Generalist Virtual Agents (GVAs) powered by Multimodal Large Language Models (MLLMs) exhibit impressive capabilities. However, their long-term learning is hampered by a core limitation: a failure to evolve beyond existing trajectories. This stems from memory systems that treat experiences as isolated fragments and rely on brittle semantic retrieval, preventing the synthesis of novel solutions from disparate knowledge. To address this, we introduce CA3Mem, a framework inspired by the human hippocampus that organizes experiences into a structured memory graph. Leveraging this graph, CA3Mem features two key innovations: 1) a generative memory recombination mechanism that synthesizes novel solutions to drive agent evolution, and 2) an associative retrieval algorithm that employs spreading activation to recall a comprehensive and contextually-aware set of experiences. Experiments on OSWorld and WebArena demonstrate that CA3Mem significantly enhances agent capabilities, leading to marked improvements in long-horizon planning, compositional generalization for novel tasks, and continuous adaptation from experience.

Jie Cao, Tianwei Lin, Bo Yuan, Rolan Yan, Hongyang He, Wenqiao Zhang, Juncheng Li, Dongping Zhang, Siliang Tang, Yueting Zhuang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Video Corpus Moment Retrieval (VCMR) requires models to efficiently retrieve and precisely locate specific moments relevant to natural language queries within a massive, untrimmed video corpus. However, existing discriminative approaches typically rely on shallow visual-textual feature matching mechanisms, which often struggle to capture fine-grained semantic differences. To address this limitation, we propose Video-GAR, a novel framework that reframes the conventional retrieval task from superficial matching to generative understanding, positing that the capability for query reconstruction evidences deep semantic comprehension. Specifically, Video-GAR orchestrates three synergistic components: To overcome the computational efficiency bottleneck, we construct a Bi-Mamba backbone that leverages the linear complexity of state-space models for efficient global context modeling. Building on these representations, we introduce a generation-augmented fusion module, in which a training-only decoder acts as a semantic regularizer to implicitly calibrate cross-modal attention without increasing inference overhead. Finally, to ensure fine-grained precision, we propose a boundary-aware localization strategy that integrates boundary modeling with categorical supervision. Experiments on two benchmark datasets demonstrate that Video-GAR significantly improves retrieval and localization accuracy while maintaining outstanding inference speed.

Knowledge editing enables multimodal large language models (MLLMs) to efficiently update outdated or incorrect information. However, existing benchmarks primarily emphasize cognitive-level modifications while lacking a focus on deeper meta-cognitive processes. To bridge this gap, we introduce CogEdit, a novel benchmark designed to evaluate MLLMs' meta-cognitive knowledge editing abilities across three levels: (1) Counterfactual-Driven Editing, assessing self-awareness of knowledge correctness changes; (2) Boundary Constraint Editing, ensuring appropriate generalization without unintended interference; and (3) Noise-Robust Editing, promoting reflective evaluation of uncertain information. To advance meta-cognitive editing, we propose MIND (Meta-cognitive INtegrated Dynamic Knowledge Editing), a framework that constructs a meta-knowledge memory for self-awareness, employs game-theoretic interactions to monitor knowledge activation, and incorporates label refinement for noise-robust updates. Extensive experiments show that MIND significantly outperforms existing cognitive editing approaches, achieving strong performance on both traditional and meta-cognitive knowledge editing benchmarks.

Large Language Models (LLMs) exhibit remarkable proficiency in understanding and managing text-based tasks.Many works try to transfer these capabilities to the video domain, which are referred to as Video-LLMs. However, current Video-LLMs can only grasp the coarse-grained semantics and are unable to efficiently handle tasks involving the comprehension or localization of specific video segments. To address these challenges, we propose Momentor, a Video-LLM designed to perform fine-grained temporal understanding tasks. To facilitate the training of Momentor, we develop an automatic data generation engine to build Moment-10M, a large-scale video instruction dataset with segment-level instruction data. Building upon the foundation of the previously published Momentor and the Moment-10M dataset, we further extend this work by introducing a Spatio-Temporal Token Consolidation (STTC) method, which can merge redundant visual tokens spatio-temporally in a parameter-free manner, thereby significantly promoting computational efficiency while preserving fine-grained visual details. We integrate STTC with Momentor to develop Momentor++ and validate its performance on various benchmarks. Momentor demonstrates robust capabilities in fine-grained temporal understanding and localization. Further, Momentor++ excels in efficiently processing and analyzing extended videos with complex events, showcasing marked advancements in handling extensive temporal contexts.

