Yuhang Wang

Yuhang Wang, Linlin Zhang, Haoxuan Ji et al.

Multimodal large language model (MLLM) unlearning methods have been proposed to remove private, sensitive, or proprietary information from well-trained models. However, most existing MLLM unlearning methods are designed for one-shot requests and fail to adequately address continual scenarios, as repeatedly applying one-shot operations leads to cumulative utility degradation, unlearning rebound, and retention drift. We introduce Merging for Continual Unlearning (MCU), an approach that dynamically merges multiple one-shot unlearning adapters into a unified adapter upon receiving each new unlearning request.Through a leave-one-out merging analysis, we reveal that these unlearning adapters exhibit strong cross-task dependencies. Such dependencies have two contrasting effects: they can facilitate cross-task unlearning transferability, but they can also introduce severe interference that degrades unlearning effectiveness and compromises retained knowledge. To address this challenge, MCU projects the adapters into a shared representation space, preserves their dominant directions, suppresses over-concentrated coordinates, and reconfigures cross-task dependencies to mitigate interference while enhancing transferability. Experiments on ICU-Bench and MLLMU-Bench demonstrate that MCU achieves superior unlearning effectiveness while preserving both retained knowledge and general multimodal utility.

Hao Zhou, Shengming Yuan, Yuhang Wang et al.

This study presents Scale-CDA, an open-hardware/open-software tool-chain that democratizes a functional version of Generative-AI-assisted Cooperative Driving Automation (CDA). Built on the community-maintained OpenDBC interface (300+ car models) and Openpilot Level-2 ADAS, Scale-CDA achieves plug-and-play retrofitting with off-the-shelf parts that cost under US \$1,000 (edge PC, webcam, CAN adapter, optional LTE/Wi-Fi radios). A lightweight Vehicle-to-Everything (V2X) stack using MQTT over Wi-Fi 6/LTE provides bidirectional connectivity. Field experiments in a 7.5 km test loop demonstrated mean round-trip latency of 5.25 ms and link speeds near 100 Mb/s, validating Wi-Fi 6 as a viable, low-cost medium for non-safety-critical CDA messaging . At the intelligence layer, an edge-deployed multimodal large-language model (MLLM) ingests synchronized vision, CAN, and V2X streams via a Model-Context-Protocol (MCP) bridge, then issues structured JSON advisories and motion primitives. A library of meta-action executors translates these high-level commands into verified Openpilot planner hooks, enabling lane changes, gap management, and emergency stops without altering the safety-certified core. In multi-vehicle road tests the full stack maintained end-to-end decision latencies below 60 ms, while preserving data privacy by keeping inference on-board. Collectively, Scale-CDA closes two critical gaps that have limited CDA R\&D: (i) affordable, interoperable hardware for large-scale field trials, and (ii) a standardized interface that lets GenAI reason and act on everyday cars. By releasing the bills-of-materials, connectivity APIs, and GenAI bridges as open resources, this work offers transportation agencies and researchers a practical blueprint for democratizing cooperative autonomy, accelerating deployments that enhance traffic efficiency and safety.