Jincheng Liu, Sijun He, Jingjing Wu, Xiangsen Wang, Yang Chen, Zhaoqi Kuang, Siqi Bao, Yuan Yao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Kun Luo, Hongjin Qian, Zheng Liu, Ziyi Xia, Shitao Xiao, Zhao Cao, Siqi Bao, Jun Zhao, Kang Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Shaoning Sun, Mingzhu Cai, Huang He, Bingjin Chen, Siqi Bao, Yujiu Yang, Hua Wu, Haifeng Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Evaluating Large Language Models (LLMs) has become increasingly important, with automatic evaluation benchmarks gaining prominence as alternatives to human evaluation.While existing research has focused on approximating model rankings, such benchmarks fail to provide users and developers with a comprehensive and fine-grained understanding of a specific model's capabilities.To fill this gap, we propose SCAN (Structured Capability Assessment and Navigation), a practical framework that enables detailed characterization of LLM capabilities through comprehensive and fine-grained evaluation.SCAN incorporates four key components: (1) TaxBuilder, which extracts capability-indicating tags from extensive queries to construct a hierarchical taxonomy automatically; (2) RealMix, a query synthesis and filtering mechanism that ensures sufficient evaluation data for each capability tag; (3) a suite of visualization and analysis tools that facilitate efficient navigation and analysis of model capabilities; and (4) a PC 2based (Pre-Comparison-derived Criteria) LLMas-a-Judge approach that achieves significantly higher accuracy (see the definition of accuracy in § D) compared to classic LLM-as-a-Judge method.Using SCAN, we conduct a comprehensive evaluation of 21 mainstream LLMs.Our detailed analysis of the GPT-OSS family reveals substantial performance variations, even within sub-capabilities belonging to the same category of capability.This finding highlights the importance of fine-grained evaluation in accurately understanding LLM behavior.Project homepage and resources are available at https://github.com/liudan193/SCAN.

The Track-1 of DSTC9 aims to effectively answer user requests or questions during task-oriented dialogues, which are out of the scope of APIs/DB. By leveraging external knowledge resources, relevant information can be retrieved and encoded into the response generation for these out-of-API-coverage queries. In this work, we have explored several advanced techniques to enhance the utilization of external knowledge and boost the quality of response generation, includingschema guided knowledge decision,negatives enhanced knowledge selection, andknowledge grounded response generation. To evaluate the performance of our proposed method, comprehensive experiments have been carried out on the publicly available dataset. Our approach was ranked as the best in human evaluation of DSTC9 Track-1.

In recent years, Internet memes have been widely used in online chatting. Compared with text-based communication, conversations become more expressive and attractive when Internet memes are incorporated. This article presents our solutions for the Meme incorporated Open-domain Dialogue (MOD) Challenge of DSTC10, where three tasks are involved: text response modeling, meme retrieval, and meme emotion classification. Firstly, we leverage a large-scale pre-trained dialogue model for coherent and informative response generation. Secondly, based on interaction-based text-matching, our approach can retrieve appropriate memes with good generalization ability. Thirdly, we propose to model the emotion flow (EF) in conversations and introduce an auxiliary task of emotion description prediction (EDP) to boost the performance of meme emotion classification. Experimental results on the MOD dataset demonstrate that our methods can incorporate Internet memes into dialogue systems effectively.

Many open-domain dialogue models pretrained with social media comments can generate coherent replies but have difficulties producing engaging responses.This phenomenon might mainly result from the deficiency of annotated human-human conversations and the misalignment with human preference.In this paper, we propose a novel and efficient framework Diamante to boost the open-domain chatbot, where two kinds of human feedback (including explicit demonstration and implicit preference) are collected and leveraged.By asking annotators to select or amend the modelgenerated candidate responses, Diamante efficiently collects the human demonstrated responses and constructs a Chinese chit-chat dataset.To enhance the alignment with human preference, Diamante leverages the implicit preference in the data collection process and introduces the generation-evaluation joint training.Comprehensive experiments indicate that the Diamante dataset and joint training paradigm can significantly boost the performance of pre-trained dialogue models.The overall engagingness of the previous state-ofthe-art model has been improved remarkably by 50% in Chinese open-domain conversations.

