Recent research reveals that a minority of high-entropy tokens significantly influence the reasoning quality of large language models (LLMs). Inspired by this, we propose Prototype Entropy Alignment (PEA), a reinforcement learning framework that models effective reasoning not as a single path but as a collection of learnable "entropy signatures." PEA identifies these signatures by clustering expert trajectories' uncertainty patterns into a diverse and continuously updated set of prototypes. The model is then rewarded for aligning its own reasoning process with these evolving targets, creating a self-improvement loop. Instead of replacing traditional outcome-based rewards, PEA provides a complementary, process-oriented signal. Our experiments show that this synergy is crucial: PEA substantially boosts performance on creative and general reasoning tasks and, when combined with outcome rewards, achieves SOTA results on structured tasks such as mathematics. By rewarding alignment with diverse and evolving reasoning structures, PEA offers a robust, verifier-free pathway to enhance reasoning's adaptability.

Guanran Luo, Wentao Qiu, Wanru Zhao, Wenhan Lv, Zhongquan Jian, Meihong Wang, Qingqiang Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Retrieval-Augmented Generation (RAG) mitigates the hallucination issues of large language models (LLMs) by integrating external knowledge.For abstractive questions involving multistep reasoning, knowledge from multiple sections is often required.To address this issue, recent research has introduced recursive summarization, which constructs a hierarchical summary tree by clustering text chunks, integrating information from various parts of the document to provide evidence for abstractive questions.However, summary trees often contain a large number of redundant summary nodes, which not only increase construction time but may also negatively impact question answering.Moreover, recursive summarization is not suitable for all types of questions.We introduce DTCRS, a method that dynamically generates summary trees based on document structure and query semantics.DTCRS determines whether a summary tree is necessary by analyzing the question type.It then decomposes the question and uses the embeddings of subquestions as initial cluster centers, reducing redundant summaries while improving the relevance between summaries and the question.Our approach significantly reduces summary tree construction time and achieves substantial improvements across three QA tasks.Additionally, we investigate the applicability of recursive summarization to different question types, providing valuable insights for future research.

Prior studies on Aspect-level Sentiment Classification (ALSC) emphasize modeling interrelationships among aspects and contexts but overlook the crucial role of aspects themselves as essential domain knowledge. To this end, we propose AGCL, a novel Aspect Graph Construction and Learning method, aimed at furnishing the model with finely tuned aspect information to bolster its task-understanding ability. AGCL’s pivotal innovations reside in Aspect Graph Construction (AGC) and Aspect Graph Learning (AGL), where AGC harnesses intrinsic aspect connections to construct the domain aspect graph, and then AGL iteratively updates the introduced aspect graph to enhance its domain expertise, making it more suitable for the ALSC task. Hence, this domain aspect graph can serve as a bridge connecting unseen aspects with seen aspects, thereby enhancing the model’s generalization capability. Experiment results on three widely used datasets demonstrate the significance of aspect information for ALSC and highlight AGL’s superiority in aspect learning, surpassing state-of-the-art baselines greatly. Code is available at https://github.com/jian-projects/agcl.

Recognizing emotions in dialogues is vital for effective human-computer interaction, yet remains a challenging task in Natural Language Processing (NLP). Previous studies in Emotion Recognition in Conversation (ERC) have primarily focused on contextual features, while overlooking the importance of emotional features in emotion recognition. To address this gap, we focus on the role of emotional features in ERC and propose a novel method, Emotional Knowledge Self-Distillation (EmoKSD1), to enhance the model’s emotional sensitivity. In EmoKSD, utterances are enriched with implicit ⟨mask⟩ tokens to represent conveyed emotions, allowing the distillation of emotional knowledge from explicit emotional tokens to implicit ⟨mask⟩ tokens, thereby enhancing the model’s ability to perceive subtle emotions within the dialogue. Through thorough evaluations on two public ERC datasets (i.e., IEMOCAP and MELD) using proposed coarse-grained utterance distillation and fine-grained token distillation techniques, EmoKSD demonstrates superior performance compared to existing methods, highlighting the significance of emotional features in ERC.

