Chronological feed of everything captured from Arvind on AI.
Infrastructure-centric world models (I-WM) leverage fixed roadside sensors' temporal depth for accumulating long-term behavioral distributions, including rare events, complementing ego-vehicle sensors' spatial breadth. The proposed three-phase approach includes generative scene understanding with uncertainty propagation, physics-informed multi-agent predictive dynamics, and collaborative V2X models via latent alignment. A dual-layer architecture uses annotation-free multi-modal perception (LiDAR to event cameras) to drive end-to-end generative models, introducing Infrastructure VLA (I-VLA) unifying perception, language, and traffic actions.
TeamFusion is a multi-agent system for open-ended domains that instantiates proxy agents from team members' preferences, facilitates structured discussions to identify agreements and disagreements, and synthesizes consensus-oriented outputs for iterative refinement. Unlike closed-domain aggregation methods, it preserves minority perspectives by resolving disagreements rather than suppressing them. Evaluations on two teamwork tasks show superior performance over direct aggregation baselines in representing individual views and achieving consensual strength across metrics, tasks, and team configurations.
Wan-Image integrates large language models with diffusion transformers in a unified multi-modal architecture to enable precise control over image generation, addressing limitations in controllability, typography, and identity preservation. It leverages large-scale multi-modal data, fine-grained annotations, and reinforcement learning to support advanced features like ultra-long text rendering, multi-subject identity preservation, 4K synthesis, and interactive editing. Human evaluations show it outperforms Seedream 5.0 Lite and GPT Image 1.5 overall, matching Nano Banana Pro on challenging tasks.
LLaDA2.0-Uni integrates multimodal understanding and generation using a discrete diffusion large language model with SigLIP-VQ tokenization, MoE-based dLLM backbone, and diffusion decoder. It applies block-level masked diffusion to both text and discretized visual inputs, enabling interleaved reasoning and generation. Prefix-aware backbone optimizations and few-step decoder distillation boost inference efficiency, matching specialized VLMs in understanding while excelling in high-fidelity image generation and editing.
Fine-tuning Gemma 3 27B on modest human-labeled Google Street View data aligns model predictions with human mean opinion scores (MOS), surpassing individual raters on SRCC and PLCC metrics. Knowledge distillation transfers performance to Gemma 3 4B (3x speedup) and vision models like EfficientNetV2-M and SwinV2-B (30x speedup) with comparable accuracy. The framework supports scalable assessment of built environment attributes via a visualization dashboard, minimizing labeling needs.