Prompt tuning, a recently emerging paradigm, adapts vision-language pre-trained models to new tasks efficiently by learning "soft prompts" for frozen models. However, in few-shot scenarios, its effectiveness is limited by sensitivity to the initialization and the time-consuming search for optimal initialization, hindering rapid adaptation. Additionally, prompt tuning risks reducing the models' generalizability due to overfitting on scarce training samples. To overcome these challenges, we introduce a novel Gradient-RegulAted Meta-prompt learning (GRAM) framework that jointly meta-learns an efficient soft prompt initialization for better adaptation and a lightweight gradient regulating function for strong cross-domain generalizability in a meta-learning paradigm using only the weakly labeled image-text pre-training data. This is achieved through a Cross-Modal Hierarchical Clustering algorithm that organizes extensive image-text data into a structured hierarchy, facilitating robust meta-learning across diverse domains. Rather than designing a specific prompt tuning method, our GRAM can be easily incorporated into various prompt tuning methods in a model-agnostic way and bring about consistent improvement for them. Further, we consider a more practical but challenging setting: test-time prompt tuning with only unlabeled test samples and propose an improved structure-induced gradient regulating function to leverage the structured semantics of the meta-learning data for zero-shot generalization. This novel approach exploits the hierarchically clustered meta-learning data to model relationships between test-time data and meta-learning prototypes, facilitating the transfer of invariant knowledge without explicit annotations. Meanwhile, we introduce a structure complexity-informed strategy for adaptively constructing meta-training tasks and generating prototypes, which fully considers the diverse semantics within hierarchical clusters of different complexities. Comprehensive experiments demonstrate the state-of-the-art few- and zero-shot generalizability of our method.

While Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) effectively address resource constraints during fine-tuning, their performance often falls short, especially in multidimensional task scenarios. To address this issue, one straightforward solution is to introduce task-specific LoRA as domain experts, leveraging the modeling of multiple capabilities of experts and thus enhancing the general capability of multi-task learning.Although promising, these additional components often add complexity to the training and inference process, contravening the efficiency that PEFT is designed to deliver. Considering this, we introduce an innovative PEFT method, TeamLoRA, consisting of a collaboration and competition module for LoRA experts, thus achieving the right balance of effectiveness and efficiency:(i) For collaboration, we introduce a novel knowledge sharing and organization mechanism designed to optimize hierarchical learning while enhancing the efficiency of model training and inference.(ii) For competition, we propose leveraging a game-theoretic interaction mechanism for experts, encouraging experts to transfer their domain-specific knowledge while facing diverse downstream tasks, thus enhancing the performance.By doing so, TeamLoRA elegantly connects the experts as a “Team” with internal collaboration and competition, enabling a faster and more accurate PEFT paradigm. Meanwhile, we curate a Comprehensive Multi-Task Evaluation (CME) benchmark to thoroughly assess the capability of multi-task learning. Experiments conducted on our CME and other benchmarks indicate the effectiveness and efficiency of TeamLoRA. Our project is available at https://github.com/DCDmllm/TeamLoRA.

The decoding strategies widely used in large language models (LLMs) today are Top-$p$ Sampling and Top-$k$ Sampling, both of which are methods situated between greedy decoding and random sampling. Inspired by the concept of loss aversion from prospect theory in behavioral economics, and the endowment effect as highlighted by Richard H. Thaler, the 2017 Nobel Memorial Prize in Economic Sciences — particularly the principle that "the negative utility of an equivalent loss is approximately twice the positive utility of a comparable gain" — we have developed a new decoding strategy called Loss Sampling. We have demonstrated the effectiveness and validity of our method on several LLMs, including Llama-2, Llama-3 and Mistral. Our approach improves text quality by 4-30% across four pure text tasks while maintaining diversity in text generation. Furthermore, we also extend our method to multimodal large models (LMs) and Beam Search, demonstrating the effectiveness and versatility of Loss Sampling with improvements ranging from 1-10%.

