arXiv subjects

Lei Yi

Publications and source records attributed to Lei Yi.

At least 19 recordsLinked to original sources

arXiv2026

Turbulence Decay Intensifies Clustering of Bubbles and Particles

Our understanding of inertial particle dynamics in turbulence is mostly based on flows held in a statistically stationary state, a particular regime that differs from many natural flows where energy input can often be intermittent or cyclic, or may abruptly cease. Here we investigate inertial particle and bubble dynamics in freely decaying turbulence through complementary experiments and direct numerical simulations. While particle accelerations decay monotonically in time, we find evidence that the clustering can exhibit a non-monotonic evolution, intensifying sharply before subsequently weakening. We demonstrate that both the acceleration and clustering behaviors can be mapped onto their counterparts in stationary turbulence using a dynamic rescaling of the evolving length and time scales of the turbulence. Validity conditions for the dynamic rescaling, satisfied by both the experimental and numerical datasets, are derived. The proposed mappings remain applicable across a broad range of density ratios, from light to heavy particles, and particle sizes spanning two orders of magnitude in Stokes number.

physics.flu-dyn↗

arXiv2026

Hydrodynamic Brachistochrone: Conflicting Paths of Time and Energy Minima within Viscous Media

We experimentally and theoretically study the hydrodynamic analog of the classical brachistochrone problem: the {\it time-} and {\it energy-minimizing} paths for a spherical particle rolling down an incline within a viscous fluid. We show that in the presence of viscous dissipation, the paths of minima diverge from the classical cycloid, into curves of opposing curvature for time and energy, and are characterized by an effective dimensionless parameter, $St_p$, representing the ratio of the particle's viscous response time scale to its gravitational time scale. Using a generalized variational framework, we show that the fastest path reduces to {nearly straight ramps}, however, beginning and terminating in localized cycloids of curvature, $\kappa_c \sim St_p^{-2}$. Remarkably, the path of fastest descent on a given energy budget requires navigating a non-monotonic path ({\it``S-shaped''}) with an interior point of inflection. Our findings reveal a unification of temporal and energetic optimality for transport through dissipative media, and expand the celebrated brachistochrone solutions to the hydrodynamic regime.

physics.flu-dyn↗

arXiv2026

HODAgent: Towards On-Demand, Responsive Humanoids for Physical World Human Interaction

We propose HODAgent, a System-2 embodied agent for humanoid robots in service settings, addressing situated intent, responsive execution, task revision, and outcome verification. Its semi-duplex architecture integrates an Env-Interactor, Planner, Executor, and hierarchical Memory to maintain coherent interaction, planning, and task state during service episodes. This allows handling new requests during motion, retaining progress, revising actions, and grounding closure in execution outcomes. A shared interface connects simulation and physical robots (Unitree G1), isolating platform-specific control. In an interactive simulation with 164 cases, HODAgent achieves 84.8% and 91.5% Joint Success under two VLM backbones, outperforming baselines by 9.8 and 18.9 points. On physical robots, pass rates are 92% (atomic), 72% (composite), and 63.3% (complete tasks). On multiple embodied benchmarks, it improves over baselines by 0.7-9.0 points. Results show a unified System-2 agent enables adaptive humanoid service across simulation and reality.