In this paper, we propose an unsupervised query enhanced approach for knowledgeintensive conversations, namely QKConv.There are three modules in QKConv: a query generator, an off-the-shelf knowledge selector, and a response generator.QKConv is optimized through joint training, which produces the response by exploring multiple candidate queries and leveraging corresponding selected knowledge.The joint training solely relies on the dialogue context and target response, getting exempt from extra query annotations or knowledge provenances.To evaluate the effectiveness of the proposed QKConv, we conduct experiments on three representative knowledgeintensive conversation datasets: conversational question-answering, task-oriented dialogue, and knowledge-grounded conversation.Experimental results reveal that QKConv performs better than all unsupervised methods across three datasets and achieves competitive performance compared to supervised methods.

Existing pipelined task-oriented dialogue systems usually have difficulties adapting to unseen domains, whereas end-to-end systems are plagued by large-scale knowledge bases in practice.In this paper, we introduce a novel querydriven task-oriented dialogue system, namely Q-TOD.The essential information from the dialogue context is extracted into a query, which is further employed to retrieve relevant knowledge records for response generation.Firstly, as the query is in the form of natural language and not confined to the schema of the knowledge base, the issue of domain adaption is alleviated remarkably in Q-TOD.Secondly, as the query enables the decoupling of knowledge retrieval from the generation, Q-TOD gets rid of the issue of knowledge base scalability.To evaluate the effectiveness of the proposed Q-TOD, we collect query annotations for three publicly available task-oriented dialogue datasets.Comprehensive experiments verify that Q-TOD outperforms strong baselines and establishes a new state-of-the-art performance on these datasets.

The rapid development of single-cell RNA-sequencing (scRNA-seq) technology has raised significant computational and analytical challenges. The application of deep learning to scRNA-seq data analysis is rapidly evolving and can overcome the unique challenges in upstream (quality control and normalization) and downstream (cell-, gene- and pathway-level) analysis of scRNA-seq data. In the present study, recent advances and applications of deep learning-based methods, together with specific tools for scRNA-seq data analysis, were summarized. Moreover, the future perspectives and challenges of deep-learning techniques regarding the appropriate analysis and interpretation of scRNA-seq data were investigated. The present study aimed to provide evidence supporting the biomedical application of deep learning-based tools and may aid biologists and bioinformaticians in navigating this exciting and fast-moving area.

Uveal melanoma (UVM) is the most common primary intraocular human malignancy with a high mortality rate. Aberrant DNA methylation has rapidly emerged as a diagnostic and prognostic signature in many cancers. However, such DNA methylation signature available in UVM remains limited. In this study, we performed a genome-wide integrative analysis of methylome and transcriptome and identified 40 methylation-driven prognostic genes (MDPGs) associated with the tumorigenesis and progression of UVM. Then, we proposed a machine-learning-based discovery and validation strategy to identify a DNA methylation-driven signature (10MeSig) composing of 10 MDPGs (AZGP1, BAI1, CCDC74A, FUT3, PLCD1, S100A4, SCN8A, SEMA3B, SLC25A38 and SLC44A3), which stratified 80 patients of the discovery cohort into two risk subtypes with significantly different overall survival (HR = 29, 95% CI: 6.7-126, P < 0.001). The 10MeSig was validated subsequently in an independent cohort with 57 patients and yielded a similar prognostic value (HR = 2.1, 95% CI: 1.2-3.7, P = 0.006). Multivariable Cox regression analysis showed that the 10MeSig is an independent predictive factor for the survival of patients with UVM. With a prospective validation study, this 10MeSig will improve clinical decisions and provide new insights into the pathogenesis of UVM.

Tumor-infiltrating immune cells (TIICs) have been recognized as crucial components of the tumor microenvironment (TME) and induced both beneficial and adverse consequences for tumorigenesis as well as outcome and therapy (particularly immunotherapy). Computer-aided investigation of immune cell components in the TME has become a promising avenue to better understand the interplay between the immune system and tumors. In this study, we presented an overview of data sources, computational methods and software tools, as well as their application in inferring the composition of tumor-infiltrating immune cells in the TME. In parallel, we explored the future perspectives and challenges that may be faced with more accurate quantitative infiltration of immune cells in the future. Together, our study provides a little guide for scientists in the field of clinical and experimental immunology to look for dedicated resources and more competent tools for accelerating the unraveling of tumor-immune interactions with the implication in precision immunotherapy.

