Leyi Pan, Shuchang Tao, Yunpeng Zhai, Zheyu Fu, Liancheng Fang, Minghua He, Lingzhe Zhang, Zhaoyang Liu, Bolin Ding, Aiwei Liu, Lijie Wen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In-context learning (ICL), a predominant trend in instruction learning, aims at enhancing the performance of large language models by providing clear task guidance and examples, improving their capability in task understanding and execution. This paper investigates ICL on Large Vision-Language Models (LVLMs) and explores the policies of multi-modal demonstration selection. Existing research efforts in ICL face significant challenges: First, they rely on pre-defined demonstrations or heuristic selecting strategies based on human intuition, which are usually inadequate for covering diverse task requirements, leading to sub-optimal solutions; Second, individually selecting each demonstration fails in modeling the interactions between them, resulting in information redundancy. Unlike these prevailing efforts, we propose a new exploration-exploitation reinforcement learning framework, which explores policies to fuse multi-modal information and adaptively select adequate demonstrations as an integrated whole. The framework allows LVLMs to optimize themselves by continually refining their demonstrations through self-exploration, enabling the ability to autonomously identify and generate the most effective selection policies for in-context learning. Experimental results verify the superior performance of our approach on four Visual Question-Answering (VQA) datasets, demonstrating its effectiveness in enhancing the generalization capability of few-shot LVLMs.

The rapid evolution of deep generative models poses a critical challenge to deepfake detection, as detectors trained on forgery-specific artifacts often suffer significant performance degradation when encountering unseen forgeries. While existing methods predominantly rely on spatial domain analysis, frequency domain operations are primarily limited to feature-level augmentation, leaving frequency-native artifacts and spatial-frequency interactions insufficiently exploited. To address this limitation, we propose a novel detection framework that integrates multiscale spatial-frequency analysis for universal deepfake detection. Our framework comprises three key components: (1) a local frequency feature extraction branch that combines block-wise discrete cosine transform with cascaded multi-scale convolutions to capture subtle spectral artifacts; (2) a global frequency feature extraction branch utilizing scale-invariant derivative accumulation to identify holistic forgery distribution patterns; and (3) a multi-stage cross-modal fusion mechanism that incorporates shallow-layer attention enhancement and deep-layer dynamic modulation to model spatial-frequency interactions. Extensive evaluations on widely adopted benchmarks demonstrate that our method outperforms state-of-the-art deepfake detection methods in both accuracy and generalizability.

This paper proposes a Semi-Attention Partition (SAP) method to learn well-aligned part features for occluded person re-identification (re-ID). Currently, the mainstream methods employ either external semantic partition or attention-based partition, and the latter manner is usually better than the former one. Under this background, this paper explores a potential that the weak semantic partition can be a good teacher for the strong attention-based partition. In other words, the attention-based student can substantially surpass its noisy semantic-based teacher, contradicting the common sense that the student usually achieves inferior (or comparable) accuracy. A key to this effect is: the proposed SAP encourages the attention-based partition of the (transformer) student to be partially consistent with the semantic-based teacher partition through knowledge distillation, yielding the so-called semi-attention. Such partial consistency allows the student to have both consistency and reasonable conflict with the noisy teacher. More specifically, on the one hand, the attention is guided by the semantic partition from the teacher. On the other hand, the attention mechanism itself still has some degree of freedom to comply with the inherent similarity between different patches, thus gaining resistance against noisy supervision. Moreover, we integrate a battery of well-engineered designs into SAP to reinforce their cooperation (e.g., multiple forms of teacher-student consistency), as well as to promote reasonable conflict (e.g., mutual absorbing partition refinement and a supervision signal dropout strategy). Experimental results confirm that the transformer student achieves substantial improvement after this semi-attention learning scheme, and produces new state-of-the-art accuracy on several standard re-ID benchmarks.

Efficient motion and appearance modeling are critical for vision-based Reinforcement Learning (RL). However, existing methods struggle to reconcile motion and appearance information within the state representations learned from a single observation encoder. To address the problem, we present Synergizing Interactive Motion-appearance Understanding (Simoun), a unified framework for vision-based RL Given consecutive observation frames, Simoun deliberately and interactively learns both motion and appearance features through a dual-path network architecture. The learning process collaborates with a structural interactive module, which explores the latent motion-appearance structures from the two network paths to leverage their complementarity. To promote sample efficiency, we further design a consistency-guided curiosity module to encourage the exploration of under-learned observations. During training, the curiosity module provides intrinsic rewards according to the consistency of environmental temporal dynamics, which are deduced from both motion and appearance network paths. Experiments conducted on Deep-Mind control suite and CARLA automatic driving benchmarks demonstrate the effectiveness of Simoun, where it performs favorably against state-of-the-art methods.