cs.RO↗

arXiv2026

Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence

Vision-language models are increasingly serving as the reasoning core of embodied agents. Robot execution is inherently iterative: each action reshapes the scene and physical state, continually renewing what must be perceived, reasoned about, and verified. Meeting these demands requires complementary capabilities that differ in supervision signals, prediction formats, and verification criteria. Existing approaches typically develop these capabilities against isolated, task-specific objectives, leaving open how they should be organized and integrated around execution as a whole. We present Capek 0.5, an embodied vision-language model built around an execution-centric capability taxonomy. Rather than organizing training by datasets or tasks, the taxonomy groups embodied capabilities according to their functional roles throughout execution and comprises four capability families: Spatial Reasoning, Temporal Understanding, Action Guidance, and State Verification. Each capability is first acquired by a dedicated specialist through reinforcement learning with verifiable rewards from a shared backbone, and the specialists are then consolidated into a single inference-time model through weight-space merging followed by routed policy-space distillation. We instantiate Capek 0.5 at the 2B and 35B-A3B scales and evaluate it from three complementary perspectives: comprehensive benchmark suites including Capek-StateBench, a new benchmark for state verification; a controlled study of capability retention from specialists to the unified model; and closed-loop evaluation in simulated embodied environments. Capek 0.5 improves the large majority of matched benchmark rows over its initialization, retains all four specialized capabilities in one checkpoint with quantified losses, and transfers to closed-loop embodied task execution.

cs.AI↗

arXiv2026

Done, But Not Sure: Disentangling World Completion from Self-Termination in Embodied Agents

Standard embodied evaluations do not independently score whether an agent correctly commits to task completion at episode closure, a capacity we call terminal commitment. Behaviorally distinct failures--never completing the task, completing it but failing to stop, and reporting success without sufficient evidence--collapse into the same benchmark failure. We introduce VIGIL, an evaluation framework that makes terminal commitment independently measurable. Under VIGIL's default protocol, agents observe only egocentric RGB, receive no action-success signals, and must end each episode with a semantic report checked deterministically against hidden world state. This yields two separate scores: world-state completion (W) and benchmark success (B), where B additionally requires a correct terminal report. This decoupling makes four outcome categories distinguishable: missed execution, post-attainment drift, unsupported commitment, and verified success. Across 20 models on 1,000 frozen episodes, systems with comparable W differ by up to 19.7 pp in B: one model converts achieved states into correct reports, while another with near-identical execution drifts past the goal without closing. An action-feedback intervention further tests the separation: execution-oriented signals improve W broadly, yet commitment failures persist in models that do not already ground terminal reports in the achieved state. VIGIL provides a protocol that makes terminal commitment independently visible and scorable.

cs.AI↗

arXiv2025

Non-reciprocal coalescence-breakup dynamics in concentrated emulsions

Dense stabilized emulsions are mixtures of immiscible fluids where the high-volume fraction droplet dispersed phase is stabilized against coalescence by steric interactions. The production of emulsions involves high-shear flows and it is well known that at a critical volume fraction the emulsion loses stability, undergoing an extremely rapid process where the fluid components in the emulsion exchange roles. This process, called catastrophic phase inversion, which resembles in several respects a dynamical phase transition, has remained widely elusive from an experimental and theoretical point of view. In this work, we present state-of-the-art experimental and numerical data to support a dynamical-system framework capable of precisely highlighting the dynamics occurring in the system as it approaches the catastrophic phase inversion. Our study clearly highlights that at high volume fractions, dynamical changes in the emulsion morphology, due to coalescence and breakup of droplets, play a critical role in determining emulsion's rheology and stability. Additionally, we show that at approaching the critical volume fractions, the dynamics can be simplified as being controlled by the dynamics of a correlation length represented, in our systems, by the size of the largest droplet. This dynamics shares a close connection with non-reciprocal phase transition where two different physical mechanisms, coalescence and breakups, can get out of balance leading to large non-symmetric periodic excursions in phase space. We clarify the phenomenology observed and quantitatively explain the essential aspect of the highly complex dynamics of stabilized emulsions undergoing catastrophic phase inversion.

cond-mat.soft↗

arXiv2025

SQLord: A Robust Enterprise Text-to-SQL Solution via Reverse Data Generation and Workflow Decomposition

