A new method called DVD proposes dynamic vector decoding to make MLLM-based perception more efficient by unifying 2D and 3D task representations into 1D vector sequences, avoiding text-based coordinate overhead and quantization…

#DVD #MLLM #Robotics #ComputerVision
https://arxiv.org/abs/2610.12266

Open-MMUnlearning offers an open-source framework that unifies target-model prep, multimodal data handling, unlearning, and evaluation across five benchmarks and eight MLLMs, aiming to bring consistency to fragmented MLLM unlearning…

#OpenSource #MLLM #AIResearch
https://arxiv.org/abs/2610.10358

OpenPhone proposes a device-cloud collaboration system for mobile GUI agents, pairing an enhanced 3B on-device model with cloud models to balance cost and capability. The approach uses two-stage SFT then GRPO training on…

#OpenSourceAI #MobileAI #GUIAgents #MLLM
https://arxiv.org/abs/2510.22009

Researchers introduced PI3D, a prompt injection attack on multimodal LLMs that works by placing text-bearing objects in 3D environments rather than editing 2D images, highlighting a new risk for robotics and situated…

#MLLM #PromptInjection #Robotics #AIsecurity
https://arxiv.org/abs/2602.07104

SONIC-O1 introduces a 60-hour, 13-domain benchmark for evaluating multimodal LLMs on real-world audio-video understanding, covering summarization, MCQ answering, and temporal localization. Results show MCQ accuracy gaps…

#SONICO1 #MLLM #AudioVideo #Benchmark
https://arxiv.org/abs/2601.21666