Yifei Li, Weidong Guo, Lingling Zhang, Rongman Xu, Muye Huang, Hui Liu, Lijiao Xu, Yu Xu, Jun Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Aligning Large Language Models (LLMs) with nuanced user instructions is critical for their effective deployment in real-world applications. While prior methods focus on enhancing data diversity and complexity, they often overlook models' sensitivity to fine-grained variations in semantically similar instructions. To address this, we introduce DeMoRecon, a data augmentation framework that decomposes complex instructions into sub-components, modifies individual elements, and reconstructs them into instruction variants. This method preserves contextual integrity while injecting targeted variability essential for fine-grained instruction-following. Based on DeMoRecon, we construct the FGIV dataset, comprising over 1,700 seed instructions and thousands of nuanced variants designed for both supervised fine-tuning and preference-based alignment. Experimental results show that LLMs trained with FGIV achieve up to +10.2% improvement on our fine-grained FGIV-Eval benchmark and up to +8.8% on existing benchmarks such as FollowBench and InfoBench. These findings highlight the value of FGIV in advancing instruction sensitivity and robustness in LLMs.

The fine-tuning of Large Language Models (LLMs) specialized in code generation has seen notable advancements through the use of open-domain coding queries. Despite the successes, existing methodologies like Evol-Instruct encounter performance limitations, impeding further enhancements in code generation tasks. This paper examines the constraints of existing prompt evolution techniques and introduces a novel approach, Instruction Fusion (IF). IF innovatively combines two distinct prompts through a hybridization process, thereby enhancing the evolution of training prompts for code LLMs. Our experimental results reveal that the proposed novel method effectively addresses the shortcomings of prior methods, significantly improving the performance of Code LLMs across five code generation benchmarks, namely HumanEval, HumanEval+, MBPP, MBPP+ and MultiPL-E, which underscore the effectiveness of Instruction Fusion in advancing the capabilities of LLMs in code generation.

The fine-tuning of Large Language Models (LLMs) specialized in code generation has seen notable advancements through the use of open-domain coding queries.Despite the successes, existing methodologies like Evol-Instruct encounter performance limitations, impeding further enhancements in code generation tasks.This paper examines the constraints of existing prompt evolution techniques and introduces a novel approach, Instruction Fusion (IF).IF innovatively combines two distinct prompts through a hybridization process, thereby enhancing the evolution of training prompts for code LLMs.Our experimental results reveal that the proposed novel method effectively addresses the shortcomings of prior methods, significantly improving the performance of Code LLMs across five code generation benchmarks, namely HumanEval, Hu-manEval+, MBPP, MBPP+ and MultiPL-E, which underscore the effectiveness of Instruction Fusion in advancing the capabilities of LLMs in code generation.

This paper introduces a novel unified representation of diffusion models for image generation and segmentation. Specifically, we use a colormap to represent entity-level masks, addressing the challenge of varying entity numbers while aligning the representation closely with the image RGB domain. Two novel modules, including the location-aware color palette and progressive dichotomy module, are proposed to support our mask representation. On the one hand, a location-aware palette guarantees the colors' consistency to entities' locations. On the other hand, the progressive dichotomy module can efficiently decode the synthesized colormap to high-quality entity-level masks in a depth-first binary search without knowing the cluster numbers. To tackle the issue of lacking large-scale segmentation training data, we employ an inpainting pipeline and then improve the flexibility of diffusion models across various tasks, including inpainting, image synthesis, referring segmentation, and entity segmentation. Comprehensive experiments validate the efficiency of our approach, demonstrating comparable segmentation mask quality to state-of-the-art and adaptability to multiple tasks.

Multimodal Sentiment Analysis aims to predict the sentiment of video content.Recent research suggests that multimodal sentiment analysis critically depends on learning a good representation of multimodal information, which should contain both modality-invariant representations that are consistent across modalities as well as modality-specific representations.In this paper, we propose ConFEDE, a unified learning framework that jointly performs contrastive representation learning and contrastive feature decomposition to enhance representation of multimodal information.It decomposes each of the three modalities of a video sample, including text, video frames, and audio, into a similarity feature and a dissimilarity feature, which are learned by a contrastive relation centered around text.We conducted extensive experiments on CH-SIMS, MOSI and MOSEI to evaluate various state-of-the-art multimodal sentiment analysis methods.Experimental results show that ConFEDE outperforms all baselines on these datasets on a range of metrics.

