The field of multimodal large language models (LLMs) is rapidly evolving, with a focus on improving reasoning capabilities, handling human annotation disagreements, and developing more inclusive NLP systems. Recent developments have highlighted the importance of adapting reinforcement learning to multimodal data and formats, addressing issues such as insufficient global context understanding and shortcut problems.

Researchers are exploring new paradigms for cooperative annotation, leveraging the strengths of both large and small language models to improve annotation accuracy and reduce costs. Additionally, there is a growing emphasis on capturing human annotation variations and disagreements, recognizing that these reflect important information such as task subjectivity and sample ambiguity.

Notably, innovative approaches include the use of multi-perspective methods, soft labels, and explainable AI to encourage the development of perspective-aware models that better approximate human label distributions. Furthermore, the introduction of new benchmarks and evaluation metrics is facilitating the assessment of model performance in complex multimodal reasoning tasks.

Some papers are particularly noteworthy, including GRPO-CARE, which proposes a consistency-aware RL framework to optimize both answer correctness and reasoning coherence, and AutoAnnotator, which presents a fully automatic annotation framework based on multi-model cooperative annotation. Commander-GPT is also notable for its modular decision routing framework inspired by military command theory, achieving state-of-the-art performance in multimodal sarcasm detection.