An accurate prognosis assessment for cancer patients could aid in guiding clinical decision-making. Reliance on traditional clinical features alone in a complex clinical environment is challenging and unsatisfactory in the era of precision medicine; thus, reliable prognostic biomarkers are urgently required to improve a patient staging system. In this study, we proposed a patient-level computational framework from mechanistic and translational perspectives to establish a personalized prognostic signature (named PLPPS) in high-grade serous ovarian carcinoma (HGSOC). The PLPPS composed of 68 immune genes achieved accurate prognostic risk stratification for 1190 patients in the meta-training cohort and was rigorously validated in multiple cross-platform independent cohorts comprising 792 HGSOC patients. Furthermore, the PLPPS was shown to be the better prognostic factor compared with clinical parameters in the univariate analysis and retained a significant independent association with prognosis after adjusting for clinical parameters in the multivariate analysis. In benchmark comparisons, the performance of PLPPS (hazard ratio (HR), 1.371; concordance index (C-index), 0.604 and area under the curve (AUC), 0.637) is comparable to or better than other published gene signatures (HR, 0.972 to 1.340; C-index, 0.495 to 0.592 and AUC, 0.48-0.624). With further validation in prospective clinical trials, we hope that the PLPPS might become a promising genomic tool to guide personalized management and decision-making of HGSOC in clinical practice.

Long noncoding RNAs (lncRNAs) have been associated with cancer immunity regulation and the tumor microenvironment (TME). However, functions of lncRNAs of tumor-infiltrating B lymphocytes (TIL-Bs) and their clinical significance have not yet been fully elucidated. In the present study, a machine learning-based computational framework is presented for the identification of lncRNA signature of TIL-Bs (named 'TILBlncSig') through integrative analysis of immune, lncRNA and clinical profiles. The TILBlncSig comprising eight lncRNAs (TNRC6C-AS1, WASIR2, GUSBP11, OGFRP1, AC090515.2, PART1, MAFG-DT and LINC01184) was identified from the list of 141 B-cell-specific lncRNAs. The TILBlncSig was capable of distinguishing worse compared with improved survival outcomes across different independent patient datasets and was also independent of other clinical covariates. Functional characterization of TILBlncSig revealed it to be an indicator of infiltration of mononuclear immune cells (i.e. natural killer cells, B-cells and mast cells), and it was associated with hallmarks of cancer, as well as immunosuppressive phenotype. Furthermore, the TILBlncSig revealed predictive value for the survival outcome and immunotherapy response of patients with anti-programmed death-1 (PD-1) therapy and added significant predictive power to current immune checkpoint gene markers. The present study has highlighted the value of the TILBlncSig as an indicator of immune cell infiltration in the TME from a noncoding RNA perspective and strengthened the potential application of lncRNAs as predictive biomarkers of immunotherapy response, which warrants further investigation.

Pre-training models have been proved effective for a wide range of natural language processing tasks.Inspired by this, we propose a novel dialogue generation pre-training framework to support various kinds of conversations, including chit-chat, knowledge grounded dialogues, and conversational question answering.In this framework, we adopt flexible attention mechanisms to fully leverage the bi-directional context and the uni-directional characteristic of language generation.We also introduce discrete latent variables to tackle the inherent one-to-many mapping problem in response generation.Two reciprocal tasks of response generation and latent act recognition are designed and carried out simultaneously within a shared network.Comprehensive experiments on three publicly available datasets verify the effectiveness and superiority of the proposed framework.

Topic modeling has been intensively studied and widely applied in both academia and industry in the last decade. In the literature, topic models usually need to be trained from scratch for each individual corpus. Hence, the wisdom of the crowd (i.e., topic models previously trained based upon other corpora) is abandoned. Since a massive amount of in-domain data, considerable computational cost, and human labour are involved in obtaining a high-quality topic model, training from scratch for each new corpus is a huge waste of resources. In this paper, we propose the novel TopicOcean framework, which aims to integrate well-trained topic models and transfer the knowledge of accumulated topics to new corpora in order to improve the quality of their topic models. We first propose a method of constructing the ever-increasing TopicOcean, and then propose a meta-learning mechanism that transfers the meta-level knowledge (i.e., topics) in TopicOcean to the scenario of topic modeling on new corpora. Comprehensive experiments validate that the TopicOcean framework can significantly outperform the state-of-the-art (53.77% perplexity improvement on a temporal-shift corpus and 29.24% improvement on a domain-shift corpus). The well-trained high-quality topic models used to construct TopicOcean have been opensourced to promote further research.11The well-trained topic models can be accessed at Github (https://github.com/baidu/Familia/blob/master/model/download_model.sh).

