Selected work

Representative Papers

VAnim: Rendering-Aware Sparse State Modeling for Structure-Preserving Vector Animation

May 02, 2026

Existing methods for text-to-SVG animation generation struggle to simultaneously preserve topological consistency and model non-rigid deformations, while also lacking support for open-domain instructions. This work proposes VAnim—the first large language model framework tailored for open-domain, text-driven SVG animation generation—which formulates animation as sparse state updates over a persistent SVG DOM tree, drastically reducing sequence length by more than 9.8×. By integrating an identity-aware motion planning mechanism and rendering-aware reinforcement learning (GRPO) with a video-perceptual hybrid reward, VAnim achieves high-fidelity dynamic generation while maintaining structural validity and identity consistency. We further introduce SVGAnim-134k, the first vector animation benchmark, and demonstrate through experiments that VAnim significantly outperforms existing approaches in semantic alignment, motion quality, and structural preservation.

0 citationsRead paper

EventMemAgent: Hierarchical Event-Centric Memory for Online Video Understanding with Adaptive Tool Use

Feb 16, 2026

This work addresses the challenge in online video understanding posed by the tension between the unbounded visual stream and the limited context window of multimodal large language models, which hinders simultaneous modeling of long-range dependencies and fine-grained details. To this end, the authors propose EventMemAgent, an active agent framework grounded in hierarchical event-centric memory. It constructs short-term memory through event boundary detection and reservoir sampling, structures long-term memory into a searchable archive, and integrates a multi-granularity perception toolkit. Furthermore, agentic reinforcement learning is introduced to internalize tool invocation and reasoning as part of the agent’s policy. The method achieves competitive performance across multiple online video understanding benchmarks and represents the first end-to-end joint optimization of hierarchical event-granular memory and active perception.

0 citationsRead paper

TwiFF (Think With Future Frames): A Large-Scale Dataset for Dynamic Visual Reasoning

Feb 11, 2026

Existing visual chain-of-thought methods struggle with temporal reasoning in dynamic scenes. To address this challenge, this work proposes the TwiFF framework and introduces TwiFF-2.7M, the first large-scale temporally aligned dataset for dynamic visual chain-of-thought reasoning. TwiFF integrates pretrained video generation and image understanding capabilities, enabling a synergistic vision-language temporal reasoning mechanism by iteratively generating future action frames while simultaneously performing textual reasoning. Experimental results demonstrate that TwiFF significantly outperforms both existing visual and purely text-based chain-of-thought approaches on dynamic visual question answering tasks, thereby validating its effectiveness and state-of-the-art performance.

0 citationsRead paper

StableMotion: Repurposing Diffusion-Based Image Priors for Motion Estimation

May 10, 2025

This paper addresses single-image motion estimation—specifically for image stitching rectification and rolling-shutter correction—by proposing a geometry-content prior transfer method leveraging a pre-trained text-to-image diffusion model (Stable Diffusion). Methodologically, it (1) reparameterizes the diffusion model as a mapping from latent space to motion fields; (2) introduces an Adaptive Ensemble Strategy (AES) to enhance consistency across multiple sampling trajectories; and (3) identifies and exploits the “Sampling-Step Disaster” (SSD) phenomenon—wherein motion-field accuracy peaks at a single denoising step—enabling one-step, high-fidelity prediction and accelerating inference by 200×. The approach achieves state-of-the-art performance on both tasks, demonstrating strong generalization, sub-pixel accuracy, and real-time capability without task-specific architectural modifications or fine-tuning.

0 citationsRead paper

FeatInsight: An Online ML Feature Management System on 4Paradigm Sage-Studio Platform

Apr 01, 2025

In online machine learning, feature management faces critical challenges including high latency, severe redundancy, difficulty in real-time updates, and inconsistent cross-source dependency resolution. To address these, we propose the first unified feature management framework that integrates feature intelligence with millisecond-scale dynamic updates, enabling consistent real-time computation over trillion-dimensional feature spaces and complex cross-table dependencies. Built upon OpenMLDB, the framework incorporates a declarative feature DSL, incremental lineage tracking, visual diagnostics, and automated validation—covering the full feature lifecycle from design and computation to verification, lineage analysis, and visualization. Deployed across 100+ production scenarios, it reduces feature update latency to the millisecond level, cuts end-to-end latency of sales forecasting services by 70%, and significantly improves model performance and iteration efficiency for recommendation and fraud detection systems.

