Large vision-language models (LVLMs) have made rapid advancements; however, existing benchmarks for evaluating their instruction-following abilities focus on purely textual constraints, e.g., format constraints and role-playing, which assess only language-level adherence and are not related to vision. To bridge this gap, we introduce VC-IF, a benchmark that involves visual constraints connecting instructions to explicit operations over image content, thereby directly probing visual spatial understanding and compositional compliance. VC-IF defines atomic constraints, systematically combining mixed settings, scaling from single- to multi-constraint scenarios, and single- and multi-turn scenarios. We evaluate 16 popular LVLMs and find that models performing well under text-only constraints may struggle when faced with visual constraints, highlighting significant gaps in their capabilities. Further, we conduct scaling analyses across varying numbers of constraints and dialogue turns, revealing performance trends as constraint complexity increases along different dimensions. Finally, we synthesize a large training dataset and propose VC-PO, a reinforcement learning strategy that uses a hybrid verifier mechanism to compute reward signals for multiple constraints. Applying VC-PO to Qwen3-VL-8B-Instruct improves its overall VC-IF score from 43.74 to 55.18 (+11.44). The VC-PO-trained model remains stable as the number of turns and constraints increases. Our data and code are available at VC-IF.

Language models trained on large-scale corpus often generate harmful responses that are harmful and contrary to human values. A prevalent approach for human alignment is reinforcement learning from human feedback (RLHF), utilizing algorithms such as proximal policy optimization (PPO). However, these methods are often characterized by complexity, instability, and substantial resource consumption. Considering that existing large language models (LLMs) like ChatGPT are already relatively well-aligned and cost-friendly, researchers propose to align the language model with human preferences from AI feedback. Nevertheless, the common practices, that unidirectionally distill the responses, are constrained by the inherent capability of LLMs. To address it, we introduce CycleAlign, a framework that distills alignment capabilities from the parameter-invisible LLMs (black-box) to the parameter-visible models (white-box) in an iterative manner. CycleAlign iteratively improves both the white-box and black-box models by integrating static and dynamic in-context learning and a belief alignment method.Empirical results illustrate that the model fine-tuned by CycleAlign remarkably exceeds existing methods, and achieves the state-of-the-art performance in alignment with human value.