Zihao Yi, Zhenqing Ling, Delong Zeng, Haohao Luo, Zhe Xu, Wei Liu, Jian Luan, Wanxia Cao, Ying Shen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Class-Incremental Learning (CIL) requires an artificial intelligence system to learn different tasks without class overlaps continually. To achieve CIL, some methods introduce the Pre-Trained Model (PTM) and leverage the generalized feature representation of PTM to learn downstream incremental tasks continually. However, the generalized feature representations of PTM are not adaptive and discriminative for these various incremental classes, which may be out of distribution for the pre-trained dataset. In addition, since the incremental classes cannot be learned at once, the class relationship cannot be constructed optimally, leading to undiscriminating feature representation for understream tasks. Thus, we propose a novel Pre-Trained Model-based Class-Incremental Learning (PTM-CIL) method to explore the potential of PTM and obtain optimal class relationships. Inspired by Neural Collapse theory, we introduce the frozen Equiangular Tight Frame classifier to construct optimal classifier structure for all seen classes, guiding the feature representation adaptation for downstream continual tasks. Specifically, Task-Related Adaptation is proposed to modulate the generalized feature representation to bridge the gap between the pre-trained dataset and various downstream datasets. Then, the Feature Compression Module is introduced to compress various features to the specific classifier weights, constructing the feature transfer pattern and satisfying the characteristic of Neural Collapse. Optimal Structural Alignment is designed to supervise the feature compression process, assisting in achieving optimal class relationships across different tasks. Sufficient experiments on seven datasets prove the effectiveness of our method.

Dialogue state tracking (DST) plays an essential role in task-oriented dialogue systems. However, user’s input may contain implicit information, posing significant challenges for DST tasks. Additionally, DST data includes complex information, which not only contains a large amount of noise unrelated to the current turn, but also makes constructing DST datasets expensive. To address these challenges, we introduce Intent-driven In-context Learning for Few-shot DST (IDIC-DST). By extracting user’s intent, we propose an Intent-driven Dialogue Information Augmentation module to augment the dialogue information, which can track dialogue states more effectively. Moreover, we mask noisy information from DST data and rewrite user’s input in the Intent-driven Examples Retrieval module, where we retrieve similar examples. We then utilize a pre-trained large language model to update the dialogue state using the augmented dialogue information and examples. Experimental results demonstrate that IDIC-DST achieves state-of-the-art performance in few-shot settings on MultiWOZ 2.1 and MultiWOZ 2.4 datasets.

Large foundation models face challenges in acquiring transferable, structured thinking abilities, especially when supervised with rigid templates or crowd-annotated instruction datasets. Unlike prior approaches, we focus on a thinking-centric data synthesis paradigm that enables models to evolve through self-generated, cognitively guided data. We propose MindGYM, a structured and scalable framework for question synthesis, composed of: (1) Cognitive Thinking Process Injection, which infuses high-level reasoning objectives to shape the model’s synthesis behavior; (2) Seed Single-Hop Question Synthesis, generating atomic questions from diverse semantic types to encourage broader thinking; and (3) Challenging Multi-Hop QA Synthesis, composing more complex multi-hop questions based on QA seeds for deeper reasoning. Detailed analysis shows that synthetic data generated by our method achieves 16.7% higher average quality and 67.91% lower quality variance compared to baseline sources, highlighting that both high-quality and self-contained data are essential for effective, thinking-oriented fine-tuning. MindGYM improves performance on six reasoning benchmarks, achieving gains of up to 16% on MathVision using only 400 data samples, and generalizable improvements across different model sizes and architectures. MindGYM underscores the viability of self-challenging mechanisms in refining large model capabilities while minimizing human intervention and resource demands.Code and data are released to promote data-centric research into self-evolving foundation models driven by their internal reasoning capabilities.

Weakly supervised natural language video localization (WS-NLVL) aims to retrieve the moment corresponding to a language query in a video with only video-language pairs utilized during training.Despite great success, existing WS-NLVL methods seldomly consider the complex temporal relations enclosing the language query (e.g., between the language query and sub-queries decomposed from it or its synonymous query), yielding illogical predictions.In this paper, we propose a novel plug-and-play method, Intrinsic Multilateral Logical Rules, namely IMLR, to exploit intrinsic temporal relations and logical rules for WS-NLVL.Specifically, we formalize queries derived from the original language query as the nodes of a directed graph, i.e., intrinsic temporal relation graph (ITRG), and the temporal relations between them as the edges.Instead of directly prompting a pre-trained language model, a relation-guided prompting method is introduced to generate ITRG in a hierarchical manner.We customize four types of multilateral temporal logical rules (i.e., identity, inclusion, synchronization, and succession) from ITRG and utilize them to train our model.Experiments demonstrate the effectiveness and superiority of our method on the Charades-STA and ActivityNet Captions datasets.

