Zhenhua Liu, Lijun Li, Ruizhe Chen, Yuxian Jiang, Tong Zhu, Zhaochen Su, Wenliang Chen, Jing Shao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Jiayu Liu, Cheng Qian, Zhaochen Su, Qing Zong, Shijue Huang, Bingxiang He, Yi R. Fung. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Process-level Reward Models (PRMs) are crucial for complex reasoning and decisionmaking tasks, where each intermediate step plays an important role in the reasoning process.Since large language models (LLMs) suffer from various types of errors during the reasoning process, PRMs are required to possess nuanced capabilities for detecting various implicit error types in real-world scenarios.However, current benchmarks primarily focus on step correctness, failing to evaluate PRMs' performance systematically.To address this gap, we introduce PRMBENCH, a processlevel benchmark specifically designed to assess the fine-grained error detection capabilities of PRMs.PRMBENCH comprises 6,216 carefully designed problems and 83,456 step-level labels, evaluating models across multiple dimensions, including simplicity, soundness, and sensitivity.In our experiments on 25 models, spanning across both open-source PRMs and LLMs prompted as critic models, we uncover significant weaknesses in current PRMs.These findings reveal the challenges inherent in processlevel evaluation and highlight key directions for future research, establishing PRMBENCH as a robust testbed for advancing research on PRM evaluation and development.

The in-depth exploration and development of knowledge organization methodologies tailored to the unique characteristics of traditional Chinese medicine (TCM) has become a pivotal issue in advancing the modernization of TCM knowledge systems and fostering scientific innovation. Through systematic review and analysis of existing TCM knowledge organization research, this study identifies current limitations. It proposes exploring integrated approaches for integrating ancient and contemporary TCM knowledge within existing classical text repositories, constructing theoretical knowledge bases, conducting ontological mapping studies between historical and modern TCM knowledge, and developing efficient human-machine collaborative annotation methods to build knowledge graphs. These initiatives aim to provide actionable strategies for optimizing TCM knowledge organization frameworks.

Objective By applying data mining technology, this study collects, organizes, and researches the ancient prescriptions for treating palpitations in the “Chinese Formula Database”, analyzes their medication rules, and provides references for modern clinical medication and new drug research and development. Methods First, the medical records related to palpitations in the database were screened, and a Microsoft Excel dataset was created. Then, the”Ancient and Modern Medical Case Cloud Platform” was used to conduct frequency statistics on the drugs, as well as their properties, flavors, meridian tropisms and efficacies. Finally, IBM SPSS Modeler 18.0 and IBM SPSS Statistics 25 data-processing software were applied to perform association rule analysis and cluster analysis on high-frequency drugs (frequency$>100$, rate$> 0.85 \%$), so as to explore the compatibility rules among them. Results A total of$\mathbf{1, 1 5 7}$prescriptions were included, involving 639 kinds of drugs. Renshen had the highest frequency of use; drugs with warm property and sweet flavor were used most frequently; drugs that enter the spleen meridian were the most common; and drugs with the effect of calming the mind and improving intelligence had the highest frequency of use. Through association rule analysis and cluster analysis,$\mathbf{1 5}$groups of core drug pairs and 4 groups of core drug combinations were obtained. Conclusion Most ancient prescriptions for treating palpitations adopt four therapeutic approaches: promoting fluid production and nourishing blood, replenishing qi and consolidating the root, simultaneous treatment of the heart and spleen, and drying dampness and resolving phlegm. These approaches accurately correspond to the core pathogeneses of palpitations, namely, malnutrition of the heart spirit, deficiency of qi and blood, and phlegm stasis disturbing the heart.

Objective To develop a unified and standardised metadata framework for ancient books of traditional Chinese medicine, aiming to achieve standardised and regulated management of such resources. This initiative seeks to enhance the efficiency of managing and utilising ancient books of traditional Chinese medicine while providing a scientific guideline for their description. Methods Using literature review, comparative analysis, and expert consultation methods, this study analyzes and compares the current state of ancient books of traditional Chinese medicine metadata and relevant domestic and international standards. Based on the Dublin Core Element Set (DC) as the basic framework and tailored to the characteristics of ancient books of traditional Chinese medicine, it proposes a research idea and framework for ancient books of traditional Chinese medicine metadata standards. Results This study proposes the research ideas and principles of metadata standards for ancient Chinese medicine books centered on the knowledge system of traditional Chinese medicine, clearly stipulates the objects and information sources of ancient Chinese medicine book resources, constructs a metadata structure model and hierarchical relationship including core metadata, extended metadata and element modifiers, and defines the core metadata elements. Conclusion The metadata standard for ancient books of traditional Chinese medicine will facilitate the integration and sharing of ancient books of traditional Chinese medicine resources, thereby advancing the inheritance and development of traditional Chinese medicine knowledge.

Zhaochen Su, Juntao Li, Jun Zhang, Tong Zhu, Xiaoye Qu, Pan Zhou, Yan Bowen, Yu Cheng, Min Zhang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Large Vision-Language Models (LVLMs) have become pivotal at the intersection of computer vision and natural language processing.However, the full potential of LVLMs' Retrieval-Augmented Generation (RAG) capabilities remains underutilized.Existing works either focus solely on the text modality or are limited to specific tasks.Moreover, most LVLMs struggle to selectively utilize retrieved information and are sensitive to irrelevant or misleading references.To address these challenges, we propose a self-refinement framework designed to teach LVLMs to Selectively Utilize Retrieved Information (SURf).Specifically, when given questions that are incorrectly answered by the LVLM backbone, we obtain references that help correct the answers (positive references) and those that do not (negative references).We then fine-tune the LVLM backbone using a combination of these positive and negative references.Our experiments across three tasks and seven datasets demonstrate that our framework significantly enhances LVLMs' ability to effectively utilize retrieved multimodal references and improves their robustness against irrelevant or misleading information.The source code is available at https://github.com/GasolSun36/SURf. * Work done during internship at Shanghai AI Laboratory.† Both are corresponding authors.How many apples in the images? VQAVanilla: There are three apples.The image depicting five apples on a tree...The picture shows 7 apples .... leaves...Ours: There are four apples.Describe this image in details. Captioning Vanilla: A person walking in snow.The image depicting a...the skier is in a crouched position... The image captures a dynamic scene ..a skier dressed in a ...

Recent research has revealed that neural language models at scale suffer from poor temporal generalization capability, i.e., language model pre-trained on static data from past years performs worse over time on emerging data.Existing methods mainly perform continual training to mitigate such a misalignment.While effective to some extent but is far from being addressed on both the language modeling and downstream tasks.In this paper, we empirically observe that temporal generalization is closely affiliated with lexical semantic change, which is one of the essential phenomena of natural languages.Based on this observation, we propose a simple yet effective lexical-level masking strategy to post-train a converged language model.Experiments on two pre-trained language models, two different classification tasks, and four benchmark datasets demonstrate the effectiveness of our proposed method over existing temporal adaptation methods, i.e., continual training with new data.Our code is available at https: //github.com/zhaochen0110/LMLM.