Recent developments in artificial intelligence benchmarking highlight the introduction of several new evaluation frameworks aimed at enhancing the assessment of language and vision models. Notably, the GameTraversalBenchmark has been launched to evaluate how well large language models (LLMs) can navigate generated 2D game environments, with findings indicating that larger models perform better. Additionally, the Holistic Evaluation of Vision Language Models (VHELM) extends the HELM framework to assess VLMs, while the newly released MMIE benchmark focuses on multimodal comprehension and generation, featuring over 20,000 examples across 12 fields. The MEGA-Bench framework has also been introduced, encompassing 505 real-world tasks and over 8,000 samples, designed to scale multimodal evaluation effectively. These frameworks aim to address vulnerabilities in existing benchmarks and improve the integrity of AI assessments, particularly in light of recent findings that LLMs can achieve high scores on benchmarks through potentially misleading methods, such as using a “null model” that outputs constant responses. This raises concerns about the reliability of current evaluation metrics in the AI research community.