Inspired by human cognitive behavior, we introduce visual modality to enhance the performance of pure text-based question-answering tasks with the development of multimodal models. However, obtaining corresponding images through manual annotation often entails high costs. Faced with this challenge, an intuitive strategy is to use search engines or use web scraping techniques to automatically obtain relevant image information. However, the images obtained by this strategy may be of low quality and may not match the context of the original task, which could fail to improve or even decrease performance on downstream tasks. In this paper, we propose a novel framework named “ITERATE”, aimed at retrieving and optimizing the quality of images to improve the alignment between text and images. Inspired by evolutionary algorithms in reinforcement learning and driven by the synergy of large language models (LLMs) and multimodal models, ITERATE employs a series of strategic actions such as filtering, optimizing, and retrieving to acquire higher quality images, and repeats this process over multiple generations to enhance the quality of the entire image cluster. Our experimental results on the ScienceQA, ARC-Easy, and OpenDataEval datasets also verify the effectiveness of our method, showing improvements of 3.5%, 5%, and 7%, respectively.

Recent endeavors in Multimodal Large Language Models (MLLMs) aim to unify visual comprehension and generation by combining LLM and diffusion models, the state-of-the-art in each task, respectively. Existing approaches rely on spatial visual tokens, where image patches are encoded and arranged according to a spatial order (e.g., raster scan). However, we show that spatial tokens lack the recursive structure inherent to languages, hence form an impossible language for LLM to master. In this paper, we build a proper visual language by leveraging diffusion timesteps to learn discrete, recursive visual tokens. Our proposed tokens recursively compensate for the progressive attribute loss in noisy images as timesteps increase, enabling the diffusion model to reconstruct the original image at any timestep. This approach allows us to effectively integrate the strengths of LLMs in autoregressive reasoning and diffusion models in precise image generation, achieving seamless multimodal comprehension and generation within a unified framework. Extensive experiments show that we achieve a new SOTA for multimodal comprehension and generation simultaneously compared with other MLLMs. Project Page: https://DDT-LLaMA.github.io/.

Video Large Language Models (Video-LLMs) have recently shown strong performance in basic video understanding tasks, such as captioning and coarse-grained question answering, but struggle with compositional reasoning that requires multi-step spatio-temporal inference across object relations, interactions, and events. The hurdles to enhancing this capability include extensive manual labor, the lack of spatio-temporal compositionality in existing training data and the absence of explicit reasoning supervision. In this paper, we propose STEP, a novel graph-guided self-training method that enables Video-LLMs to generate reasoning-rich fine-tuning data from any raw videos to improve itself. Specifically, we first induce Spatio-Temporal Scene Graph (STSG) representation of diverse videos to capture fine-grained, multi-granular video semantics. Then, the STSGs guide the derivation of multi-step reasoning Question-Answer (QA) data with Chain-of-Thought (CoT) rationales. Both answers and rationales are integrated as training objective, aiming to enhance model’s reasoning abilities by supervision over explicit reasoning steps. Experimental results demonstrate the effectiveness of STEP across models of varying scales, with a significant 21.3% improvement in tasks requiring three or more reasoning steps. Furthermore, it achieves superior performance with a minimal amount of self-generated rationale-enriched training samples in both compositional reasoning and comprehensive understanding benchmarks, highlighting the broad applicability and vast potential.

