Developed a crossmodal poetry generation system using large models, with functions including high-quality poetry generation, analysis and scoring, and optimization. The whole process all provides user-friendly interpretation for better user experience, e.g., allusion annotation and appreciation. The system firstly utilized CLIP, miniGPT4 and ERNIE (paper accepted), and finally swichted to DeepSeek’s series models for better performance.

Trained RNNs on the same-different task under varying noise levels by neurogym, optimizing the code for readability and extensibility. Analyzed normalized averages and principal components of RNN hidden states, performed linear fitting of activities at different time points to stimuli values, and analyzed the temporal scope.

Explored various effects of parameter settings on the working memory capacity of RNNs.

Pre-processed word data for forward flow tasks, inserting seed words, removing duplicates, and generating embeddings. Analyzed the correlation between participants’ scale scores and statistical indicators.

An interactive poetry learning assistant based on LLM for different ages supports different chatting styles, oneline search, RAG, and image generation.