arXiv subjects

Zhenquan Li

Publications and source records attributed to Zhenquan Li.

8 recordsLinked to original sources

arXiv2026

Think in 360°: Benchmarking and Bootstrapping Panoramic Video Reasoning

A panoramic video captures all directions around the camera, but relevant evidence can appear in different directions and at different times. Understanding it requires models to connect observations across the viewing sphere and time. We introduce PanoVideoBench, the first manually annotated benchmark for comprehensive panoramic video understanding. Its 1,000 questions span five categories and 15 question types, with diagnostic annotations for target search and panoramic capabilities. Evaluation of 11 proprietary, open-source, and panoramic-specialized models identifies cross-view tracking and sphere-time evidence integration as their weakest capabilities. To strengthen these capabilities, we construct PanoVideoCoT, the first reasoning dataset grounded in real panoramic video, through sphere-time-grounded synthesis that converts verified panoramic observations and successful agentic explorations into 13,465 QA-CoT pairs. Starting from a general video MLLM, PanoVideoReasoner learns from this corpus through Panoramic CoT SFT and Evidence-Privileged On-Policy Self-Distillation, a panoramic instantiation of Vision-OPD's OPSD in which the teacher uses verified perspective views to guide student-generated trajectories. It reaches 55.7% on PanoVideoBench, improving the backbone by 12.5 points and the previous SOTA by 2.6. The gains concentrate on the same spatial, cross-view, and temporal capabilities exposed by the benchmark, showing that grounded sphere-time supervision can equip a general video MLLM with panoramic reasoning.

cs.CV↗

arXiv2026

PhysCount: Do MLLMs Count Visual Appearances or Physical Entities?

Images reveal appearances, but counting often asks about physical objects. A single entity may be visible directly, in several reflections, or through glass, so detecting every visible occurrence does not determine how many entities are present. Existing visual counting benchmarks rarely separate these two operations. We introduce PhysCount, a benchmark for evaluating whether multimodal large language models (MLLMs) can distinguish what appears in an image from what physically exists in the scene. PhysCount contains 546 real-world images and 4,595 human-written questions organized into 651 target-specific groups. For each group, annotators locate the relevant appearances, label how each is observed, and identify appearances of the same entity. Four complementary question families test the enumeration of appearances, the enumeration of physical entities, image-formation-specific counting, and appearance--entity correspondence. We measure both accuracy on each family and the proportion of groups answered entirely correctly. Across 40 MLLMs, counting visible appearances is substantially easier than recovering the identities that connect them. Even the strongest model answers every question correctly in only 13.36\% of groups. Performance declines when one entity produces many appearances or when its appearances cross observation modes. Higher image resolution improves access to visual evidence but brings limited gains in complete-group accuracy, whereas answering related questions jointly produces a marked improvement. These results expose a gap between recognizing what is visible and correctly recovering the physical entities behind those observations.

cs.CV↗

arXiv2026

Towards Multimodal Lifelong Understanding: A Dataset and Agentic Baseline

While datasets for video understanding have scaled to hour-long durations, they typically consist of densely concatenated clips that differ from natural, unscripted daily life. To bridge this gap, we introduce MM-Lifelong, a dataset designed for Multimodal Lifelong Understanding. Comprising 181.1 hours of footage, it is structured across Day, Week, and Month scales to capture varying temporal densities. Extensive evaluations reveal two critical failure modes in current paradigms: end-to-end MLLMs suffer from a Working Memory Bottleneck due to context saturation, while representative agentic baselines experience Global Localization Collapse when navigating sparse, month-long timelines. To address this, we propose the Recursive Multimodal Agent (ReMA), which employs dynamic memory management to iteratively update a recursive belief state, significantly outperforming existing methods. Finally, we establish dataset splits designed to isolate temporal and domain biases, providing a rigorous foundation for future research in supervised learning and out-of-distribution generalization.

cs.CV↗

arXiv2004

The accurate and comprehensive model of thin fluid flows with inertia on curved substrates

Consider the 3D flow of a viscous Newtonian fluid upon a curved 2D substrate when the fluid film is thin as occurs in many draining, coating and biological flows. We derive a comprehensive model of the dynamics of the film, the model being expressed in terms of the film thickness and the average lateral velocity. Based upon centre manifold theory, we are assured that the model accurately includes the effects of the curvature of substrate, gravitational body force, fluid inertia and dissipation. The model may be used to resolve wave-like phenomena in the dynamics of viscous fluid flows over arbitrarily curved substrates such as cylinders, tubes and spheres. We briefly illustrate its use in simulating drop formation on cylindrical fibres, wave transitions, Faraday waves, viscous hydraulic jumps, and flow vortices in a compound channel. These models are the most complete models for thin film flow of a Newtonian fluid; many other thin film models can be obtained by different truncations of the dynamical equations given herein.

chao-dyn↗

arXiv2003

Low-dimensional modelling of a generalized Burgers equation

Burgers equation is one of the simplest nonlinear partial differential equations-it combines the basic processes of diffusion and nonlinear steepening. In some applications it is appropriate for the diffusion coefficient to be a time-dependent function. Using a Wayne's transformation and centre manifold theory, we derive l-mode and 2-mode centre manifold models of the generalised Burgers equations for bounded smooth time dependent coefficients. These modellings give some interesting extensions to existing results such as the similarity solutions using the similarity method.

math-ph↗

arXiv2000

A flexible error estimate for the application of centre manifold theory

In applications of centre manifold theory we need more flexible error estimates than that provided by, for example, the Approximation Theorem~3 by Carr (1981,1983). Here we extend the theory to cover the case where the order of approximation in parameters and that in dynamical variables may be completely different. This allows, for example, the effective evaluation of low-dimensional dynamical models at finite parameter values.

math.DS↗

arXiv1999

Modelling the dynamics of turbulent floods

Consider the dynamics of turbulent flow in rivers, estuaries and floods. Based on the widely used k-epsilon model for turbulence, we use the techniques of centre manifold theory to derive dynamical models for the evolution of the water depth and of vertically averaged flow velocity and turbulent parameters. This new model for the shallow water dynamics of turbulent flow: resolves the vertical structure of the flow and the turbulence; includes interaction between turbulence and long waves; and gives a rational alternative to classical models for turbulent environmental flows.

patt-sol↗

arXiv1998

The Accurate Modelling of Thin 3D Fluid Flows with Inertia on Curved Substrates

Models and simulations of the flow of thin films of fluids have many important applications in industrial and natural processes. We consider the motion of a thin layer of an incompressible, Newtonian fluid over an arbitrary solid, stationary curved substrate. Earlier work by Roy, et al has modeled slow fluid flows in which inertia is negligible. However, in many applications a model expressed in terms of both the fluid layer thickness and the average lateral velocity is needed to resolve faster dynamics. Here we use centre manifold techniques to derive such a non-dimensional, approximate model of the three dimensional flow.

patt-sol↗