arXiv subjects

Jingmin Chen

Publications and source records attributed to Jingmin Chen.

13 recordsLinked to original sources

arXiv2026

ReflectWorld-MM: An Entity-Oriented Multimodal Memory System for Open-Ended Video Streams

Building assistants that can continually watch the world, remember what they see, and reason over their accumulated experience is a long-standing goal, and recently multimodal agents equipped with long-term memory over video streams have attracted increasing interest. Unfortunately, existing systems either keep their memory inside the model context or in a flat feature store, and organize it around frames rather than around the persistent entities a stream is really about, which confines them to bounded videos and weakens their ability to track who and what reappears over time. In this paper, we propose ReflectWorld-MM, an entity-oriented multimodal memory system for open-ended video streams. It consists of three parts. The first is a perception front-end that turns an audiovisual stream into entity-resolved observations under a bounded short-term memory. The second is a hierarchical long-term memory, grounded in human memory theory, that couples a multi-scale episodic memory, an evolving entity-centric semantic memory, and a procedural memory. The third is a complete realization, built for real-world operation, that ingests arbitrary streams and plugs into off-the-shelf assistants. Across six long-video and lifelong-memory benchmarks, ReflectWorld-MM achieves the best accuracy on all six, outperforming strong memory agents and a frontier model.

cs.CV↗

arXiv2026

DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos

World models for deformable objects should recover not only geometry and appearance, but also underlying physical dynamics, interaction grounding, and material behavior. Learning such a model from real videos is challenging because deformable linear, planar, and volumetric objects evolve under high-dimensional deformation, noisy interactions, and complex material response. The model must therefore infer a physical state from visual observations, roll it forward under new interactions, and render the resulting dynamics with high visual fidelity. We present DeformMaster, a video-derived interactive physics-neural world model that turns real interaction videos into an online interactive model of deformable objects within a unified dynamics-and-appearance framework. DeformMaster preserves structured physical rollout while using a neural residual to compensate for unmodeled effects, grounds sparse hand motion as distributed compliant actuator for hand-continuum interaction, represents material response with spatially varying constitutive experts, and drives high-fidelity 4D appearance from the predicted physical evolution. Experiments on real-world deformable-object sequences demonstrate DeformMaster's ability to roll out future dynamics and render dynamic appearance, outperforming state-of-the-art baselines while supporting novel action rollout, material-parameter variation, and dynamic novel-view synthesis. Project page: https://can-lee.github.io/deformmaster-web/

cs.CV↗

arXiv2026

Move-Then-Operate: Behavioral Phasing for Human-Like Robotic Manipulation

We present Move-Then-Operate, a Vision language action framework that explicitly decouples robotic manipulation into two distinct behavioral phases: coarse relocation (move) and contact-critical interaction (operate). Unlike monolithic policies that conflate these heterogeneous regimes, our architecture employs a dual-expert policy routed by a learnable phase selector, introducing a structural inductive bias that isolates phase-specific dynamics. Phase labels are automatically generated via an MLLM-based pipeline conditioned on lightweight contextual cues such as end-effector velocity and subtask decomposition to ensure alignment with human motor patterns. Evaluated on the RoboTwin2 benchmark, our method achieves an average success rate of $68.9\%$, outperforming the monolithic $\pi_0$ baseline by $24\%$. It matches or exceeds models trained on $10\times$ more data and reaches peak performance in $40\%$ fewer training steps, demonstrating that architectural disentanglement of move and operate phases is a highly effective and efficient strategy for mastering high-precision manipulation.

cs.RO↗

arXiv2026

Gaussian Sequences with Multi-Scale Dynamics for 4D Reconstruction from Monocular Casual Videos

Understanding dynamic scenes from casual videos is critical for scalable robot learning, yet four-dimensional (4D) reconstruction under strictly monocular settings remains highly ill-posed. To address this challenge, our key insight is that real-world dynamics exhibits a multi-scale regularity from object to particle level. To this end, we design the multi-scale dynamics mechanism that factorizes complex motion fields. Within this formulation, we propose Gaussian sequences with multi-scale dynamics, a novel representation for dynamic 3D Gaussians derived through compositions of multi-level motion. This layered structure substantially alleviates ambiguity of reconstruction and promotes physically plausible dynamics. We further incorporate multi-modal priors from vision foundation models to establish complementary supervision, constraining the solution space and improving the reconstruction fidelity. Our approach enables accurate and globally consistent 4D reconstruction from monocular casual videos. Experiments of dynamic novel-view synthesis (NVS) on benchmark and real-world manipulation datasets demonstrate considerable improvements over existing methods.

