs.LG↗

arXiv2026

Auto Researching, not hyperparameter tuning: Convergence Analysis of 10,000 Experiments

When LLM agents autonomously design ML experiments, do they perform genuine architecture search -- or do they default to hyperparameter tuning within a narrow region of the design space? We answer this question by analyzing 10,469 experiments executed by two LLM agents (Claude Opus and Gemini 2.5 Pro) across a combinatorial configuration space of 108,000 discrete cells for dashcam collision detection over 27 days. Through ANOVA decomposition, we find that \textbf{architectural choices explain 94\% of performance variance} ($F = 1324$, $η^2 = 0.94$), while hyperparameter variation within a fixed architecture explains only 6\%. Cross-task validation on a second collision dataset confirms this finding (75\% architecture-explained variance) with a \emph{different} winning backbone, confirming genuine architecture discovery. The agents' key contribution is discovering that V-JEPA\,2 video features with Zipformer temporal encoders achieve 0.9245 AP -- a configuration no human proposed -- and concentrating search on productive architectural regions: at $N = 50$, LLM-guided search reaches AP $= 0.985$ versus $0.965$ for from-scratch random search. Post-bugfix convergence follows a power law ($c = 0.11$, $R^2 = 0.93$); the low exponent reflects the cost of broad exploration, not inefficiency, since the LLM discovers qualitatively better regions than random or Bayesian baselines. We characterize multi-agent search dynamics via entropy cycles and Jensen--Shannon specialization, providing the first large-scale empirical framework for LLM-guided combinatorial ML experiment design.

cs.LG↗

arXiv2025

A new class of one-step A-stable and L-stable schemes of high-order accuracy for parabolic type equations

Recently, a new class of BDF schemes proposed in [F. Huang and J. Shen, SIAM J Numer. Anal., 62.4, 1609--1637] for the parabolic type equations are studied in this paper. The basic idea is based on the Taylor expansions at time $t^{n+β}$ with $β>1$ being a tunable parameter. These new BDF schemes allow larger time steps at higher order r for stiff problems than that which allowed with a usual higher-order scheme. However, multi-step methods like BDF exhibit inherent disadvantages relative to one-step methods in practical implementations. In this paper, inspired by their excellent work, we construct a new class of high-order one-step schemes for linear parabolic-type equations. These new schemes, with several suitable $β_i$, can achieve A-stable, or even L-stable. Specially, the new scheme with special parameters $β_i$ can be regarded as the classical one-step Runge-Kutta scheme with a stabilized term. Besides, we provide two different techniques to construct the one-step high-order schemes: the first one is by choosing different parameters $β_i$, and the second one is by increasing the number of intermediate layers. Both methods have been proven to be highly effective and even exhibit superconvergence property. Finally, we also conducted several numerical experiments to support our conclusions.

math.NA↗

arXiv2023

Lagrangian Trajectory Modeling of Lunar Dust Particles

A mathematical model and software implementation developed to predict trajectories of single lunar dust particles acted on by a high velocity gas flow is discussed. The model uses output from a computation fluid dynamics (CFD) or direct simulation Monte Carlo (DSMC) simulation of a rocket nozzle hot gas jet. The gas density, velocity vector field, and temperature predicted by the CFD/DSMC simulations, provide the data necessary to compute the forces and accelerations acting on a single particle of regolith. All calculations of trajectory assume that the duration of particle flight is much shorter than the change in gas properties, i.e., the particle trajectory calculations take into account the spatial variation of the gas jet, but not the temporal variation. This is a reasonable first-order assumption. Final results are compared to photogrammetry derived estimates of dust angles form Apollo landing videos.

astro-ph.EP↗

arXiv2021

Scaling of Erosion Rate in Subsonic Jet Experiments and Apollo Lunar Module Landings

Small scale jet-induced erosion experiments are useful for identifying the scaling of erosion with respect to the various physical parameters (gravity, grain size, gas velocity, gas density, grain density, etc.), and because they provide a data set for benchmarking numerical flow codes. We have performed experiments varying the physical parameters listed above (e.g., gravity was varied in reduced gravity aircraft flights). In all these experiments, a subsonic jet of gas impinges vertically on a bed of sand or lunar soil simulant forming a localized scour hole beneath the jet. Videography captures the erosion and scour hole formation processes, and analysis of these videos post-test identifies the scaling of these processes. This has produced important new insights into the physics of erosion. Based on these insights, we have developed an erosion rate model that can be applied to generalized situations, such as the erosion of soil beneath a horizontal gas flow on a planetary surface. This is important to lunar exploration because the rate of erosion beneath the rocket exhaust plume of a landing spacecraft will determine the amount of sand-blasting damage that can be inflicted upon surrounding hardware. Although the rocket exhaust plume at the exit of the nozzle is supersonic, the boundary layer on the lunar surface where erosion occurs is subsonic. The model has been benchmarked through comparison with the Apollo landing videos, which show the blowing lunar soil, and computational fluid dynamics simulations of those landings.

astro-ph.EP↗

arXiv2021

ISRU Implications for Lunar and Martian Plume Effects

Experiments, analyses, and simulations have shown that the engine exhaust plume of a Mars lander large enough for human spaceflight will create a deep crater in the martian soil, blowing ejecta to approximately 1 km distance, damaging the bottom of the lander with high-momentum rock impacts, and possibly tilting the lander as the excavated hole collapses to become a broad residual crater upon engine cutoff. Because of this, we deem that we will not have adequate safety margins to land humans on Mars unless we robotically stabilize the soil to form in situ landing pads prior to the mission. It will take a significant amount of time working in a harsh off-planet environment to develop and certify the new technologies and procedures for in situ landing pad construction. The only place to reasonably accomplish this is on the Moon.

astro-ph.EP↗

arXiv2020

A Scenario-Based Development Framework for Autonomous Driving

This article summarizes the research progress of scenario-based testing and development technology for autonomous vehicles. We systematically analyzed previous research works and proposed the definition of scenario, the elements of the scenario ontology, the data source of the scenario, the processing method of the scenario data, and scenario-based V-Model. Moreover, we summarized the automated test scenario construction method by random scenario generation and dangerous scenario generation.

cs.DC↗

arXiv2018

Phase field and level set methods for modeling solute precipitation and/or dissolution

The dynamics of solid-liquid interfaces controlled by solute precipitation and/or dissolution due to the chemical reaction at the interface were computed in two dimensions using a phase field models. Sharp-interface asymptotic analysis demonstrated that the phase field solutions should converge to the proper sharp-interface precipitation/dissolution limit. For the purpose of comparison, the numerical solution of the sharp-interface model for solute precipitation/dissolution was directly solved using a level set method. In general, the phase field results are found in good agreement with the level set results for all reaction rates and geometry configurations investigated. Present study supports the applications of both methods to more complicated and realistic reactive systems, including the nuclear waste release and mineral precipitation and dissolution.

physics.comp-ph↗