Transforming natural language into SQL queries (NL2SQL) is crucial for data-driven business applications. Existing frameworks, trained on open-source datasets, struggle with complex business logic and lack domain-specific data for fine-tuning. Additionally, evaluation methods often require annotated data and executable database environments, which are scarce in real-world scenarios. To address these challenges, we propose SQLord, an enterprise-level NL2SQL framework. First, SQLord introduces a data reverse generation approach to convert raw SQL statements into annotated data for supervised fine-tuning (SFT). Second, it proposes a decomposition method for complex queries using an automated workflow generator. Additionally, SQLord features a comprehensive GPT-Judge evaluation framework, including Execution Evaluation (EXE), Query-SQL Evaluation (QSE), and SQL-SQL Evaluation (SSE), tailored to diverse scenarios. Offline tests significantly outperform state of the art baselines, and online accuracy consistently exceeds 90, highlighting SQLord's advantages and effectiveness in complex real world scenarios. SQLord has been successfully applied across multiple scenarios on the world's largest B2B e-commerce platform.

cs.DB↗

arXiv2025

ASPO: Adaptive Sentence-Level Preference Optimization for Fine-Grained Multimodal Reasoning

Direct Preference Optimization (DPO) has gained significant attention for its simplicity and computational efficiency in aligning large language models (LLMs). Recent advancements have extended DPO to multimodal scenarios, achieving strong performance. However, traditional DPO relies on binary preference optimization, rewarding or penalizing entire responses without considering fine-grained segment correctness, leading to suboptimal solutions. The root of this issue lies in the absence of fine-grained supervision during the optimization process. To address this, we propose Adaptive Sentence-level Preference Optimization (ASPO), which evaluates individual sentences for more precise preference optimization. By dynamically calculating adaptive rewards at the sentence level based on model predictions, ASPO enhances response content assessment without additional models or parameters. This significantly improves the alignment of multimodal features. Extensive experiments show that ASPO substantially enhances the overall performance of multimodal models.

cs.CL↗

arXiv2025

Experimental study of velocity statistics in wall-bounded turbulent emulsions

Turbulent emulsions are ubiquitous in chemical engineering, food processing, pharmaceuticals, and other fields. However, our experimental understanding of this area remains limited due to the multi-scale nature of turbulent flow and the presence of extensive interfaces, which pose significant challenges to optical measurements. In this study, we address these challenges by precisely matching the refractive indices of the continuous and dispersed phases, enabling us to measure local velocity information at high volume fractions. The emulsion is generated in a turbulent Taylor-Couette flow, with velocity measured at two radial locations: near the inner cylinder (boundary layer) and in the middle gap (bulk region). Near the inner cylinder, the presence of droplets suppresses the emission of angular velocity plumes, which reduces the mean azimuthal velocity and its root mean squared fluctuation. The former effect leads to a higher angular velocity gradient in the boundary layer, resulting in greater global drag on the system. In the bulk region, although droplets suppress turbulence fluctuations, they enhance the cross-correlation between azimuthal and radial velocities, leaving the angular velocity flux contributed by the turbulent flow nearly unchanged. In both locations, droplets suppress turbulence at scales larger than the average droplet diameter and increase the intermittency of velocity increments. However, the effects of the droplets are more pronounced near the inner cylinder than in the bulk, likely because droplets fragment in the boundary layer but are less prone to breakup in the bulk. Our study provides experimental insights into how dispersed droplets modulate global drag, coherent structures, and the multi-scale characteristics of turbulent flow.

physics.flu-dyn↗

arXiv2024

CoF: Coarse to Fine-Grained Image Understanding for Multi-modal Large Language Models

The impressive performance of Large Language Model (LLM) has prompted researchers to develop Multi-modal LLM (MLLM), which has shown great potential for various multi-modal tasks. However, current MLLM often struggles to effectively address fine-grained multi-modal challenges. We argue that this limitation is closely linked to the models' visual grounding capabilities. The restricted spatial awareness and perceptual acuity of visual encoders frequently lead to interference from irrelevant background information in images, causing the models to overlook subtle but crucial details. As a result, achieving fine-grained regional visual comprehension becomes difficult. In this paper, we break down multi-modal understanding into two stages, from Coarse to Fine (CoF). In the first stage, we prompt the MLLM to locate the approximate area of the answer. In the second stage, we further enhance the model's focus on relevant areas within the image through visual prompt engineering, adjusting attention weights of pertinent regions. This, in turn, improves both visual grounding and overall performance in downstream tasks. Our experiments show that this approach significantly boosts the performance of baseline models, demonstrating notable generalization and effectiveness. Our CoF approach is available online at https://github.com/Gavin001201/CoF.

