multimodality

TypeSafe 发布 Jev 1.14.0 模型更新，并通过 "danger zone" 实验性功能引入图像支持、新的 box 边界框类型以及更长的上下文长度，面向开发者与创意用户开放预览。

Randal Balestrero announces his keynote talk at MICCAI-FOMO26 on recent progress in Joint-Embedding Predictive Architectures (JEPAs) and the need for more principled approaches to data imbalance, noise, multimodality, and reasoning.

This paper formalizes multimodality in generative behavioral cloning and finds, through theory and extensive experiments on LIBERO, Meta-World, and synthetic benchmarks, that common simulation benchmarks are largely unimodal and that simple deterministic action transformers remain competitive with expressive generative policies such as diffusion, flow, and CVAE-based methods.

The paper proposes a general asynchronous LLM framework that lets users define inference coroutines with overlapping memory states, demonstrating that Qwen3.x models can handle streaming video, video games, and monitoring tasks asynchronously without task-specific training.

The paper examines how vision-language models handle conflicts between in-context and parametric knowledge, revealing asymmetric biases where models favor in-context information for text entities but parametric information for image entities due to slower visual processing.

This paper investigates the use of conversational temporal dynamics (turn-pair timing) as a lightweight modality for automatic depression detection from dyadic clinical interviews, showing that a compact 24-dimensional timing module achieves strong performance and complements standard acoustic and semantic features when fused.