Large Multimodal Models (LMMs) have demonstrated impressive capabilities in multimodal understanding and generation, pushing forward advancements in text-to-image generation. However, achieving accurate text-image alignment for LMMs, particularly in compositional scenarios, remains challenging. Existing approaches, such as layout planning for multi-step generation and learning3 from human feedback or AI feedback, depend heavily on prompt engineering, costly human annotations, and continual upgrading, limiting flexibility and scalability. In this work, we introduce a model-agnostic iterative self-improvement framework (SILMM) that can enable LMMs to provide helpful and scalable self-feedback and optimize text-image alignment via Direct Preference Optimization (DPO). DPO can readily applied to LMMs that use discrete visual tokens as intermediate image representations; while it is less suitable for LMMs with continuous visual features, as obtaining generation probabilities is challenging. To adapt SILMM to LMMs with continuous features, we propose a diversity mechanism to obtain diverse representations and a kernel-based continuous DPO for alignment. Extensive experiments on three compositional text-to-image generation benchmarks validate the effectiveness and superiority of SILMM, showing improvements exceeding 30% on T2I-CompBench++ and around 20% on DPG-Bench. The code is available at https://silmm.github.io/.

The latest emerged 4D Panoptic Scene Graph (4D-PSG) provides an advanced-ever representation for comprehensively modeling the dynamic 4D visual real world. Unfortunately, current pioneering 4D-PSG research can primarily suffer from data scarcity issues severely, as well as the resulting out-of-vocabulary problems; also, the pipeline nature of the benchmark generation method can lead to suboptimal performance. To address these challenges, this paper investigates a novel framework for 4D-PSG generation that leverages rich 2D visual scene annotations to enhance 4D scene learning. First, we introduce a 4D Large Language Model (4D-LLM) integrated with a 3D mask decoder for end-to-end generation of 4D-PSG. A chained SG inference mechanism is further designed to exploit LLMs’ open-vocabulary capabilities to infer accurate and comprehensive object and relation labels iteratively. Most importantly, we propose a 2D-to-4D visual scene transfer learning framework, where a spatial-temporal scene transcending strategy effectively transfers dimension-invariant features from abundant 2D SG annotations to 4D scenes, effectively compensating for data scarcity in 4D-PSG. Extensive experiments on the benchmark data demonstrate that we strikingly outperform baseline models by a large margin, highlighting the effectiveness of our method. The project page is https://sqwu.top/PSG-4D-LLM/.

Instruction-based image editing aims to modify specific image elements with natural language instructions. However, current models in this domain often struggle to execute complex user instructions accurately, as they are trained on low-quality data with limited editing types. We present AnyEdit, a comprehensive multi-modal instruction editing dataset, comprising 2.5 million high-quality editing pairs spanning over 20 editing types and five domains. We ensure the diversity and quality of the AnyEdit collection through three aspects: initial data diversity, adaptive editing process, and automated selection of editing results. Using the dataset, we further train a novel AnyEdit Stable Diffusion with task-aware routing and learnable task embedding for unified image editing. Comprehensive experiments on three benchmark datasets show that AnyEdit consistently boosts the performance of diffusion-based editing models. This presents prospects for developing instruction-driven image editing models that support human creativity.

Recent advancements in reward signal usage for Large Language Models (LLMs) are remarkable. However, significant challenges exist when transitioning reward signal to the multimodal domain, including labor-intensive annotations, over-reliance on one-step rewards, and inadequate evaluation. To address these issues, we propose SVIP, a novel approach to train a step-level multi-dimensional Chain-of-Thought (CoT) reward model automatically. It generates code for solving visual tasks and transforms the analysis of code blocks into the evaluation of CoT step as training samples. Then, we train SVIP-Reward model using a multi-head attention mechanism called TriAtt-CoT. The advantages of SVIP-Reward are evident throughout the entire process of MLLM. We also introduce a benchmark for CoT reward model training and testing. Experimental results demonstrate that SVIP-Reward improves MLLM performance across training and inference-time scaling, yielding better results on benchmarks while reducing hallucinations and enhancing reasoning ability.