cs.CV↗

arXiv2024

Enhancing Fine-Grained Vision-Language Pretraining with Negative Augmented Samples

Existing Vision-Language Pretraining (VLP) methods have achieved remarkable improvements across a variety of vision-language tasks, confirming their effectiveness in capturing coarse-grained semantic correlations. However, their capability for fine-grained understanding, which is critical for many nuanced vision-language applications, remains limited. Prevailing VLP models often overlook the intricate distinctions in expressing different modal features and typically depend on the similarity of holistic features for cross-modal interactions. Moreover, these models directly align and integrate features from different modalities, focusing more on coarse-grained general representations, thus failing to capture the nuanced differences necessary for tasks demanding a more detailed perception. In response to these limitations, we introduce Negative Augmented Samples(NAS), a refined vision-language pretraining model that innovatively incorporates NAS to specifically address the challenge of fine-grained understanding. NAS utilizes a Visual Dictionary(VD) as a semantic bridge between visual and linguistic domains. Additionally, it employs a Negative Visual Augmentation(NVA) method based on the VD to generate challenging negative image samples. These samples deviate from positive samples exclusively at the token level, thereby necessitating that the model discerns the subtle disparities between positive and negative samples with greater precision. Comprehensive experiments validate the efficacy of NAS components and underscore its potential to enhance fine-grained vision-language comprehension.

cs.CV↗

arXiv2024

How interfacial tension enhances drag in turbulent Taylor-Couette flow with neutrally buoyant and equally viscous droplets

The presence of dispersed-phase droplets can result in a notable increase in the system's drag. However, our understanding of the mechanism underlying this phenomenon remains limited. In this study, we use three-dimensional direct numerical simulations with a modified multi-marker volume-of-fluid method to investigate liquid-liquid two-phase turbulence in a Taylor-Couette geometry. The dispersed phase has the same density and viscosity as the continuous phase. The Reynolds number $Re\equiv r_i\omega_i d/\nu$ is fixed at 5200, the volume fraction of the dispersed phase is up to $40\%$, and the Weber number $We\equiv \rho u^2_\tau d/\sigma$ is around 8. It is found that the increase in the system's drag originates from the contribution of interfacial tension. Specifically, droplets experience significant deformation and stretching in the streamwise direction due to shear near the inner cylinder. Consequently, the rear end of the droplets lags behind the fore head. This causes opposing interfacial tension effects on the fore head and rear end of the droplets. For the fore head of the droplets, the effect of interfacial tension appears to act against the flow direction. For the rear end, the effect appears to act in the flow direction. The increase in the system's drag is primarily attributed to the effect of interfacial tension on the fore head of the droplets which leads to the hindering effect of the droplets on the surrounding continuous phase. This hindering effect disrupts the formation of high-speed streaks, favoring the formation of low-speed ones, which are generally associated with higher viscous stress and drag of the system. This study provides new insights into the mechanism of drag enhancement reported in our previous experiments.

physics.flu-dyn↗

arXiv2024

Numerical study on the mechanism of drag modulation by dispersed drops in two-phase Taylor-Couette turbulence

