Large Language Models (LLMs) have demonstrated strong reasoning capabilities, yet achieving nuanced personality alignment for simulating human conversation remains challenging. This difficulty stems from two bottlenecks: the scarcity of scalable, privacy-respecting training data and the limited ability of conventional reinforcement learning (RL) rewards to capture subtle personality-relevant signals. In this work, we introduce a unified framework to address these issues. First, we propose the Personality Alignment Choice Forge (PACF), an automatic pipeline that transforms existing conversational datasets into structured, multiple-choice benchmarks, constraining the model’s action space and enabling scalable evaluation grounded in observed human choices. Second, we present Group Sentence-Specified Sequence Policy Optimization (GSSPO), an RL algorithm designed to address the coarse credit assignment problem in methods that apply a uniform reward across all tokens. By integrating stable sequence-level updates with differentiated, rubric-guided token-level feedback, GSSPO assigns reward signals to more localized spans of the response. This targeted approach mitigates optimization instabilities caused by undifferentiated reward signals, leading to more effective training in our experiments. On the Qwen3 model family, GSSPO achieves the best personality-alignment accuracy on every benchmark and scale in the SFT+RL regime, and leads the other RL methods on five of six in the SFT-Free regime. These results suggest a scalable route toward training LLMs that better match measured personality-alignment preferences. Our core code is available at https://github.com/HKUST-FintechLab/GSSPO

The International Classification of Diseases (ICD) is the foundation of global health statistics and epidemiology. The ICD is designed to translate health conditions into alphanumeric codes. A number of approaches have been proposed for automatic ICD coding, since manual coding is labor-intensive and there is a global shortage of healthcare workers. However, existing studies did not exploit the discourse structure of clinical notes, which provides rich contextual information for code assignment. In this paper, we exploit the discourse structure by leveraging section type classification and section type embeddings. We also focus on the class-imbalanced problem and the heterogeneous writing style between clinical notes and ICD code definitions. The proposed reconciled embedding approach is able to tackle them simultaneously. Experimental results on the MIMIC dataset show that our model outperforms all previous state-of-the-art models by a large margin. The source code is available at https://github.com/discnet2022/discnet