Proposed LGBench, the first large-scale benchmark for long-horizon multi-goal image generation and editing, featuring 2,000 tasks and 29,000+ structured goals. Developed VisionDirector, a training-free VLM-driven closed-loop framework. Applied GRPO to reduce editing rounds by 26%.

We present TraveLLaMA, a specialized multimodal language model for comprehensive travel assistance. Our contributions include: (1) TravelQA, a novel dataset of 265k QA pairs combining 160k text QA, 100k vision-language QA, and 5k expert-annotated CoT reasoning examples; (2) Travel-CoT, a structured reasoning framework that decomposes travel queries into spatial, temporal, and practical dimensions, improving accuracy by 10.8%; and (3) an interactive agent system achieving a SUS score of 82.5 in user studies with 500 participants.

We present VRBench, the first long narrative video benchmark for evaluating multi-step reasoning capabilities. It comprises 960 long videos (avg. 1.6 hours), 8,243 human-labeled multi-step QA pairs, and 25,106 reasoning steps with timestamps. We develop a human-AI collaborative framework generating coherent reasoning chains spanning seven types. VRBench designs a multi-phase evaluation pipeline assessing models at both outcome and process levels.

Proposed a graph-based memory agent that tracks entities and relations across long videos. Developed a temporal reasoning framework achieving 56.3% / 73.3% accuracy while analyzing only 8.2 frames on average, significantly improving efficiency.

Constructed GeoText-1652, an image–text–bounding box benchmark for drone-view vision-language understanding. Proposed a spatial-aware image–text matching method, outperforming ALBEF and X-VLM. Achieved a 31.2% Recall@10, a significant improvement over prior baselines.

Developed a teleoperation system integrating multisensory fusion with haptic and visual feedback under IoT framework. Published in IEEE Internet of Things Journal with Impact Factor 8.2.

Capybara is a unified visual creation framework for high-quality image and video generation as well as instruction-based editing. It supports text-to-image, text-to-video, image editing, and video editing within one open-source system.