Xuan Hu

Xuhao Hu, Dongrui Liu, Hao Li et al.

Safety concerns of Multimodal large language models (MLLMs) have gradually become an important problem in various applications. Surprisingly, previous works indicate a counterintuitive phenomenon that using textual unlearning to align MLLMs achieves comparable safety performances with MLLMs aligned with image text pairs. To explain such a phenomenon, we discover a Visual Safety Information Leakage (VSIL) problem in existing multimodal safety benchmarks, i.e., the potentially risky content in the image has been revealed in the textual query. Thus, MLLMs can easily refuse these sensitive image-text pairs according to textual queries only, leading to unreliable cross-modality safety evaluation of MLLMs. We also conduct a further comparison experiment between textual alignment and multimodal alignment to highlight this drawback. To this end, we construct multimodal Visual Leakless Safety Bench (VLSBench) with 2.2k image-text pairs through an automated data pipeline. Experimental results indicate that VLSBench poses a significant challenge to both open-source and close-source MLLMs, e.g., LLaVA, Qwen2-VL and GPT-4o. Besides, we empirically compare textual and multimodal alignment methods on VLSBench and find that textual alignment is effective enough for multimodal safety scenarios with VSIL, while multimodal alignment is preferable for safety scenarios without VSIL. Code and data are released under https://github.com/AI45Lab/VLSBench

Jingyi Yang, Guanxu Chen, Xuhao Hu et al.

Masked diffusion language models (MDLMs) have recently emerged as a promising alternative to autoregressive (AR) language models, offering properties such as parallel decoding, flexible generation orders, and the potential for fewer inference steps. Despite these advantages, decoding strategies and reinforcement learning (RL) algorithms tailored for MDLMs remain underexplored. A naive approach is to directly transfer techniques well-established for AR models to MDLMs. However, this raises an immediate question: Is such a naive transfer truly optimal? For example, 1) Block-wise and semi-AR decoding strategies are not employed during the training of MDLMs, so why do they outperform full diffusion-style decoding during inference? 2) Applying RL algorithms designed for AR models directly to MDLMs exhibits a training-inference inconsistency, since MDLM decoding are non-causal (parallel). This results in inconsistencies between the rollout trajectory and the optimization trajectory. To address these challenges, we propose EOS Early Rejection (EOSER) and Ascending Step-Size (ASS) decoding scheduler, which unlock the potential of MDLMs to perform full diffusion-style decoding, achieving competitive performance with fewer decoding steps. Additionally, we introduce Consistency Trajectory Group Relative Policy Optimization (CJ-GRPO) for taming MDLMs, which emphasizes the consistency between rollout trajectory and optimization trajectory, and reduces the optimization errors caused by skip-step optimization. We conduct extensive experiments on reasoning tasks, such as mathematical and planning benchmarks, using LLaDA-8B-Instruct. The results demonstrate that the proposed EOSER and ASS mechanisms, together with CJ-GRPO, hold significant promise for effectively and efficiently taming MDLMs. Code: https://github.com/yjyddq/EOSER-ASS-RL.

Wesley H. Brigner, Naimul Hassan, Xuan Hu et al.

Complementary metal oxide semiconductor (CMOS) devices display volatile characteristics, and are not well suited for analog applications such as neuromorphic computing. Spintronic devices, on the other hand, exhibit both non-volatile and analog features, which are well-suited to neuromorphic computing. Consequently, these novel devices are at the forefront of beyond-CMOS artificial intelligence applications. However, a large quantity of these artificial neuromorphic devices still require the use of CMOS, which decreases the efficiency of the system. To resolve this, we have previously proposed a number of artificial neurons and synapses that do not require CMOS for operation. Although these devices are a significant improvement over previous renditions, their ability to enable neural network learning and recognition is limited by their intrinsic activation functions. This work proposes modifications to these spintronic neurons that enable configuration of the activation functions through control of the shape of a magnetic domain wall track. Linear and sigmoidal activation functions are demonstrated in this work, which can be extended through a similar approach to enable a wide variety of activation functions.

Wesley H. Brigner, Naimul Hassan, Xuan Hu et al.

Neuromorphic computing promises revolutionary improvements over conventional systems for applications that process unstructured information. To fully realize this potential, neuromorphic systems should exploit the biomimetic behavior of emerging nanodevices. In particular, exceptional opportunities are provided by the non-volatility and analog capabilities of spintronic devices. While spintronic devices have previously been proposed that emulate neurons and synapses, complementary metal-oxide-semiconductor (CMOS) devices are required to implement multilayer spintronic perceptron crossbars. This work therefore proposes a new spintronic neuron that enables purely spintronic multilayer perceptrons, eliminating the need for CMOS circuitry and simplifying fabrication.