👋

Hi, I'm

Jing

A study of test-time thinking in large vision-language models that shows more reasoning is not always better, then introduces uncertainty-guided lookback to improve visual grounding and decoding performance.

A memory-based approach for interactive action planning from egocentric demonstrations that models affordances to select contextually appropriate actions and detect deviations during task execution.

A language-driven framework for turning procedural text into coherent visual instructions using structured text encoding, discourse coherence modeling, and diffusion-based generation.

A comprehensive survey examining reasoning techniques in both textual and multimodal LLMs, addressing the challenge of integrating visual and textual inputs while resolving ambiguities across modalities.

The first benchmark explicitly designed to assess the reasoning path of MLLMs in visual reasoning tasks with novel metrics that assess reasoning fidelity beyond accuracy.

Novel attention mechanisms for improving visual perception in deep learning models with focus on spatial and temporal dynamics through attention head analysis.

A multimodal large language model capable of aligning audio-visual events with temporal intervals and text tokens. Built on PU-VALOR dataset with over 114,000 pseudo-untrimmed videos, AVicuna excels in temporal localization and time-aware dialogue capabilities for audio-visual understanding.

A comprehensive dataset and benchmark for evaluating multimodal large language models on object state captioning and state change representation. OSCaR consists of 14,084 annotated video segments with nearly 1,000 unique objects from various egocentric video collections, setting a new testbed for understanding dynamic environments and object state changes.

A video-based multimodal large language model fine-tuned for egocentric video content using comprehensive EAGLE-400K dataset comprising 400K visual instruction-tuning data from diverse sources.

A comprehensive survey examining the emergent capabilities of Vid-LLMs in video understanding, covering open-ended multi-granularity reasoning and categorizing approaches into three main types: Video Analyzer x LLM, Video Embedder x LLM, and (Analyzer + Embedder) x LLM.

An innovative augmented reality system that harnesses LLMs to assimilate information from visual, auditory, and contextual modalities, focusing on task performance quantification in AR through egocentric video, speech, and context analysis.

Novel approach combining Bayesian Inference and Model-based Imitation Learning to learn goal-directed action planning from instructional videos, capturing both long-term action associations and short-term action separations.

Hierarchical policy framework that addresses expert reaction delay in Learning from Interventions (LfI) through novel backtracking interpolation and sub-goal prediction for safe autonomous learning.