ReasonX: MLLM-Guided Intrinsic Image Decomposition

The introduction of Omniguard presents a novel approach to AI safety moderation by enhancing the detection of harmful prompts across various languages and modalities, addressing the vulnerabilities of large language models (LLMs) to misuse. This method improves classification accuracy by 11.57% over existing baselines, marking a significant advancement in AI safety protocols.

The introduction of UniME-V2, a novel Universal Multimodal Embedding model, aims to enhance representation learning by leveraging the advanced capabilities of Multimodal Large Language Models (MLLMs). This model addresses limitations in existing approaches, particularly in capturing subtle semantic differences and improving the diversity of negative samples in embedding tasks.

A novel real-time stereo matching network, Generalized Geometry Encoding Volume (GGEV), has been proposed to enhance generalization in stereo matching applications, addressing the limitations of existing methods that focus primarily on in-domain performance. The GGEV employs depth-aware features and a Depth-aware Dynamic Cost Aggregation module to improve matching accuracy in unseen scenes.