Nested Named Entity Recognition (NNER), as a subarea of Named Entity Recognition, has presented longstanding challenges to researchers. In NNER, one entity may be part of a larger entity, which can occur at multiple levels. These nested structures prevent traditional sequence labeling methods from properly recognizing all entities. While recent research has focused on designing better recognition methods for NNER in various languages, Chinese Nested Named Entity Recognition (CNNER) is still underdeveloped, largely due to a lack of freely available CNNER benchmarks. To support CNNER research, in this paper, we introduce ChiNesE, a CNNER dataset comprising 20,000 sentences from online passages in multiple domains and containing 117,284 entities that fall into 10 categories, of which 43.8% are nested named entities. Based on ChiNesE, we propose Mulco, a novel method that can recognize named entities in nested structures through multiple scopes. Each scope uses a scope-based sequence labeling method that predicts an anchor and the length of a named entity to recognize it. Experimental results show that Mulco outperforms state-of-the-art baseline methods with different recognition schemes on ChiNesE and ACE 2005 Chinese corpus.

Dense image segmentation tasks (e.g., semantic, panoptic) are useful for image editing, but existing methods can hardly generalize well in an in-the-wild setting where there are unrestricted image domains, classes, and image resolution & quality variations. Motivated by these observations, we construct a new entity segmentation dataset, with a strong focus on high-quality dense segmentation in the wild. The dataset contains images spanning diverse image domains and entities, along with plentiful high-resolution images and high-quality mask annotations for training and testing. Given the high-quality and -resolution nature of the dataset, we propose CropFormer which is designed to tackle the intractability of instance-level segmentation on high-resolution images. It improves mask prediction by fusing high-res image crops that provides more fine-grained image details and the full image. CropFormer is the first query-based Transformer architecture that can effectively fuse mask predictions from multiple image views, by learning queries that effectively associate the same entities across the full image and its crop. With CropFormer, we achieve a significant AP gain of 1.9 on the challenging entity segmentation task. Furthermore, CropFormer consistently improves the accuracy of traditional segmentation tasks and datasets. The dataset and code are released at http://luqi.info/entityv2.github.io/.

Single-stage Human-Object Interaction (HOI) detection methods have attracted considerable attention due to their high efficiency. Existing methods tend to concentrate the detection of humans and objects in one decoder without considering the differences between them, which causes tremendous pressure on a single decoder and affects the detection effect. This paper aims to decouple the detection decoder of humans and objects. In particular, we advocate and propose a novel human-object decoupling network (HOD) that divides the decoder into three tasks: human detection, object detection, and action classification. The network uses the random erasure training strategy to improve the model’s generalization ability and introduces pose features to handle the long-tailed problem. In addition, we design a pose fusion branch to alleviate the semantic gap between pose and HOI datasets. The experimental results suggest that our method achieves consistent improvements over the state-of-theart across different datasets, utilizing only image information. Specifically for HICO-Det, our method outperforms existing methods by a large margin, with a significant relative mAP gain of 7.8%. Our source code will be publicly available upon acceptance.

Despite the progress of image segmentation for accurate visual entity segmentation, completing the diverse requirements of image editing applications for different-level region-of-interest selections remains unsolved. In this paper, we propose a new task, All-Inclusive Multi-Level Segmentation (AIMS), which segments visual regions into three levels: part, entity, and relation (two entities with some semantic relationships). We also build a unified AIMS model through multi-dataset multi-task training to address the two major challenges of annotation inconsistency and task correlation. Specifically, we propose task complementarity, association, and prompt mask encoder for three-level predictions. Extensive experiments demonstrate the effectiveness and generalization capacity of our method compared to other state-of-the-art methods on a single dataset or the concurrent work on segment anything. We will make our code and training model publicly available.

Multi-modal video similarity evaluation is important for video recommendation systems such as video de-duplication, relevance matching, ranking, and diversity control. However, there still lacks a benchmark dataset that can support supervised training and accurate evaluation. In this paper, we propose the Tencent-MVSE dataset, which is the first benchmark dataset for the multi-modal video similarity evaluation task. The Tencent-MVSE dataset contains video pairs similarity annotations, and diverse metadata including Chinese title, automatic speech recognition (ASR) text, as well as human-annotated categories/tags. We provide a simple baseline with a multi-modal Transformer architecture to perform supervised multi-modal video similarity evaluation. We also explore pre-training strategies to make use of the unpaired data. The whole dataset as well as our baseline will be released to promote the development of the multi-modal video similarity evaluation. The dataset has been released in https://tencent-mvse.github.io/.

Headline generation is a task of generating an appropriate headline for a given article, which can be further used for machine-aided writing or enhancing the click-through ratio.Current works only use the article itself in the generation, but have not taken the writing style of headlines into consideration.In this paper, we propose a novel Seq2Seq model called CLH3G (Contrastive Learning enhanced Historical Headlines based Headline Generation) which can use the historical headlines of the articles that the author wrote in the past to improve the headline generation of current articles.By taking historical headlines into account, we can integrate the stylistic features of the author into our model, and generate a headline not only appropriate for the article, but also consistent with the author's style.In order to efficiently learn the stylistic features of the author, we further introduce a contrastive learning based auxiliary task for the encoder of our model.Besides, we propose two methods to use the learned stylisic features to guide both the pointer and the decoder during the generation.Experimental results show that historical headlines of the same user can improve the headline generation significantly, and both the contrastive learning module and the two style features fusion methods can further boost the performance.

