Reinforcement learning is increasingly used to align multimodal large language models (MLLMs). However, higher proxy rewards do not necessarily lead to better task performance, especially when visual evidence is difficult to verify with text-only or weakly grounded reward signals. We systematically study reward hacking in MLLM reinforcement learning by varying reward design, data ambiguity, model scale, and RL algorithm. To separate RL-induced failures from pre-existing weaknesses, we introduce the Newly Rewarded Failure Rate (NRFR), which measures the hacking rate among samples where RL achieves higher proxy reward than the SFT baseline. Our analysis yields three main findings. First, RL can actively create new failures rather than merely amplify existing ones, and the effect is especially pronounced near ambiguous decision boundaries. Second, robustness depends jointly on model scale, reward design, and RL algorithm, and scaling mitigates hacking but cannot replace reliable rewards. Third, incorporating visual evidence helps only when verification is reliable, since keyword-based verification introduces additional exploitation opportunities whereas VLM-based semantic verification can reduce reward hacking. Overall, multimodal reward hacking is a systematic consequence of optimizing imperfect reward signals, highlighting the need for rewards and verifiers that remain reliable under optimization pressure.

With the widespread popularisation of intelligent technology, task-based dialogue systems (TOD) are increasingly being applied to a wide variety of practical scenarios. As the key tasks in dialogue systems, named entity recognition and slot filling play a crucial role in the completeness and accuracy of information extraction. This paper is an evaluation paper for Sere-TOD 2022 Workshop challenge (Track 1 Information extraction from dialog transcripts). We proposed a multi-model fusion approach based on GlobalPointer, combined with some optimisation tricks, finally achieved an entity F1 of 60.73, an entity-slot-value triple F1 of 56, and an average F1 of 58.37, and got the highest score in SereTOD 2022 Workshop challenge

Learning social media content is the basis of many real-world applications, including information retrieval and recommendation systems, among others. In contrast with previous works that focus mainly on single modal or bi-modal learning, we propose to learn social media content by fusing jointly textual, acoustic, and visual information (JTAV). Effective strategies are proposed to extract fine-grained features of each modality, that is, attBiGRU and DCRNN. We also introduce cross-modal fusion and attentive pooling techniques to integrate multi-modal information comprehensively. Extensive experimental evaluation conducted on real-world datasets demonstrate our proposed model outperforms the state-of-the-art approaches by a large margin.