Digital agents are increasingly employed to automate tasks in interactive digital environments such as web pages, software applications, and operating systems. While text-based agents built on Large Language Models (LLMs) often require frequent updates due to platform-specific APIs, visual agents leveraging Multimodal Large Language Models (MLLMs) offer enhanced adaptability by interacting directly with Graphical User Interfaces (GUIs). However, these agents face significant challenges in visual perception, particularly when handling high-resolution, visually complex digital environments. This paper introduces Iris, a foundational visual agent that addresses these challenges through two key innovations: Information-Sensitive Cropping (ISC) and Self-Refining Dual Learning (SRDL). ISC dynamically identifies and prioritizes visually dense regions using a edge detection algorithm, enabling efficient processing by allocating more computational resources to areas with higher information density. SRDL enhances the agent's ability to handle complex tasks by leveraging a dual-learning loop, where improvements in referring (describing UI elements) reinforce grounding (locating elements) and vice versa, all without requiring additional annotated data. Empirical evaluations demonstrate that Iris achieves state-of-the-art performance across multiple benchmarks with only 850K GUI annotations, outperforming methods using 10x more training data. These improvements further translate to significant gains in both web and OS agent downstream tasks.

Instruction tuning fine-tunes pre-trained Multi-modal Large Language Models (MLLMs) to handle real-world tasks. However, the rapid expansion of visual instruction datasets introduces data redundancy, leading to excessive computational costs. We propose a collaborative framework, DataTailor, which leverages three key principles--informativeness, uniqueness, and representativeness--for effective data selection. We argue that a valuable sample should be informative of the task, non-redundant, and represent the sample distribution (i.e., not an outlier). We further propose practical ways to score against each principle, which automatically adapts to a given dataset without tedious hyperparameter tuning. Comprehensive experiments on various benchmarks demonstrate that DataTailor achieves 101.3% of the performance of full-data fine-tuning with only 15% of the data, significantly reducing computational costs while maintaining superior results. This exemplifies the "Less is More" philosophy in MLLM development. The code and data is available in this \href{https://github.com/Yuqifan1117/DataTailor}{URL}.

The Multimodal Large Language Model (MLLM) is currently experiencing rapid growth, driven by the advanced capabilities of language-based LLMs. Unlike their specialist predecessors, existing MLLMs are evolving towards a Multimodal Generalist paradigm. Initially limited to understanding multiple modalities, these models have advanced to not only comprehend but also generate across modalities. Their capabilities have expanded from coarse-grained to fine-grained multimodal understanding and from supporting singular modalities to accommodating a wide array of or even arbitrary modalities. To assess the capabilities of various MLLMs, a diverse array of benchmark test sets has been proposed. This leads to a critical question: Can we simply assume that higher performance across tasks indicates a stronger MLLM capability, bringing us closer to human-level AI? We argue that the answer is not as straightforward as it seems. In this project, we introduce an evaluation framework to delineate the capabilities and behaviors of current multimodal generalists. This framework, named General-Level, establishes 5-scale levels of MLLM performance and generality, offering a methodology to compare MLLMs and gauge the progress of existing systems towards more robust multimodal generalists and, ultimately, towards AGI (Artificial General Intelligence). Central to our framework is the use of Synergy as the evaluative criterion, categorizing capabilities based on whether MLLMs preserve synergy across comprehension and generation, as well as across multimodal interactions. To evaluate the comprehensive abilities of various generalists, we present a massive multimodal benchmark, General-Bench, which encompasses a broader spectrum of skills, modalities, formats, and capabilities, including over 700 tasks and 325,800 instances. The evaluation results that involve over 100 existing state-of-the-art MLLMs uncover the capability rankings of generalists, highlighting the challenges in reaching genuine AI. We expect this project to pave the way for future research on next-generation multimodal foundation models, providing a robust infrastructure to accelerate the realization of AGI. Project Page: https://generalist.top/, Leaderboard: https://generalist.top/leaderboard/, Benchmark: https://huggingface.co/General-Level/.