Although conceptualization has been widely studied in semantics and knowledge representation, it is still challenging to find the most accurate concept terms to tag fast-growing social media content. This is partly attributed to the fact that most traditional knowledge bases contain general terms of the world, such as trees and cars, which are not interesting to users, and do not have the defining power for social media content. Another reason is that the intricate use of tense, negation and grammar in social media content may change the logic or emphasis of the content, thus focusing on different main ideas. In this paper, we present TAG, a high-quality concept matching dataset consisting of 10,000 labeled pairs of fine-grained concepts and web-styled natural language sentences, mined from open-domain social media content. The concepts we provide are the trending terms on social media and have the right granularity to define user interests, e.g., highly educated actors instead of just actors. In the meantime, TAG offers a concept graph which interconnects these fine-grained concepts and entities to provide contextual information. We evaluate a wide range of neural text matching models as well as pre-trained language models for the concept matching task on TAG, and point out their insufficiency to tag social media content to characterize its main idea. We further propose a novel graph-graph matching framework that demonstrates superior abstraction and generalization performance by better utilizing both the structural information in the concept graph and logic interactions between semantic units in the natural language sentence via syntactic dependency parsing.

Concepts naturally constitute an abstraction for fine-grained entities and knowledge in the open domain. They enable search engines and recommendation systems to enhance user experience by discovering high-level abstraction of a search query and the user intent behind it. In this paper, we study the problem of query conceptualization, which is to find the most appropriate matching concepts for any given search query from a large pool of pre-defined concepts. We propose a coarse-to-fine approach to first reduce the search space for each query through a shortlisting scheme and then identify the matching concepts using pre-trained language models, which are meta-tuned to our query-concept matching task. Our shortlisting scheme involves using a GRU-based Relevant Words Generator (RWG) to first expand and complete the context of the given query and then shortlisting the candidate concepts through a scoring mechanism based on word overlaps. To accurately identify the most appropriate matching concepts for a query, even when the concepts may have zero verbatim overlaps with the query, we meta-fine-tune a BERT pairwise text-matching model under the Reptile meta-learning algorithm, which achieves zero-shot transfer learning on the conceptualization problem. Our two-stage framework can be trained with data completely derived from a search click graph, without requiring any human labelling efforts. For evaluation, we have constructed a large click graph based on more than $7$ million instances of the click history recorded in Tencent QQ browser and performed the query conceptualization task based on a large ontology with $159,148$ unique concepts. Results from a range of evaluation methods, including an offline evaluation procedure on the click graph, human evaluation, online A/B testing and case studies, have demonstrated the superiority of our approach over a number of competitive pre-trained language models and fine-tuned neural network baselines.

Understanding what online users may pay attention to on the web is key to content recommendation and search services. These services will benefit from a highly structured and web-scale ontology of entities, concepts, events, topics and categories. While existing knowledge bases and taxonomies embody a large volume of entities and categories, we argue that they fail to discover properly grained concepts, events and topics in the language style of online users. Neither is a logically structured ontology maintained among these notions. In this paper, we present GIANT, a mechanism to construct a user-centered, web-scale, structured ontology, containing a large number of natural language phrases conforming to user attentions at various granularities, mined from the vast volume of web documents and search click logs. Various types of edges are also constructed to maintain a hierarchy in the ontology. We present our detailed techniques used in GIANT, and evaluate the proposed models and methods as compared to a variety of baselines, as well as deploy the resulted Attention Ontology in real-world applications, involving over a billion users, to observe its effect on content recommendation. The online performance of the ontology built by GIANT proves that it can significantly improve the click-through rate in news feeds recommendation.

Concepts embody the knowledge of the world and facilitate the cognitive processes of human beings. Mining concepts from web documents and constructing the corresponding taxonomy are core research problems in text understanding and support many downstream tasks such as query analysis, knowledge base construction, recommendation, and search. However, we argue that most prior studies extract formal and overly general concepts from Wikipedia or static web pages, which are not representing the user perspective.

