Despite remarkable capability in multi-modal understanding, deploying Multi-modal Large Language Models (MLLMs) in open-ended conversational scenarios introduces safety risks that remain poorly addressed by existing alignment methods. Unlike simple malicious visual question and answer (VQA) pairs , multi-turn interactions enable adversaries to incrementally reconstruct harmful intent across dialogues, progressively bypassing safety constraints in ways that are difficult to detect at any individual turn. Meanwhile, conventional reinforcement learning from human feedback (RLHF) approaches are unsuitable for this situation: designed primarily for VQA tasks, they neither capture cross-turn risk dynamics nor scale efficiently without costly manual preference annotation. To close this gap, we introduce MINT-Safe, an open-source visual multi-turn training dataset comprising 11,270 multi-image dialogues and 500 refusal VQA pairs, constructed via multi-agent interaction with text-to-image (T2I) tool-call augmentation. Building on MINT-Safe, we propose TAD-Align, a dialogue safety alignment framework centered on a turn-aware dual-objective reward function. Rather than treating all dialogue turns uniformly, TAD-Align leverages rollout-based safety score variance to dynamically identify turns where the model exhibits inconsistent safety behavior, and adaptively up-weights these turns during optimization. Experiments on Qwen2.5-VL-7B-Instruct and LLaVA-NeXT-7B demonstrate reductions of over 10% in Attack Success Rate (ASR), alongside improvements of at least 8% in harmlessness and 13% in helpfulness on multi-modal multi-turn safety benchmarks, while preserving general model capabilities.

Large language models (LLMs) have revolutionized various domains with their remarkable capabilities, but their massive parameter sizes pose significant challenges for fine-tuning and inference, especially in resource-constrained environments. Conventional compression methods often result in substantial performance degradation within LLMs and struggle to restore model quality during fine-tuning. To address this challenge, we present Bayesian Knowledge Distillation (BayesKD), a novel distillation framework meticulously designed for compact LLMs in resource-constrained fine-tuning scenarios. Departing from conventional LLM distillation methods that introduce time-consuming paradigms and fail to generalize in compressed LLM fine-tuning scenarios, our BayesKD develops the Logits Dual-Scaling, Knowledge Alignment Module, and Bayesian Distillation Optimization. In particular, our Logits Dual-Scaling strategy adaptively aligns the strength of the teacher’s knowledge transfer, while the Knowledge Alignment Module bridges the gap between the teacher and student models by projecting their knowledge representations into a shared interval. Additionally, we employ Logits-Aware Bayesian Optimization to swiftly identify optimal settings based on these strategies, thereby enhancing model performance. Extensive experiments across diverse tasks demonstrate that BayesKD consistently outperforms baseline methods on various state-of-the-art LLMs, including LLaMA, Qwen2, Bloom, and Vicuna. Notably, our BayesKD achieves average accuracy gains of 2.99% and 4.05% over standard KD for the 8B parameter LLaMA and Qwen2 model. Codes are available in the supplementary materials.