Emerging evidence revealed the critical roles of long non-coding RNAs (lncRNAs) in maintaining genomic instability. However, identification of genome instability-associated lncRNAs and their clinical significance in cancers remain largely unexplored. Here, we developed a mutator hypothesis-derived computational frame combining lncRNA expression profiles and somatic mutation profiles in a tumor genome and identified 128 novel genomic instability-associated lncRNAs in breast cancer as a case study. We then identified a genome instability-derived two lncRNA-based gene signature (GILncSig) that stratified patients into high- and low-risk groups with significantly different outcome and was further validated in multiple independent patient cohorts. Furthermore, the GILncSig correlated with genomic mutation rate in both ovarian cancer and breast cancer, indicating its potential as a measurement of the degree of genome instability. The GILncSig was able to divide TP53 wide-type patients into two risk groups, with the low-risk group showing significantly improved outcome and the high-risk group showing no significant difference compared with those with TP53 mutation. In summary, this study provided a critical approach and resource for further studies examining the role of lncRNAs in genome instability and introduced a potential new avenue for identifying genomic instability-associated cancer biomarkers.

In this paper, a novel Generation-Evaluation framework is developed for multi-turn conversations with the objective of letting both participants know more about each other.For the sake of rational knowledge utilization and coherent conversation flow, a dialogue strategy which controls knowledge selection is instantiated and continuously adapted via reinforcement learning.Under the deployed strategy, knowledge grounded conversations are conducted with two dialogue agents.The generated dialogues are comprehensively evaluated on aspects like informativeness and coherence, which are aligned with our objective and human instinct.These assessments are integrated as a compound reward to guide the evolution of dialogue strategy via policy gradient.Comprehensive experiments have been carried out on the publicly available dataset, demonstrating that the proposed method outperforms the other state-of-the-art approaches significantly.

In this paper, a novel 3D deep learning network is proposed for brain MR image segmentation with randomized connection, which can decrease the dependency between layers and increase the network capacity. The convolutional LSTM and 3D convolution are employed as network units to capture the long-term and short-term 3D properties respectively. To assemble these two kinds of spatial-temporal information and refine the deep learning outcomes, we further introduce an efficient graph-based node selection and label inference method. Experiments have been carried out on two publicly available databases and results demonstrate that the proposed method can obtain competitive performances as compared with other state-of-the-art methods.

In this paper, a novel label fusion method is proposed for brain magnetic resonance image segmentation. This label fusion method is formulated on a graph, which embraces both label priors from atlases and anatomical priors from target image. To represent a pixel in a comprehensive way, three kinds of feature vectors are generated, including intensity, gradient, and structural signature. To select candidate atlas nodes for fusion, rather than exact searching, randomized k-d tree with spatial constraint is introduced as an efficient approximation for high-dimensional feature matching. Feature sensitive label prior (FSLP), which takes both the consistency and variety of different features into consideration, is proposed to gather atlas priors. As FSLP is a non-convex problem, one heuristic approach is further designed to solve it efficiently. Moreover, based on the anatomical knowledge, parts of the target pixels are also employed as the graph seeds to assist the label fusion process, and an iterative strategy is utilized to gradually update the label map. The comprehensive experiments carried out on two publicly available databases give results to demonstrate that the proposed method can obtain better segmentation quality.

In this paper, a novel unified framework for atlas-based segmentation is proposed, consisting of two main components: forward deformation and label refinement. A newly designed distance constraint on mesh edges is enforced with contrast sensitivity in forward deformation based on Markov random field. With the edge distance constraint, the object shapes in the atlas and the target images can remain similar during deformation. Considering the shape variations caused by individual difference, we then develop a label refinement process embracing patch registration and label fusion to compensate the small variations around the structural surfaces. As the anatomical correspondences determined in forward deformation can differ from that in label refinement, the conventional one-to-one correspondence constraint can be relaxed in our framework. Experiments on two publicly available databases IBSR and LPBA40 demonstrate that our method can obtain better performance as compared with other state-of-the-art methods.

In this paper, a novel label inference method encoded with local and global patch priors is introduced for the segmentation of subcortical structures in brain MR images. Due to the serious overlap of intensity profiles among different tissues in brain MR images, the conventional patch prior estimated with similarity measurement can be adversely impacted and become misleading during the final label inference procedure. As such, to obtain a more discriminative patch representation, we propose to capture local patch prior using sparse learning. Besides the local and low-level patch prior, the high-level structural properties of each subcortical structure are also taken into consideration and global patch prior is extracted with Convolutional Neural Networks. Experiments have been carried out on two publicly available datasets and results indicate that the proposed method can obtain the best performance as compared with other state-of-the-art methods.