As multimodal large language models (MLLMs) advance, MLLM-based virtual agents have demonstrated remarkable performance. However, existing benchmarks face significant limitations, including uncontrollable task complexity, extensive manual annotation, and a lack of multidimensional evaluation. In response to these challenges, we introduce OmniBench, a self-generating, graph-based benchmark with an automated pipeline for synthesizing tasks of controllable complexity through subtask composition. To evaluate the diverse capabilities of virtual agents on the graph, we further present OmniEval, a multidimensional evaluation framework that includes subtask-level evaluation, graph-based metrics, and comprehensive tests across 10 capabilities. Our synthesized dataset contains 36k graph-structured tasks across 20 scenarios, achieving a 91% human acceptance rate. Training on our graph-structured data shows that it improves generalization across environments. We conduct multidimensional evaluations for virtual agents, revealing their performance across various capabilities and paving the way for future advancements. Our project is available at https://omni-bench.github.io.

The development of Generalist Virtual Agents (GVAs) has shown significant promise in autonomous task execution. However, current training paradigms face critical limitations, including reliance on outcome supervision and labor-intensive human annotations. To address these challenges, we propose Similar, a step-wise multi-dimensional generalist reward model, which offers fine-grained signals for agent training and can choose better actions for inference-time scaling. Specifically, we begin by systematically defining five dimensions for evaluating agent actions. Building on this framework, we design an MCTS-P algorithm to automatically collect and annotate step-wise, five-dimensional agent execution data. Using this data, we train Similar with our crafted Triple-M strategy. Furthermore, we introduce the first benchmark in the virtual agent domain for step-wise, multi-dimensional reward model training and evaluation, named SRM. This benchmark consists of two components: SRMTrain, which serves as the training set for Similar, and SRMEval, a manually selected test set for evaluating the reward model. Experimental results demonstrate that Similar, through its step-wise, multi-dimensional assessment and synergistic gain, provides GVAs with effective intermediate signals during both training and inference-time scaling. The code is available at https://github.com/antgroup/Similar.

Online form filling is a common yet labor-intensive task involving extensive keyboard and mouse interactions. Despite the long-standing vision of automating this process with ''one click,'' existing tools remain largely rule-based and lack generalizable, generative capabilities. Recent advances in Multimodal Large Language Models (MLLMs) have enabled promising agents for GUI-related tasks in general-purpose scenarios. However, they struggle with the unique challenges of form filling, such as flexible layouts and the difficulty of aligning textual instructions with on-screen fields. To bridge this gap, we formally define the form-filling task and propose FormFactory-an interactive benchmarking suite comprising a web-based interface, backend evaluation module, and carefully constructed dataset. Our benchmark covers diverse real-world scenarios, incorporates various field formats, and simulates high-fidelity form interactions. We conduct a comprehensive evaluation of state-of-the-art MLLMs and observe that no model surpasses 5% accuracy, underscoring the inherent difficulty of the task. These findings also reveal significant limitations in current models' visual layout reasoning and field-value alignment abilities.

As researchers continue to optimize AI agents for more effective task execution within operating systems, they often overlook a critical security concern: the ability of these agents to detect ''impostors'' within their environment. Through an analysis of the agents' operational context, we identify a significant threat-attackers can disguise malicious attacks as environmental elements, injecting active disturbances into the agents' execution processes to manipulate their decision-making. We define this novel threat as the Active Environment Injection Attack (AEIA). Focusing on the interaction mechanisms of the Android OS, we conduct a risk assessment of AEIA and identify two critical security vulnerabilities: (1) Adversarial content injection in multimodal interaction interfaces, where attackers embed adversarial instructions within environmental elements to mislead agent decision-making; and (2) Reasoning gap vulnerabilities in the agent's task execution process, which increase susceptibility to AEIA attacks during reasoning. To evaluate the impact of these vulnerabilities, we propose AEIA-MN, an attack scheme that exploits interaction vulnerabilities in mobile operating systems to assess the robustness of MLLM-based agents. Experimental results show that even advanced MLLMs are highly vulnerable to this attack, achieving a maximum attack success rate of 93% on the AndroidWorld benchmark by combining two vulnerabilities.