Vision-based reinforcement learning (RL) depends on discriminative representation encoders to abstract the observation states. Despite the great success of increasing CNN parameters for many supervised computer vision tasks, reinforcement learning with temporal-difference (TD) losses cannot benefit from it in most complex environments. In this paper, we analyze that the training instability arises from the oscillating self-overfitting of the heavy-optimizable encoder. We argue that serious oscillation will occur to the parameters when enforced to fit the sensitive TD targets, causing uncertain drifting of the latent state space and thus transmitting these perturbations to the policy learning. To alleviate this phenomenon, we propose a novel asymmetric interactive cooperation approach with the interaction between a heavy-optimizable encoder and a supportive light-optimizable encoder, in which both their advantages are integrated including the highly discriminative capability as well as the training stability. We also present a greedy bootstrapping optimization to isolate the visual perturbations from policy learning, where representation and policy are trained sufficiently by turns. Finally, we demonstrate the effectiveness of our method in utilizing larger visual models by first-person highway driving task CARLA and Vizdoom environments.

Decentralized multi-agent cooperative learning is a practical task due to the partially observed setting both in training and execution. Every agent learns to cooperate without access to the observations and policies of others. However, the decentralized training of multi-agent is of great difficulty due to non-stationarity, especially when other agents' policies are also in learning during training. To overcome this, we propose to learn a dynamic policy belief for each agent to predict the current policies of other agents and accordingly condition the policy of its own. To quickly adapt to the development of others' policies, we introduce a historical context to learn the belief inference according to a few recent action histories of other agents and a latent variational inference to model their policies by a learned distribution. We evaluate our method on the StarCraft II micro management task (SMAC) and demonstrate its superior performance in the decentralized training settings and comparable results with the state-of-the-art CTDE methods.

Occluded person re-identification (re-ID) is a challenging task as different human parts may become invisible in cluttered scenes, making it hard to match person images of different identities. Most existing methods address this challenge by aligning spatial features of body parts according to semantic information (e.g. human poses) or feature similarities but this approach is complicated and sensitive to noises. This paper presents Matching on Sets (MoS), a novel method that positions occluded person re-ID as a set matching task without requiring spatial alignment. MoS encodes a person image by a pattern set as represented by a `global vector’ with each element capturing one specific visual pattern, and it introduces Jaccard distance as a metric to compute the distance between pattern sets and measure image similarity. To enable Jaccard distance over continuous real numbers, we employ minimization and maximization to approximate the operations of intersection and union, respectively. In addition, we design a Jaccard triplet loss that enhances the pattern discrimination and allows to embed set matching into deep neural networks for end-to-end training. In the inference stage, we introduce a conflict penalty mechanism that detects mutually exclusive patterns in the pattern union of image pairs and decreases their similarities accordingly. Extensive experiments over three widely used datasets (Market1501, DukeMTMC and Occluded-DukeMTMC) show that MoS achieves superior re-ID performance. Additionally, it is tolerant of occlusions and outperforms the state-of-the-art by large margins for Occluded-DukeMTMC.

Domain adaptive person re-identification (re-ID) is a challenging task, especially when person identities in target domains are unknown. Existing methods attempt to address this challenge by transferring image styles or aligning feature distributions across domains, whereas the rich unlabeled samples in target domains are not sufficiently exploited. This paper presents a novel augmented discriminative clustering (AD-Cluster) technique that estimates and augments person clusters in target domains and enforces the discrimination ability of re-ID models with the augmented clusters. AD-Cluster is trained by iterative density-based clustering, adaptive sample augmentation, and discriminative feature learning. It learns an image generator and a feature encoder which aim to maximize the intra-cluster diversity in the sample space and minimize the intra-cluster distance in the feature space in an adversarial min-max manner. Finally, AD-Cluster increases the diversity of sample clusters and improves the discrimination capability of re-ID models greatly. Extensive experiments over Market-1501 and DukeMTMC-reID show that AD-Cluster outperforms the state-of-the-art with large margins.

RGB-Infrared (IR) cross-modality person re-identification (re-ID), which aims to search an IR image in RGB gallery or vice versa, is a challenging task due to the large discrepancy between IR and RGB modalities. Existing methods address this challenge typically by aligning feature distributions or image styles across modalities, whereas the very useful similarities among gallery samples of the same modality (i.e. intra-modality sample similarities) are largely neglected. This paper presents a novel similarity inference metric (SIM) that exploits the intra-modality sample similarities to circumvent the cross-modality discrepancy targeting optimal cross-modality image matching. SIM works by successive similarity graph reasoning and mutual nearest-neighbor reasoning that mine cross-modality sample similarities by leveraging intra-modality sample similarities from two different perspectives. Extensive experiments over two cross-modality re-ID datasets (SYSU-MM01 and RegDB) show that SIM achieves significant accuracy improvement but with little extra training as compared with the state-of-the-art.