The field of vision-language models is moving towards more comprehensive and accurate evaluation protocols, with a focus on true visual reasoning and multimodal understanding. Researchers are exploring new benchmarks and frameworks that can effectively assess the capabilities of these models, such as evaluating their ability to generate detailed and accurate descriptions of videos and images. Noteworthy papers in this area include: Evaluating Multimodal Large Language Models on Video Captioning via Monte Carlo Tree Search, which proposes an automatic framework for constructing diverse descriptive sentences for video captioning. EasyARC, a vision-language benchmark that requires multi-image, multi-step reasoning and self-correction, setting a new standard for evaluating true reasoning and test-time scaling capabilities in vision-language models. VL-GenRM, an iterative training framework that leverages vision experts and Chain-of-Thought rationales to enhance vision-language verification. PeRL, a general reinforcement learning approach tailored for interleaved multimodal tasks, which achieves state-of-the-art performance on multi-image benchmarks. video-SALMONN 2, an advanced audio-visual large language model that significantly enhances video captioning accuracy through directed preference optimization. DiscoSG, a new task and dataset for discourse-level text scene graph parsing, which improves downstream vision-language tasks like discourse-level caption evaluation and hallucination detection. Dual-Stage Value-Guided Inference with Margin-Based Reward Adjustment, a two-stage inference framework that improves both efficiency and output fidelity by combining a temporal-difference value model with a margin-aware reward adjustment.