We study the problem of search query inference from web documents, where a short, comprehensive natural language query is inferred from a long article. Search query generation or inference is of great value to search engines and recommenders in terms of locating potential target users and ranking content. Despite being closely related to other NLP tasks like abstract generation and keyword extraction, we point out that search query inference is, in fact, a new problem, in that the generated natural language query, which consists of a few words, is expected to be comprehensive enough to lead to the click-through of the corresponding document. Therefore, query generation requires an accurate inference of query words, as well as a deeper level of understanding on document semantic structures. Toward this end, we propose a novel generative model called the Graph-augmented Sequence to Attention (G-S2A) network. Adopting an Encoder-Decoder architecture, G-S2A incorporates a sentence-level Graph Convolutional Network (GCN), a keyword-level GCN, as well as a hierarchical recurrent neural network (RNN) into the encoder to generate structural document representations. An attentional Transformer decoder is then applied to combine different types of encoded features to generate a target query. On a query-document dataset from a real-world search engine, our model outperforms several neural generative models on a wide range of metrics.

In this study, the authors investigate the outage probability and bit‐error rate (BER) of distributed antenna system in downlink multicell environment with blanket transmission. Different from the most existing works, the variance of interference plus noise is treated as a random variable other than constant, and it is influenced by the short term fading when propagation pathloss and transmit power are given. From the perspective of information theory, the closed‐form and approximate analytical expressions of downlink outage probability and average BER in the cellular system are derived for no shadowing and shadowing scenarios, respectively. Extensive simulation results validate the theoretical analysis and demonstrate that the system performances can be significantly improved for cell‐edge users. Moreover, the proposed analytical method can obtain more accurate system performances.

Distributed antenna systems (DAS) have been paid much attention for low outage probability in future wireless communications, since it can reduce the radio transmission distance between the transmitter and the receiver. In this study, the authors present a study on DAS for two different transmission strategies: blanket transmission scheme and single selection transmission scheme (SSTS). The outage probabilities of the two schemes are re‐analysed over Rayleigh fading adopted non‐central limit theorem (NCLT) and compared with central limit theorem (CLT) method. Moreover, the authors derive analytical expressions of the outage probability for both no shadowing and shadowing scenarios, respectively. Then, an energy efficient power allocation strategy is proposed. Finally, extensive simulation results validate the theoretical analysis and demonstrate the proposed NCLT method can describe the system performances more accurately.

Given a collection of web images with the corresponding textual descriptions, in this paper, we propose a novel cross-domain learning method to classify these web multimedia objects by transferring the correlation knowledge among different information sources. Here, the knowledge is extracted from unlabeled objects through unsupervised learning and applied to perform supervised classification tasks. To mine more meaningful correlation knowledge, instead of using commonly used visual words in the traditional bag-of-visual-words (BoW) model, we discover higher level visual components (words and phrases) to incorporate the spatial and semantic information into our image representation model, i.e., bag-of-visual-phrases (BoP). By combining the enriched visual components with the textual words, we calculate the frequently co-occurring pairs among them to construct a cross-domain correlated graph in which the correlation knowledge is mined. After that, we investigate two different strategies to apply such knowledge to enrich the feature space where the supervised classification is performed. By transferring such knowledge, our cross-domain transfer learning method can not only handle large scale web multimedia objects, but also deal with the situation that the textual descriptions of a small portion of web images are missing. Empirical experiments on two different datasets of web multimedia objects are conducted to demonstrate the efficacy and effectiveness of our proposed cross-domain transfer learning method.

In this paper we study on haptic collaboration in a maze game over computer network. Two players located at separated places operate each haptic device to collaboratively finish the game task. Herein, a new collaboration mode of manipulation separate DOF control is proposed for the first time. Separate DOF control here means each player controls one DOF of an object or a task independently in collaborative virtual environment. Mutual guidance is also proposed which provides guidance force to each player. We setup an experiment to evaluate its efforts on cooperation performance and co-presence. Twelve participants did the experiment. The results revealed that this collaboration mode is effective. Ten of the twelve participants believed that they performed well in the experiment and thought the collaboration way was very interesting. The presented results motivated a new haptic collaboration mode in the fields of game design, education and cooperative assembly.

In this study, the authors investigate the downlink performance of distributed antenna systems (DAS) in multicell environment with blanket transmission. In most existing works, the interference plus noise is treated as Gaussian random variable with fixed variance by the central limit theorem. However, doing this will ignore the effect of the short-term fading on interference that is called as restrictive condition. To avoid the restrictive condition, the authors consider the variance of interference plus noise as a random variable with changeable variance that is influenced by the short-term fading when propagation pathloss and transmit power are given. From the perspective of information theoretic, the closed-form and approximate analytical expressions of downlink achievable throughput, outage probability and average bit error rate in the cellular system are derived for no-shadowing and shadowing scenarios, respectively. Extensive simulation results validate the theoretical analysis and demonstrate that the system performances can be significantly improved for cell-edge users. Moreover, the proposed analytical method can obtain more accurate system performances.