cs.CV↗

arXiv2025

Egocentric Instruction-oriented Affordance Prediction via Large Multimodal Model

Affordance is crucial for intelligent robots in the context of object manipulation. In this paper, we argue that affordance should be task-/instruction-dependent, which is overlooked by many previous works. That is, different instructions can lead to different manipulation regions and directions even for the same object. According to this observation, we present a new dataset comprising fifteen thousand object-instruction-affordance triplets. All scenes in the dataset are from an egocentric viewpoint, designed to approximate the perspective of a human-like robot. Furthermore, we investigate how to enable large multimodal models (LMMs) to serve as affordance predictors by implementing a ``search against verifiers'' pipeline. An LMM is asked to progressively predict affordances, with the output at each step being verified by itself during the iterative process, imitating a reasoning process. Experiments show that our method not only unlocks new instruction-oriented affordance prediction capabilities, but also achieves outstanding performance broadly.

cs.RO↗

arXiv2025

Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective

Existing Multimodal Large Language Models (MLLMs) process a large number of visual tokens, leading to significant computational costs and inefficiency. Instruction-related visual token compression demonstrates strong task relevance, which aligns well with MLLMs ultimate goal of instruction following. Previous works generally assume that visual tokens achieve better vision-language alignment in the shallow layers of LLMs, which have led to task-related token compression being primarily applied in intermediate LLM layers. In contrast, our study reveals that with proper selection, task-related token compression is feasible at the input stage of LLM with negligible performance loss. This new paradigm significantly reduces task-irrelevant visual tokens and its model-agnostic design enables application without modifying the LLM architecture. Specifically, we suggest that explainability methods for transformer-based architechtures can evaluate the global importance of each visual token with respect to the given instruction, which can effectively guide the task-related token compression for MLLMs. Furthermore, we propose to learn a mapping from the attention map of the first LLM layer to the explanation results, thereby avoiding the need for a full inference pass. Interestingly, this mapping can be learned using a simple and lightweight convolutional network, whose training is efficient and independent of MLLMs. Extensive experiments on 13 image and video benchmarks across three leading MLLMs (Qwen2-VL, LLaVA-OneVision, and VILA1.5) demonstrate the remarkable effectiveness and strong generalization of our approach. Additionally, our new compression paradigm achieves faster inference with reductions in both prefilling time and KV cache memory.

cs.CV↗

arXiv2024

Stimulating Imagination: Towards General-purpose "Something Something Placement"

General-purpose object placement is a fundamental capability of an intelligent generalist robot: being capable of rearranging objects following precise human instructions even in novel environments. This work is dedicated to achieving general-purpose object placement with ``something something'' instructions. Specifically, we break the entire process down into three parts, including object localization, goal imagination and robot control, and propose a method named SPORT. SPORT leverages a pre-trained large vision model for broad semantic reasoning about objects, and learns a diffusion-based pose estimator to ensure physically-realistic results in 3D space. Only object types (movable or reference) are communicated between these two parts, which brings two benefits. One is that we can fully leverage the powerful ability of open-set object recognition and localization since no specific fine-tuning is needed for the robotic scenario. Moreover, the diffusion-based estimator only need to ``imagine" the object poses after the placement, while no necessity for their semantic information. Thus the training burden is greatly reduced and no massive training is required. The training data for the goal pose estimation is collected in simulation and annotated by using GPT-4. Experimental results demonstrate the effectiveness of our approach. SPORT can not only generate promising 3D goal poses for unseen simulated objects, but also be seamlessly applied to real-world settings.

cs.RO↗

arXiv2024

VideoTetris: Towards Compositional Text-to-Video Generation

Diffusion models have demonstrated great success in text-to-video (T2V) generation. However, existing methods may face challenges when handling complex (long) video generation scenarios that involve multiple objects or dynamic changes in object numbers. To address these limitations, we propose VideoTetris, a novel framework that enables compositional T2V generation. Specifically, we propose spatio-temporal compositional diffusion to precisely follow complex textual semantics by manipulating and composing the attention maps of denoising networks spatially and temporally. Moreover, we propose an enhanced video data preprocessing to enhance the training data regarding motion dynamics and prompt understanding, equipped with a new reference frame attention mechanism to improve the consistency of auto-regressive video generation. Extensive experiments demonstrate that our VideoTetris achieves impressive qualitative and quantitative results in compositional T2V generation. Code is available at: https://github.com/YangLing0818/VideoTetris

cs.CV↗

arXiv2022

Who is next: rising star prediction via diffusion of user interest in social networks