Emotional Support Conversation (ESC) systems are pivotal in providing empathetic interactions, aiding users through negative emotional states by understanding and addressing their unique experiences. In this paper, we tackle two key challenges in ESC: enhancing contextually relevant and empathetic response generation through dynamic demonstration retrieval, and advancing cognitive understanding to grasp implicit mental states comprehensively. We introduce Dynamic Demonstration Retrieval and Cognitive-Aspect Situation Understanding (D2RCU), a novel approach that synergizes these elements to improve the quality of support provided in ESCs. By leveraging in-context learning and persona information, we introduce an innovative retrieval mechanism that selects informative and personalized demonstration pairs. We also propose a cognitive understanding module that utilizes four cognitive relationships from the ATOMIC knowledge source to deepen situational awareness of help-seekers' mental states. Our supportive decoder integrates information from diverse knowledge sources, underpinning response generation that is both empathetic and cognitively aware. The effectiveness of D2RCU is demonstrated through extensive automatic and human evaluations, revealing substantial improvements over numerous state-of-the-art models, with up to 13.79% enhancement in overall performance of ten metrics. Our codes are available for public access to facilitate further research and development.

Class-Incremental Unsupervised Domain Adaptation (CI-UDA) requires the model can continually learn several steps containing unlabeled target domain samples, while the source-labeled dataset is available all the time. The key to tackling CI-UDA problem is to transfer domain-invariant knowledge from the source domain to the target domain, and preserve the knowledge of the previous steps in the continual adaptation process. However, existing methods introduce much biased source knowledge for the current step, causing negative transfer and unsatisfying performance. To tackle these problems, we propose a novel CI-UDA method named Pseudo-Label Distillation Continual Adaptation (PLDCA). We design Pseudo-Label Distillation module to leverage the discriminative information of the target domain to filter the biased knowledge at the class- and instance-level. In addition, Contrastive Alignment is proposed to reduce domain discrepancy by aligning the class-level feature representation of the confident target samples and the source domain, and exploit the robust feature representation of the unconfident target samples at the instance-level. Extensive experiments demonstrate the effectiveness and superiority of PLDCA. Code is available at code.

Weakly supervised temporal action localization (WSTAL), which aims to locate the time interval of actions in an untrimmed video with only video-level action labels, has attracted increasing research interest in the past few years. However, a model trained with such labels will tend to focus on segments that contributions most to the video-level classification, leading to inaccurate and incomplete localization results. In this paper, we tackle the problem from a novel perspective of relation modeling and propose a method dubbed Bilateral Relation Distillation (BRD). The core of our method involves learning representations by jointly modeling the relation at the category and sequence levels. Specifically, category-wise latent segment representations are first obtained by different embedding networks, one for each category. We then distill knowledge obtained from a pre-trained language model to capture the category-level relations, which is achieved by performing correlation alignment and category-aware contrast in an intra- and inter-video manner. To model the relations among segments at the sequence-level, we elaborate a gradient-based feature augmentation method and encourage the learned latent representation of the augmented feature to be consistent with that of the original one. Extensive experiments illustrate that our approach achieves state-of-the-art results on THUMOS14 and ActivityNet1.3 datasets.

Given an untrimmed video and a language query, Video Temporal Grounding (VTG) aims to locate the time interval in the video semantically relevant to the query. Existing fully-supervised VTG methods require accurate annotations of temporal boundary, which is time-consuming and expensive to obtain. On the other hand, weakly-supervised VTG methods where only paired videos and queries are available during training lag far behind the fully-supervised ones. In this paper, we introduce point supervision to narrow the performance gap with affordable annotating cost and propose a novel method dubbed Point-Supervised Video Temporal Grounding (PS-VTG). Specifically, an attention-based grounding network is first employed to obtain a language activation sequence (LAS). Then pseudo segment-level label is generated based on the LAS and the given point supervision to assist the training process. In addition, multi-level distribution calibration and cross-modal contrast are framed to obtain discriminative feature representations and precisely highlight the language-relevant video segments. Experiments on three benchmarks demonstrate that our method trained with point supervision can significantly outperform weakly-supervised approaches and achieve comparable performance with fully-supervised ones.

Video Temporal Grounding (VTG) aims to locate the time interval in a video that is semantically relevant to a language query. Existing VTG methods interact the query with entangled video features and treat the instances in a dataset independently. However, intra-video entanglement and inter-video connection are rarely considered in these methods, leading to mismatches between the video and language. To this end, we propose a novel method, dubbed Hierarchically Semantic Associating (HiSA), which aims to precisely align the video with language and obtain discriminative representation for further location regression. Specifically, the action factors and background factors are disentangled from adjacent video segments, enforcing precise multimodal interaction and alleviating the intra-video entanglement. In addition, cross-guided contrast is elaborately framed to capture the inter-video connection, which benefits the multimodal understanding to locate the time interval. Extensive experiments on three benchmark datasets demonstrate that our approach significantly outperforms the state-of-the-art methods. The project page is available at: https://github.com/zhexu1997/HiSA.

Zero-Shot Learning (ZSL) aims to recognize images belonging to unseen classes that are unavailable in the training process, while Generalized Zero-Shot Learning (GZSL) is a more realistic variant that both seen and unseen classes appear during testing. Most GZSL approaches achieve knowledge transfer based on the features of samples that inevitably contain information irrelevant to recognition, bringing negative influence for the performance. In this work, we propose a novel method, dubbed Disentangled-VAE, which aims to disentangle category-distilling factors and category-dispersing factors from visual as well as semantic features, respectively. In addition, a batch re-combining strategy on latent features is introduced to guide the disentanglement, encouraging the distilling latent features to be more discriminative for recognition. Extensive experiments demonstrate that our method outperforms the state-of-the-art approaches on four challenging benchmark datasets