In the RAG paradigm, document ranking determines the evidence available to downstream generators. Through controlled analysis, we identify two phenomena underexplored by existing rankers: (i) downstream response quality depends not only on relevance but also on the composition and ordering of selected documents, and (ii) such preferences differ systematically across generators. However, existing rankers are trained purely on query–document relevance, leaving both phenomena unmodeled. To close this gap, we construct PRISM, a bilingual preference-aligned dataset built through a four-stage pipeline that compresses the combinatorial subset-and-ordering space by roughly four orders of magnitude and produces response-quality preference supervision conditioned on seven downstream generators. On a 13k-query subset of PRISM, we train Rank4Gen, a generator-aware ranker that performs joint document set selection and ordering. Experiments on five challenging RAG benchmarks show that Rank4Gen improves downstream QA quality on most evaluated generators, with per-generator F1 gains of up to +2.08 over the strongest set-selection baseline. Code is available at https://github.com/JOHNNY-fans/Rank4Gen.

Training large language models for domain adaptation poses a significant challenge in balancing the acquisition of domain knowledge with the retention of general abilities, often leading to catastrophic forgetting. While curriculum learning offers a promising direction, conventional methods typically rely on a single dimension of knowledge or task, which is insufficient to navigate the trade-off between knowledge breadth and task depth. In this paper, we propose a two-dimensional curriculum learning framework that coordinates model training along two orthogonal axes: the knowledge dimension and the task dimension. We first reconstruct the dataset by clustering instances according to their semantic similarity to general-domain data, and subsequently annotate them with a task hierarchy. Then, we design an integrated curriculum that develops from general to domain-specific knowledge clusters, and within each cluster, from lower- to higher-order cognitive tasks. Compared with the second-best method, our method improves accuracy on medical evaluations by 2.49% and on financial evaluations by 1.2%. Ablation and cross-domain experiments further demonstrate our method as a scalable and effective framework for structured domain adaptation in large language model fine-tuning. We have released the code in an anonymous repository at https://github.com/Melo-1017/Balancing-Knowledge-Breadth-and-Task-Depth.

Medical insurance plays a pivotal role in the healthcare system, where effective supervision relies on the precise execution of complex medical insurance rules. Existing solutions for medical insurance violation detection mainly rely on costly manual labor, limiting efficiency. Although current Large Language Models (LLMs) have demonstrated significant potential in general Natural Language Processing (NLP) tasks, they still struggle to understand these rules, and a corresponding evaluation benchmark is lacking. Hence, we propose T2ER, a benchmark to evaluate the capability of LLMs in converting texts into executable rules in medical insurance. T2ER consists of 2,115 rules and 5,636 synthetic Electronic Medical Record (EMR) test cases for execution verification. Furthermore, we design a dual evaluation framework spanning rule and execution levels to comprehensively assess transformation quality. We conduct extensive experiments on advanced LLMs. The results reveal a significant human-AI performance gap, indicating that this conversion remains a formidable challenge. The dataset is available at https://github.com/YHJ1223/T2ER.

Large language models (LLMs) have demonstrated notable advancements in psychological counseling. However, existing models generally do not explicitly model seekers’ emotion shifts across counseling sessions, a core focus in classical psychological schools. Moreover, how to align counselor models’ responses with these emotion shifts while proactively mitigating safety risks remains underexplored. To bridge these gaps, we propose PsychēChat, which explicitly integrates emotion shift tracking and safety risk analysis for psychological counseling. Specifically, we employ interactive role-playing to synthesize counselor-seeker dialogues, incorporating two modules: Emotion Management Module, to capture seekers’ current emotions and emotion shifts; and Risk Control Module, to anticipate seekers’ subsequent reactions and identify potential risks. Furthermore, we introduce two modeling paradigms. The Agent Mode structures emotion management, risk control, and counselor responses into a collaborative multi-agent pipeline. The LLM Mode integrates these stages into a unified chain-of-thought for end-to-end inference, balancing efficiency and performance. Experiments with simulated interaction, dialogue-level evaluation, ablation studies, and expert assessment indicate that PsychēChat improves emotion-shift awareness and risk-aware response generation in psychological counseling support. Our data and code are available at https://github.com/XNeil12138/PsycheChat.