The presence of a dispersed phase can significantly modulate the drag in turbulent systems. We derived a conserved quantity that characterizes the radial transport of azimuthal momentum in the fluid-fluid two-phase Taylor-Couette turbulence. This quantity consists of contributions from advection, diffusion, and two-phase interface, which are closely related to density, viscosity, and interfacial tension, respectively. We found that the presence of the two-phase interface consistently produces a positive contribution to the momentum transport and leads to drag enhancement, while decreasing the density and viscosity ratios of the dispersed phase to the continuous phase reduces the contribution of local advection and diffusion terms to the momentum transport, respectively, resulting in drag reduction. Therefore, we concluded that the decreased density ratio and the decreased viscosity ratio work together to compete with the presence of two-phase interface for achieving drag modulation in fluid-fluid two-phase turbulence.

physics.flu-dyn↗

arXiv2023

Divergence of critical fluctuations on approaching catastrophic phase inversion in turbulent emulsions

Catastrophic phase inversion, the sudden breakdown of a dense emulsion, occurs when the dispersed majority phase irreversibly exchanges role with the continuous minority phase. This common process has been extensively studied over the past decades and yet its fundamental physical mechanism has remained largely unexplored. Here we experimentally and numerically study the dynamics of catastrophic phase inversion as it occurs when the volume fraction of the dispersed phase exceeds a critical volume fraction (typically around 92% in experiments). Our data accurately quantify the abrupt change of both the global torque and average droplet size at approaching and across the phase inversion point, exhibiting strong hysteresis. Most importantly, we reveal that the fluctuations in the global torque diverge as a power-law while approaching the critical volume-fraction and we connect their growth to the formation of highly heterogeneous spatial droplet structures. The present finding, unveiling the tight connection between fluctuations in dynamic heterogeneity and the critical divergence of torque fluctuation, paves the way to a quantitative description of catastrophic phase inversion as an out-of-equilibrium critical-like phenomena.

physics.flu-dyn↗

arXiv2022

Turbulence drag modulation by dispersed droplets in Taylor-Couette flow: the effects of the dispersed phase viscosity

The dispersed phase in turbulence can vary from almost inviscid fluid to highly viscous fluid. By changing the viscosity of the dispersed droplet phase, we experimentally investigate how the deformability of dispersed droplets affects the global transport quantity of the turbulent emulsion. Different kinds of silicone oil are employed to result in the viscosity ratio, $ζ$, ranging from $0.53$ to $8.02$. The droplet volume fraction, $ϕ$, is varied from 0\% to 10\% with a spacing of 2\%. The global transport quantity, quantified by the normalized friction coefficient $c_{f,ϕ}/c_{f,ϕ=0}$, shows a weak dependence on the turbulent intensity due to the vanishing finite-size effect of the droplets. The interesting fact is that, with increasing $ζ$, the $c_{f,ϕ}/c_{f,ϕ=0}$ first increases and then saturates to a plateau value which is similar to that of the rigid particle suspension. By performing image analysis, this drag modification is interpreted from the aspect of droplet deformability, which is responsible for the breakup and coalescence effect of the droplets. The statistics of the droplet size distribution show that, with increasing $ζ$, the stabilizing effect induced by interfacial tension comes to be substantial and the pure inertial breakup process becomes dominant. The measurement of the droplet distribution along the radial direction of the system shows a bulk-clustering effect, which can be attributed to the non-negligible coalescence effect of the droplet. It is found that the droplet coalescence effect could be suppressed as the $ζ$ increases, thereby affecting the contribution of interfacial tension to the total stress, and accounting for the observed emulsion rheology.

physics.flu-dyn↗

arXiv2022

Recent developments of turbulent emulsions in Taylor-Couette flow

Emulsions are common in many natural and industrial settings. Recently, much attention has been put on understanding the dynamics of turbulent emulsions. This paper reviews some recent studies of emulsions in turbulent Taylor-Couette flow, mainly focusing on the statistics of the dispersed phase and the global momentum transport of the system. We first study the size distribution and the breakup mechanism of the dispersed droplets for turbulent emulsions with a low volume-fraction (dilute) of the dispersed phase. For systems with a high volume-fraction of the dispersed phase (dense), we address the detailed response of the global transport (effective viscosity) of the turbulent emulsion and its connection to the droplet statistics. Finally, we will discuss catastrophic phase inversions, which can happen when the volume fraction of the dispersed phase exceeds a critical value during dynamic emulsification. We end the manuscript with a summary and an outlook including some open questions for future research. This article is part of the theme issue `Taylor-Couette and Related Flows on the Centennial of Taylor's Seminal Philosophical Transactions Paper'.