Pre-trained Language Models (PLMs) have achieved remarkable performance in various Natural Language Processing (NLP) tasks, including Aspect-based Sentiment Analysis (ABSA). Therefore, numerous ABSA models based on PLMs have been proposed, primarily focusing on module design to exploit the inherent connections between aspects and contexts. However, the core factor driving performance improvements, the PLM’s powerful semantic understanding capabilities, has not been fully considered, raising the question of how to further unlock their potential for downstream tasks. To this end, we introduce a novel training strategy, called CCL1, which integrates the strengths of Curriculum Learning (CurL) and Contrastive Learning (ConL) to facilitate the learning of robust feature representations. For the ABSA task, we use aspect similarities to develop the CurL strategy, grouping samples with similar aspects into batches. This allows ConL to learn more robust representations by providing related samples within each batch. The superiority of CCL is demonstrated through extensive experiments on two public ABSA datasets, with ablation studies validating the effectiveness of combining CurL and ConL in enhancing aspect understanding.

This research proposes the METORIE dataset1, a novel resource designed to improve the reasoning capabilities of large language models (LLMs), such as LLaMA3 and GLM4, in information extraction (IE) tasks. The METORIE dataset is derived from brain teasers that incorporate complex logical and metaphorical elements and is designed to train LLMs to navigate intricate reasoning paths and interpret layered expressions. Our findings demonstrate that the METORIE dataset markedly enhances LLMs’ performance across both general and specialized IE tasks. The results of fine-tuning with the METORIE dataset, mixed with a small number of IE datasets, are close to, if not exceeding, those of LLMs of the same parametric size on IE tasks using much larger datasets. Through controlled experiments, we establish that metaphors of medium complexity optimize IE performance, while higher complexities tend to overstretch LLMs’ inference limits. METORIE-fine-tuned LLMs also demonstrate exceptional performance in legal and medical domains, suggesting that enhanced metaphor understanding and logical deduction are key to improving LLMs’ adaptability and efficiency in vertical domains.

In an emotional conversation, emotions are causally transmitted among communication participants, constituting a fundamental conversational feature that can facilitate the comprehension of intricate changes in emotional states during the conversation and contribute to neutralizing emotional semantic bias in utterance caused by the absence of modality information. Therefore, emotional transition (ET) plays a crucial role in the task of Emotion Recognition in Conversation (ERC) that has not received sufficient attention in current research. In light of this, an Emotional Transition-based Emotion Recognizer (EmoTrans) is proposed in this paper. Specifically, we concatenate the most recent utterances with their corresponding speakers to construct the model input, known as samples, each with several placeholders to implicitly express the emotions of contextual utterances. Based on these placeholders, two components are developed to make the model sensitive to emotions and effectively capture the ET features in the sample. Furthermore, an ET-based Contrastive Learning (CL) is developed to compact the representation space, making the model achieve more robust sample representations. We conducted exhaustive experiments on four widely used datasets and obtained competitive experimental results, especially, new state-of-the-art results obtained on MELD and IEMOCAP, demonstrating the superiority of EmoTrans.

Effective extraction and integration of valuable contextual information is the core of models for the Emotion Recognition in Conversation (ERC) task. However, a significant amount of irrelevant information is inevitably introduced when integrating long-range contextual information, perplexing the model greatly and resulting in incorrect emotion identification. To this end, we proposed a Conversation Clique-based Model (CCM), designed to extract the most efficacious contextual information to bolster the semantic quality of utterances. Specifically, we devise an utterance spatial relationship module (SpaRel) to explicitly model structural-level correlations among utterances by using GAT, and an emotion temporal relationship module (TemRel) to implicitly capture the emotion sequence constraints by employing HMM. We conduct extensive experiments on the publicly available MELD dataset, and the experimental results indicate the effectiveness of our proposed model, achieving new state-of-the-art results.

In this paper, we present the Intra- and Inter-Human Relation Networks I²R-Net for Multi-Person Pose Estimation. It involves two basic modules. First, the Intra-Human Relation Module operates on a single person and aims to capture Intra-Human dependencies. Second, the Inter-Human Relation Module considers the relation between multiple instances and focuses on capturing Inter-Human interactions. The Inter-Human Relation Module can be designed very lightweight by reducing the resolution of feature map, yet learn useful relation information to significantly boost the performance of the Intra-Human Relation Module. Even without bells and whistles, our method can compete or outperform current competition winners. We conduct extensive experiments on COCO, CrowdPose, and OCHuman datasets. The results demonstrate that the proposed model surpasses all the state-of-the-art methods. Concretely, the proposed method achieves 77.4% AP on CrowPose dataset and 67.8% AP on OCHuman dataset respectively, outperforming existing methods by a large margin. Additionally, the ablation study and visualization analysis also prove the effectiveness of our model.

