Improving large language models (LLMs) for electronic health record (EHR) reasoning is essential for enabling accurate and generalizable clinical predictions. While LLMs excel at medical text understanding, they underperform on EHR-based prediction tasks due to challenges in modeling temporally structured, high-dimensional data. Existing approaches often rely on hybrid paradigms, where LLMs serve merely as frozen prior retrievers while downstream deep learning (DL) models handle prediction, failing to improve the LLM’s intrinsic reasoning capacity and inheriting the generalization limitations of DL models. To this end, we propose EAG-RL, a novel two-stage training framework designed to intrinsically enhance LLMs’ EHR reasoning ability through expert attention guidance, where expert EHR models refer to task-specific DL models trained on EHR data. Concretely, EAG-RL first constructs high-quality, stepwise reasoning trajectories using expert-guided Monte Carlo Tree Search to effectively initialize the LLM’s policy. Then, EAG-RL further optimizes the policy via reinforcement learning by aligning the LLM’s attention with clinically salient features identified by expert EHR models. Extensive experiments on two real-world EHR datasets show that EAG-RL improves the intrinsic EHR reasoning ability of LLMs by an average of 14.62%, while also enhancing robustness to feature perturbations and generalization to unseen clinical domains. These results demonstrate the practical potential of EAG-RL for real-world deployment in clinical prediction tasks.

Hongxin Ding, Baixiang Huang, Yue Fang, Weibin Liao, Xinke Jiang, Jinyang Zhang, Yinghao Zhu, Zheng Li, Liantao Ma, Junfeng Zhao, Yasha Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

While agentic AI systems rely on LLMs to translate user intent into structured function calls, this process is fraught with computational redundancy, leading to high inference latency that hinders real-time applications. This paper identifies and addresses three key redundancies: (1) the redundant processing of a large library of function descriptions for every request; (2) the redundant use of a large, slow model to generate an entire, often predictable, token sequence; and (3) the redundant generation of fixed, boilerplate parameter syntax. We introduce HyFunc, a novel framework that systematically eliminates these inefficiencies. HyFunc employs a hybrid-model cascade where a large model distills user intent into a single ''soft token.'' This token guides a lightweight retriever to select relevant functions and directs a smaller, prefix-tuned model to generate the final call, thus avoiding redundant context processing and full-sequence generation by the large model. To eliminate syntactic redundancy, our ''dynamic templating'' technique injects boilerplate parameter syntax on-the-fly within an extended vLLM engine. To avoid potential limitations in generalization, we evaluate HyFunc on an unseen benchmark dataset, BFCL. Experimental results demonstrate that HyFunc achieves an excellent balance between efficiency and performance. It achieves an inference latency of 0.828 seconds, outperforming all baseline models, and reaches a performance of 80.1%, surpassing all models with a comparable parameter scale. These results suggest that HyFunc offers a more efficient paradigm for agentic AI. Our code is publicly available at https://github.com/MrBlankness/HyFunc.

Hongxin Ding, Yue Fang, Runchuan Zhu, Xinke Jiang, Jinyang Zhang, Yongxin Xu, Weibin Liao, Xu Chu, Junfeng Zhao, Yasha Wang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

In the domain of complex reasoning tasks, such as mathematical reasoning, recent advancements have proposed the use of Direct Preference Optimization (DPO) to suppress output of dispreferred responses, thereby enhancing the long-chain reasoning capabilities of large language models (LLMs). To this end, these studies employed LLMs to generate preference trees via Tree-of-thoughts (ToT) and sample the paired preference responses required by the DPO algorithm. However, the DPO algorithm based on binary preference optimization is unable to learn multiple responses with varying degrees of preference/dispreference that provided by the preference trees, resulting in incomplete preference learning. In this work, we introduce Tree Preference Optimization (TPO), that does not sample paired preference responses from the preference tree; instead, it directly learns from the entire preference tree during the fine-tuning. Specifically, TPO formulates the language model alignment as a Preference List Ranking problem, where the policy can potentially learn more effectively from a ranked preference list of responses given the prompt. In addition, to further assist LLMs in identifying discriminative steps within long-chain reasoning and increase the relative reward margin in the preference list, TPO utilizes Adaptive Step Reward to adjust the reward values of each step in trajectory for performing fine-grained preference optimization. We carry out extensive experiments on mathematical reasoning tasks to evaluate TPO. The experimental results indicate that TPO consistently outperforms DPO across five public large language models on four datasets.

