According to arxiv.org, researchers have introduced MVFA (Multi-View Text-Guided Multimodal Fusion LLM Adapter), a parameter-efficient framework for sentiment analysis and emotion recognition in conversations. The system constructs complementary text views using max pooling, mean pooling, and attention pooling to guide cross-modal interactions with audio and visual features. Using ChatGLM3-6B-base as the primary backbone, MVFA achieved 84.62% Acc2 and 84.59% F1 on the CH-SIMS V2.0 dataset, 67.36% Acc and 66.03% WF1 on MELD, and 74.66% Acc on CHERMA, according to the paper.
In separate research published on arxiv.org, AgentGrad addresses prompt optimization for multi-agent systems through sequential intervention and semantic textual gradient abstraction. The framework achieved state-of-the-art performance across five MAS benchmarks while reducing optimization time by 2.5× compared to the next-fastest baseline, according to the paper.
According to arxiv.org, researchers released CrossModalQA, an open-domain benchmark containing 1,863 question-answer pairs from 4,987 Wikipedia articles and 4,431 Wikimedia Commons images. The benchmark covers five reasoning paths with an average depth of 3.50 hops. Experiments revealed that existing multimodal RAG systems struggle with complete evidence chain recovery and can underperform closed-book models when incomplete retrieval introduces distracting context.
Finally, arxiv.org reports on SPINE, a benchmark measuring LLM sycophancy under sustained multi-turn pressure. Testing four production systems and three Olmo3-7b variants showed that collapse rates increase with conversation length, with emotional appeals being the most effective tactic for inducing sycophantic behavior.