Mobile devices have generated massive textual data with geographical locations. These data are applied to the topic discovery, event detection and user behavior analysis in enterprise systems. Therefore, geospatial data mining has become a very important and challenging research topic in such systems. In this paper, we develop a geospatial data mining system called GDMS to support the retrieval and analysis of textual data with geographical locations. The system contains three components: data collection, data analysis and data visualization. First, a large number of geospatical data are collected from our implemented mobile APP that is used by community residents. Residents can use the APP to upload abnormal events by text descriptions with geographical locations. All these events are processed and stored in a server. In the data analysis component, we focus on the problem of finding textual topics of clusters containing text descriptions with geographical locations. The key is how to combine clustering techniques with topic-retrieval models to integrate both geo-location information and text information. We investigated methods that combine clustering methods with the knowledge graph to discover topics of clusters of documents with geo-locations. Finally, we demonstrate an effective visualization tool that shows detected textual topics on the map in our mobile APP that is used by government staffs.

Crowdsourcing is quite cheap and effective to get a data set labeled by multiple annotators in a short amount of time. Although there are many traditional methods focusing on quality control of crowdsourcing, little pays close attention to the intrinsic ambiguity in dataset, which is hard for workers to make the right decision. In this paper, we introduce DCRB, a collaborative crowdsourcing framework that can obtain high quality output even for ambiguous tasks. We enable users' disagreement and encourage them to provide in-depth explanations when the disagreement occurs. We use these explanations to analyze the ambiguous tasks and design collaborative crowdsourcing to improve the output. In addition, we also design “reward brave” incentive mechanism to encourage users' valuable explanations. The experimental results show that our method significantly improves the accuracy of crowdsourcing, especially for those ambiguous crowdsourcing tasks.

Identification of gas-oil reservoir is always important but rather difficult in global gas-oil exploration. It is of the great significance to improve the accuracy of reservoir recognition. Seismic exploration is one of the most valuable methods of gas-oil exploration, and the huge amounts of seismic attribute data can be useful for gas-oil exploration. One limitation of the Generative Topographic Mapping (GTM) algorithm is that it cannot determine the classifications of the data points with close probabilities accurately, and it would be more likely to result in confused clarification and fuzzy boundary. To overcome the limitation, an advanced GTM algorithm with Euclidean Distance (GTM-ED) is proposed in this paper, and we use Euclidean Distance to compute the distance from the edge points to the neighbor centroids, and classify it to the closet class to avoid the problems of confused classification. And then the GTM-ED algorithm is used in the research of reservoir identification model, experiments are made with actual seismic data set. First of all, the GTM algorithm is discussed, and then the GTM-ED algorithm is introduced. And afterwards, many experiments are made. In the experiments, the log data and geological data are selected as the labels, and the comparison and analysis are made through three aspects, including relative criteria, absolute criteria, and run-time, and then the results of each model are visualized. The experimental results indicate that the GTM-ED can achieve better results in reservoir clustering and unknown reservoir identification. And in the actual application, the visualization of the GTM-ED can behave better than the GTM in showing the geological characteristics of paleochannel, the string of beads-like reservoirs and linear lava.

Google Maps is an excellent GIS (Geographic Information System) product, but its visual effect is unsatisfactory under the specific environment. As is known to all, visualization is an important aspect to evaluate the quality of electronic map. Firstly, we analyze factors of visual effect in the paper. Combining implementation mechanisms of RFID (Radio Frequency Identification Devices) and Google Maps, we designed the system of electronic map which owns a reasonable amount of displayed information, beautiful legends and fast reaction rate. After that, we made an example to verify the effectiveness of the system's performance.

This paper summarizes the 2010 CLSP Summer Workshop on speech recognition at Johns Hopkins University. The key theme of the workshop was to improve on state-of-the-art speech recognition systems by using Segmental Conditional Random Fields (SCRFs) to integrate multiple types of information. This approach uses a state of-the-art baseline as a springboard from which to add a suite of novel features including ones derived from acoustic templates, deep neural net phoneme detections, duration models, modulation features, and whole word point-process models. The SCRF framework is able to appropriately weight these different information sources to produce significant gains on both die Broadcast News and Wall Street Journal tasks.

We apply the framework of kernel dimension reduction, originally designed for supervised problems, to unsupervised dimensionality reduction. In this framework, kernel-based measures of independence are used to derive low-dimensional representations that maximally capture information in covariates in order to predict responses. We extend this idea and develop similarly motivated measures for unsupervised problems where covariates and responses are the same. Our empirical studies show that the resulting compact representation yields meaningful and appealing visualization and clustering of data. Furthermore, when used in conjunction with supervised learners for classification, our methods lead to lower classification errors than state-of-the-art methods, especially when embedding data in spaces of very few dimensions.