Analyzing the health status of patients based on Electronic Health Records (EHR) is a fundamental research problem in medical informatics. The presence of extensive missing values in EHR makes it challenging for deep neural networks (DNNs) to directly model the patient's health status. Existing DNNs training protocols, including Impute-then-Regress Procedure and Jointly Optimizing of Impute-n-Regress Procedure, require the additional imputation models to reconstruction missing values. However, Impute-then-Regress Procedure introduces the risk of injecting imputed, non-real data into downstream clinical prediction tasks, resulting in power loss, biased estimation, and poorly performing models, while Jointly Optimizing of Impute-n-Regress Procedure is also difficult to generalize due to the complex optimization space and demanding data requirements. Inspired by the recent advanced literature of learnable prompt in the fields of NLP and CV, in this work, we rethought the necessity of the imputation model in downstream clinical tasks, and proposed Learnable Prompt as Pseudo-Imputation (PAI) as a new training protocol to assist EHR analysis. PAI no longer introduces any imputed data but constructs a learnable prompt to model the implicit preferences of the downstream model for missing values, resulting in a significant performance improvement for all state-of-the-arts EHR analysis models on four real-world datasets across two clinical prediction tasks. Further experimental analysis indicates that PAI exhibits higher robustness in situations of data insufficiency and high missing rates. More importantly, as a plug-and-play protocol, PAI can be easily integrated into any existing or even imperceptible future EHR analysis models. The code of this work is deployed publicly available at https://github.com/MrBlankness/PAI to help the research community reproduce the results and assist the EHR analysis tasks.

Medical Lay Language Generation (MLLG) plays a vital role in improving the accessibility of complex scientific content for broader audiences. Recent literature to MLLG commonly employ parameter-efficient fine-tuning methods such as Low-Rank Adaptation (LoRA) to fine-tuning large language models (LLMs) using paired expert-lay language datasets. However, LoRA struggles with the challenges posed by multi-source heterogeneous MLLG datasets. Specifically, through a series of exploratory experiments, we reveal that standard LoRA fail to meet the requirement for semantic fidelity and diverse lay-style generation in MLLG task. To address these limitations, we propose Magical, an asymmetric LoRA architecture tailored for MLLG under heterogeneous data scenarios. Magical employs a shared matrix A for abstractive summarization, along with multiple isolated matrices B for diverse lay-style generation. To preserve semantic fidelity during the lay language generation process, Magical introduces a Semantic Invariance Constraint to mitigate semantic subspace shifts on matrix A. Furthermore, to better adapt to diverse lay-style generation, Magical incorporates the Recommendation-guided Switch, an externally interface to prompt the LLM to switch between different matrices B. Experimental results on three real-world lay language generation datasets demonstrate that Magical consistently outperforms prompt-based methods, vanilla LoRA, and its recent variants, while also reducing trainable parameters by 31.66%. Our code is publicly available at https://github.com/tianlwang/Magical.git.

Acquiring reviewers for academic submissions is a challenging recommendation scenario. Recent graph learning-driven models have made remarkable progress in the field of recommendation, but their performance in the academic reviewer recommendation task may suffer from a significant false negative issue. This arises from the assumption that unobserved edges represent negative samples. In fact, the mechanism of anonymous review results in inadequate exposure of interactions between reviewers and submissions, leading to a higher number of unobserved interactions compared to those caused by reviewers declining to participate. Therefore, investigating how to better comprehend the negative labeling of unobserved interactions in academic reviewer recommendations is a significant challenge. This study aims to tackle the ambiguous nature of unobserved interactions in academic reviewer recommendations. Specifically, we propose an unsupervised Pseudo Neg-Label strategy to enhance graph contrastive learning (GCL) for recommending reviewers for academic submissions, which we call RevGNN. RevGNN utilizes a two-stage encoder structure that encodes both scientific knowledge and behavior using Pseudo Neg-Label to approximate review preference. Extensive experiments on three real-world datasets demonstrate that RevGNN outperforms all baselines across four metrics. Additionally, detailed further analyses confirm the effectiveness of each component in RevGNN.