The most recent research uses reinforcement learning (RL) to post-train Multi-modal Large Language Models (MLLMs) such that these models are able to iteratively call search engines to dynamically access external knowledge when handling complex Visual Question Answering (VQA) tasks. However, existing methods face two major limitations in effectiveness and efficiency: i) For effectiveness, the objective of these methods, which only considers the correctness of the generated final response, overlooks the quality of intermediate search results, thus leading to suboptimal search strategies. ii) For efficiency, existing methods often unnecessarily invoke search calls during reasoning, making the inference inefficient. To address these issues, we propose , a customized dual-objective reinforcement learning framework to improve the search strategies of MLLMs, enhancing their search quality yet minimizing search frequency. The key ideas include (1) a reward function that promotes correct reasoning trajectories with fewer search calls; and (2) dual optimization objectives that jointly optimize search quality and answer correctness. Extensive experiments on 3 real-world datasets demonstrate that DORA outperforms state-of-the-art methods, achieving up to 8.4% higher accuracy while reducing the number of search calls by 9.7%.

Despite detection of suicidal ideation on social media has made great progress in recent years, people’s implicitly and anti-real contrarily expressed posts still remain as an obstacle, constraining the detectors to acquire higher satisfactory performance. Enlightened by the hidden “tree holes” phenomenon on microblog, where people at suicide risk tend to disclose their inner real feelings and thoughts to the microblog space whose authors have committed suicide, we explore the use of tree holes to enhance microblog-based suicide risk detection from the following two perspectives. (1) We build suicide-oriented word embeddings based on tree hole contents to strength the sensibility of suicide-related lexicons and context based on tree hole contents. (2) A two-layered attention mechanism is deployed to grasp intermittently changing points from individual’s open blog streams, revealing one’s inner emotional world more or less. Our experimental results show that with suicide-oriented word embeddings and attention, microblog-based suicide risk detection can achieve over 91% accuracy. A large-scale well-labelled suicide data set is also reported in the paper.