Ruofan Wang

Borja Ocejo, Ruofan Wang, Ke Liu et al.

Notifications are an important communication channel for delivering timely and relevant information. Optimizing their delivery involves addressing complex sequential decision-making challenges under constraints such as message utility and user fatigue. Offline reinforcement learning (RL) methods, such as Conservative Q-Learning (CQL), have been applied to this problem but face practical challenges at scale, including instability, sensitivity to distribution shifts, limited reproducibility, and difficulties with explainability in high-dimensional recommendation settings. We present a Decision Transformer (DT) based framework that reframes policy learning as return-conditioned supervised learning, improving robustness, scalability, and modeling flexibility. Our contributions include a real-world comparison with CQL, a multi-reward design suitable for non-episodic tasks, a quantile regression approach to return-to-go conditioning, and a production-ready system with circular buffer-based sequence processing for near-real-time inference. Extensive offline and online experiments in a deployed notification system show that our approach improves notification utility and overall session activity while minimizing user fatigue. Compared to a multi-objective CQL-based agent, the DT-based approach achieved a +0.72% increase in sessions for notification decision-making at LinkedIn by making notification recommendation more relevant.

Ruofan Wang, Xiang Zheng, Xiaosen Wang et al.

Vision-Language Models (VLMs) are vulnerable to jailbreak attacks, where adversaries bypass safety mechanisms to elicit harmful outputs. In this work, we examine an insidious variant of this threat: toxic continuation. Unlike standard jailbreaks that rely solely on malicious instructions, toxic continuation arises when the model is given a malicious input alongside a partial toxic output, resulting in harmful completions. This vulnerability poses a unique challenge in multimodal settings, where even subtle image variations can disproportionately affect the model's response. To this end, we propose RedDiffuser (RedDiff), the first red teaming framework that uses reinforcement learning to fine-tune diffusion models into generating natural-looking adversarial images that induce toxic continuations. RedDiffuser integrates a greedy search procedure for selecting candidate image prompts with reinforcement fine-tuning that jointly promotes toxic output and semantic coherence. Experiments demonstrate that RedDiffuser significantly increases the toxicity rate in LLaVA outputs by 10.69% and 8.91% on the original and hold-out sets, respectively. It also exhibits strong transferability, increasing toxicity rates on Gemini by 5.1% and on LLaMA-Vision by 26.83%. These findings uncover a cross-modal toxicity amplification vulnerability in current VLM alignment, highlighting the need for robust multimodal red teaming. We will release the RedDiffuser codebase to support future research.