0 citationsRead paper

Recent publications

Latest Papers

VAnim: Rendering-Aware Sparse State Modeling for Structure-Preserving Vector Animation

May 02, 2026

Existing methods for text-to-SVG animation generation struggle to simultaneously preserve topological consistency and model non-rigid deformations, while also lacking support for open-domain instructions. This work proposes VAnim—the first large language model framework tailored for open-domain, text-driven SVG animation generation—which formulates animation as sparse state updates over a persistent SVG DOM tree, drastically reducing sequence length by more than 9.8×. By integrating an identity-aware motion planning mechanism and rendering-aware reinforcement learning (GRPO) with a video-perceptual hybrid reward, VAnim achieves high-fidelity dynamic generation while maintaining structural validity and identity consistency. We further introduce SVGAnim-134k, the first vector animation benchmark, and demonstrate through experiments that VAnim significantly outperforms existing approaches in semantic alignment, motion quality, and structural preservation.

0 citationsRead paper

EventMemAgent: Hierarchical Event-Centric Memory for Online Video Understanding with Adaptive Tool Use

Feb 16, 2026

This work addresses the challenge in online video understanding posed by the tension between the unbounded visual stream and the limited context window of multimodal large language models, which hinders simultaneous modeling of long-range dependencies and fine-grained details. To this end, the authors propose EventMemAgent, an active agent framework grounded in hierarchical event-centric memory. It constructs short-term memory through event boundary detection and reservoir sampling, structures long-term memory into a searchable archive, and integrates a multi-granularity perception toolkit. Furthermore, agentic reinforcement learning is introduced to internalize tool invocation and reasoning as part of the agent’s policy. The method achieves competitive performance across multiple online video understanding benchmarks and represents the first end-to-end joint optimization of hierarchical event-granular memory and active perception.

0 citationsRead paper

TwiFF (Think With Future Frames): A Large-Scale Dataset for Dynamic Visual Reasoning

Feb 11, 2026

Existing visual chain-of-thought methods struggle with temporal reasoning in dynamic scenes. To address this challenge, this work proposes the TwiFF framework and introduces TwiFF-2.7M, the first large-scale temporally aligned dataset for dynamic visual chain-of-thought reasoning. TwiFF integrates pretrained video generation and image understanding capabilities, enabling a synergistic vision-language temporal reasoning mechanism by iteratively generating future action frames while simultaneously performing textual reasoning. Experimental results demonstrate that TwiFF significantly outperforms both existing visual and purely text-based chain-of-thought approaches on dynamic visual question answering tasks, thereby validating its effectiveness and state-of-the-art performance.

0 citationsRead paper

StableMotion: Repurposing Diffusion-Based Image Priors for Motion Estimation

May 10, 2025

This paper addresses single-image motion estimation—specifically for image stitching rectification and rolling-shutter correction—by proposing a geometry-content prior transfer method leveraging a pre-trained text-to-image diffusion model (Stable Diffusion). Methodologically, it (1) reparameterizes the diffusion model as a mapping from latent space to motion fields; (2) introduces an Adaptive Ensemble Strategy (AES) to enhance consistency across multiple sampling trajectories; and (3) identifies and exploits the “Sampling-Step Disaster” (SSD) phenomenon—wherein motion-field accuracy peaks at a single denoising step—enabling one-step, high-fidelity prediction and accelerating inference by 200×. The approach achieves state-of-the-art performance on both tasks, demonstrating strong generalization, sub-pixel accuracy, and real-time capability without task-specific architectural modifications or fine-tuning.

0 citationsRead paper

FeatInsight: An Online ML Feature Management System on 4Paradigm Sage-Studio Platform

Apr 01, 2025

In online machine learning, feature management faces critical challenges including high latency, severe redundancy, difficulty in real-time updates, and inconsistent cross-source dependency resolution. To address these, we propose the first unified feature management framework that integrates feature intelligence with millisecond-scale dynamic updates, enabling consistent real-time computation over trillion-dimensional feature spaces and complex cross-table dependencies. Built upon OpenMLDB, the framework incorporates a declarative feature DSL, incremental lineage tracking, visual diagnostics, and automated validation—covering the full feature lifecycle from design and computation to verification, lineage analysis, and visualization. Deployed across 100+ production scenarios, it reduces feature update latency to the millisecond level, cuts end-to-end latency of sales forecasting services by 70%, and significantly improves model performance and iteration efficiency for recommendation and fraud detection systems.

0 citationsRead paper