ProactiveBench: Benchmarking Proactiveness in Multimodal Large Language Models

A new benchmark called ProactiveBench measures whether multimodal large language models know when to request human help, finding that 22 tested models generally lack this capability and that larger models do not perform better [1]. The benchmark, built from seven repurposed datasets, tests proactiveness across tasks such as recognizing occluded objects, enhancing image quality, and interpreting coarse sketches [1]. Thomas De Min and colleagues submitted the work to the arXiv preprint repository, which hosts open-access e-prints in fields including computer science and has seen submission rates of about 24,000 articles per month as of late 2024 [1][6]. Evaluations across the 22 multimodal large language models revealed that proactiveness does not correlate with model capacity [1]. The researchers also found that providing conversation histories and in-context learning introduced negative biases that hindered performance, while hinting at proactiveness produced only marginal gains [1]. A simple fine-tuning strategy based on reinforcement learning suggested that proactiveness can be learned and may even generalize to unseen scenarios [1]. The authors publicly released ProactiveBench as a first step toward building proactive multimodal models, with the initial submission on March 19, 2026, weighing 9,278 KB and a revised version on June 26, 2026, at 9,299 KB [1]. The paper appears on arXiv, an open-access repository that began in 1991 and surpassed two million articles by the end of 2021 [6]. The platform supports community-built tools through arXivLabs, a framework launched in 2020 that allows collaborators to develop features such as bibliographic explorers and recommenders while adhering to values of openness and user data privacy [5].