Finding items with potential to increase sales is of great importance in online market. In this paper, we propose to study this novel and practical problem: rising star prediction. We call these potential items Rising Star, which implies their ability to rise from low-turnover items to best-sellers in the future. Rising stars can be used to help with unfair recommendation in e-commerce platform, balance supply and demand to benefit the retailers and allocate marketing resources rationally. Although the study of rising star can bring great benefits, it also poses challenges to us. The sales trend of rising star fluctuates sharply in the short-term and exhibits more contingency caused by some external events (e.g., COVID-19 caused increasing purchase of the face mask) than other items, which cannot be solved by existing sales prediction methods. To address above challenges, in this paper, we observe that the presence of rising stars is closely correlated with the early diffusion of user interest in social networks, which is validated in the case of Taocode (an intermediary that diffuses user interest in Taobao). Thus, we propose a novel framework, RiseNet, to incorporate the user interest diffusion process with the item dynamic features to effectively predict rising stars. Specifically, we adopt a coupled mechanism to capture the dynamic interplay between items and user interest, and a special designed GNN based framework to quantify user interest. Our experimental results on large-scale real-world datasets provided by Taobao demonstrate the effectiveness of our proposed framework.

cs.SI↗

arXiv2021

How Powerful are Interest Diffusion on Purchasing Prediction: A Case Study of Taocode

A taocode is a kind of specially coded text-link on Taobao(the world's biggest online shopping website), through which users can share messages about products with each other. Analyzing taocodes can potentially facilitate understanding of the social relationships between users and, more excitingly, their online purchasing behaviors under the influence of taocode diffusion. This paper innovatively investigates the problem of online purchasing predictions from an information diffusion perspective, with taocode as a case study. Specifically, we conduct profound observational studies on a large-scale real-world dataset from Taobao, containing over 100M Taocode sharing records. Inspired by our observations, we propose InfNet, a dynamic GNN-based framework that models the information diffusion across Taocode. We then apply InfNet to item purchasing predictions. Extensive experiments on real-world datasets validate the effectiveness of InfNet compared with 8 state-of-the-art baselines.

cs.SI↗

arXiv2020

Exploiting Behavioral Consistence for Universal User Representation

User modeling is critical for developing personalized services in industry. A common way for user modeling is to learn user representations that can be distinguished by their interests or preferences. In this work, we focus on developing universal user representation model. The obtained universal representations are expected to contain rich information, and be applicable to various downstream applications without further modifications (e.g., user preference prediction and user profiling). Accordingly, we can be free from the heavy work of training task-specific models for every downstream task as in previous works. In specific, we propose Self-supervised User Modeling Network (SUMN) to encode behavior data into the universal representation. It includes two key components. The first one is a new learning objective, which guides the model to fully identify and preserve valuable user information under a self-supervised learning framework. The other one is a multi-hop aggregation layer, which benefits the model capacity in aggregating diverse behaviors. Extensive experiments on benchmark datasets show that our approach can outperform state-of-the-art unsupervised representation methods, and even compete with supervised ones.

cs.LG↗

arXiv2020

On the Uniqueness of Clifford Torus with Prescribed Isoperimetric Ratio

The Marques-Neves theorem asserts that among all the torodial (i.e. genus 1) closed surfaces, the Clifford torus has the minimal Willmore energy $\int H^2 \, dA$. % It is a natural conjecture that if one prescribes the isoperimetric Since the Willmore energy is invariant M{ö}bius transformations, it can be shown that there is a one-parameter family, up to homotheties, of genus 1 Willmore minimizers. It is then a natural conjecture that such a minimizer is unique if one prescribes its isoperimetric ratio. In this article, we show that this conjecture can be reduced to the positivity question of a polynomial recurrence.

math.DG↗

arXiv2020

Numerical Methods for Biomembranes: conforming subdivision methods versus non-conforming PL methods

The Canham-Helfrich-Evans models of biomembranes consist of a family of geometric constrained variational problems. In this article, we compare two classes of numerical methods for these variational problems based on piecewise linear (PL) and subdivision surfaces (SS). Since SS methods are based on spline approximation and can be viewed as higher order versions of PL methods, one may expect that the only difference between the two methods is in the accuracy order. In this paper, we prove that a numerical method based on minimizing any one of the `PL Willmore energies' proposed in the literature would fail to converge to a solution of the continuous problem, whereas a method based on minimization of the bona fide Willmore energy, well-defined for SS but not PL surfaces, succeeds. Motivated by this analysis, we propose also a regularization method for the PL method based on techniques from conformal geometry. We address a number of implementation issues crucial for the efficiency of our solver. A software package called Wmincon accompanies this article, provides parallel implementations of all the relevant geometric functionals. When combined with a standard constrained optimization solver, the geometric variational problems can then be solved numerically. To this end, we realize that some of the available optimization algorithms/solvers are capable of preserving symmetry, while others manage to break symmetry; we explore the consequences of this observation.

math.NA↗