physics.flu-dyn↗

arXiv2022

Physical mechanisms for droplet size and effective viscosity asymmetries in turbulent emulsions

By varying the oil volume fraction, the microscopic droplet size and the macroscopic rheology of emulsions are investigated in a Taylor-Couette (TC) turbulent shear flow. Although here oil and water in the emulsions have almost the same physical properties (density and viscosity), unexpectedly, we find that oil-in-water (O/W) and water-in-oil (W/O) emulsions have very distinct hydrodynamic behaviors, i.e., the system is clearly asymmetric. By looking at the micro-scales, the average droplet diameter hardly changes with the oil volume fraction neither for O/W nor for W/O. However, for W/O it is about 50% larger than that of O/W. At the macro-scales, the effective viscosity of O/W is higher when compared to that of W/O. These asymmetric behaviors can be traced back to the presence of surface-active contaminants in the system. By introducing an oil-soluble surfactant at high concentration, remarkably, we recover the symmetry (droplet size and effective viscosity) between O/W and W/O emulsions. Based on this, we suggest a possible mechanism responsible for the initial asymmetry. Next, we discuss what sets the droplet size in turbulent emulsions. We uncover a -6/5 scaling dependence of the droplet size on the Reynolds number of the flow. Combining the scaling dependence and the droplet Weber number, we conclude that the droplet fragmentation, which determines the droplet size, occurs within the boundary layer and is controlled by the dynamic pressure caused by the gradient of the mean flow, as proposed by Levich (1962), instead of the dynamic pressure due to turbulent fluctuations, as proposed by Kolmogorov (1949). The present findings provide an understanding of both the microscopic droplet formation and the macroscopic rheological behaviors in dynamic emulsification, and connects them.

physics.flu-dyn↗

arXiv2021

How do the finite-size particles modify the drag in Taylor-Couette turbulent flow

We experimentally investigate the drag modification by neutrally buoyant finite-size particles with various aspect ratios in a Taylor-Couette (TC) turbulent flow. The current Reynolds number, $Re$, ranges from $6.5\times10^3$ to $2.6\times10^4$, and the particle volume fraction, $Φ$, is up to $10\%$. Particles with three kinds of aspect ratio, $λ$, are used: $λ=1/3$ (oblate), $λ=1$ (spherical) and $λ=3$ (prolate). Unlike the case of bubbly TC flow, we find that the suspended finite-size particles increase the drag of the TC system regardless of their aspect ratios. In addition, the normalized friction coefficient, $c_{f,Φ}/c_{f,Φ=0}$, decreases with increasing $Re$, the reason could be that in the current low volume fractions the turbulent stress becomes dominant at higher $Re$. As $Re$ increases, the particles distribute more evenly in the entire system, which results from both the greater turbulence intensity and the more pronounced finite-size effects of the particles at higher $Re$. Moreover, it is found that the variation of the particle aspect ratios leads to different particle collective effects. The suspended spherical particles, which tend to cluster near the walls and form a particle layer, significantly affect the boundary layer and result in maximum drag modification. The minimal drag modification is found in the oblate case, where the particles preferentially cluster in the bulk region, and thus the particle layer is absent. Based on the optical measurement results, it can be concluded that, in the low volume fraction ranges ($Φ=0.5\%$ and $Φ= 2\%$ here), the larger drag modification is connected to the near-wall particle clustering. The present findings suggest that the particle shape plays a significant role in drag modification.

physics.flu-dyn↗