Multimodal Large Language Models (MLLMs) have garnered significant attention for their strong visual-semantic understanding. Most existing chart benchmarks evaluate MLLMs' ability to parse information from charts to answer questions. However, they overlook the inherent output biases of MLLMs, where models rely on their parametric memory to answer questions rather than genuinely understanding the chart content. To address this limitation, we introduce a novel Chart Hypothetical Question Answering (HQA) task, which imposes assumptions on the same question to compel models to engage in counterfactual reasoning based on the chart content. Furthermore, we introduce HAI, a human-AI interactive data synthesis approach that leverages the efficient text-editing capabilities of LLMs alongside human expert knowledge to generate diverse and high-quality HQA data at a low cost. Using HAI, we construct Chart-HQA, a challenging benchmark synthesized from publicly available data sources. Evaluation results on 18 MLLMs of varying model sizes reveal that current models face significant generalization challenges and exhibit imbalanced reasoning performance on the HQA task. Our codebase and newly generated datasets are available at https://github.com/chenxn2020/Chart-HQA.

Short-video misinformation detection has attracted wide attention in the multi-modal domain, aiming to accurately identify the misinformation in the video format accompanied by the corresponding audio. Despite significant advancements, current models in this field, trained on particular domains (source domains), often exhibit unsatisfactory performance on unseen domains (target domains) due to domain gaps. To effectively realize such domain generalization on the short-video misinformation detection task, we propose deep insights into the characteristics of different domains: (1) The detection on various domains may mainly rely on different modalities (i.e., mainly focusing on videos or audios). To enhance domain generalization, it is crucial to achieve optimal model performance on all modalities simultaneously. (2) For some domains focusing on cross-modal joint fraud, a comprehensive analysis relying on cross-modal fusion is necessary. However, domain biases located in each modality (especially in each frame of videos) will be accumulated in this fusion process, which may seriously damage the final identification of misinformation. To address these issues, we propose a new DOmain generalization model via ConsisTency and invariance learning for shORt-video misinformation detection (named DOCTOR), which contains two characteristic modules: (1) We involve the cross-modal feature interpolation to map multiple modalities into a shared space and the interpolation distillation to synchronize multi-modal learning; (2) We design the diffusion model to add noise to retain core features of multi modal and enhance domain invariant features through cross-modal guided denoising. Extensive experiments demonstrate the effectiveness of our proposed DOCTOR model. Our code is publicly available at https://github.com/ghh1125/DOCTOR.

Multimodal Industrial Anomaly Detection (MIAD)-fusing 3D point clouds and 2D RGB for product defect detection-is critical to quality inspection. However, existing MIAD methods assume all modalities are available and paired, overlooking real-scenario modality-missing and risking overfitting to incomplete data. To address these, we conduct the first comprehensive study on Modality-Incomplete Industrial Anomaly Detection (MIIAD) and establish MIIAD Bench , a benchmark covering diverse missing settings. Meanwhile, we propose RADAR, a robust two-stage Robust modAlity-instructive fusing & Detecting frAmewoRk. RADAR integrates i) a Modality-Incomplete Instruction mechanism-guiding the multimodal Transformer to focus more on available modal info, and ii) a Double-Pseudo Hybrid Module to highlight unique modality combinations and reduce overfitting. Our results show RADAR outperforms prior methods markedly on MIIAD Bench.

Semantic retrieval is crucial for modern applications yet remains underexplored in current research. Existing datasets are limited to single languages, single images, or singular retrieval conditions, often failing to fully exploit the expressive capacity of visual information as evidenced by maintained performance when images are replaced with captions. However, practical retrieval scenarios frequently involve interleaved multi-condition queries with multiple images. Hence, this paper introduces MERIT, the first multilingual dataset for interleaved multi-condition semantic retrieval, comprising 320,000 queries with 135,000 products in 5 languages, covering 7 distinct product categories. Extensive experiments on MERIT identify existing models's critical limitation: focusing solely on global semantic information while neglecting specific conditional elements in queries. Consequently, we propose Coral, a novel fine-tuning framework that adapts pre-trained MLLMs by integrating embedding reconstruction to preserve fine-grained conditional elements and contrastive learning to extract comprehensive global semantics. Experiments demonstrate that Coral achieves a 45.9% performance improvement over conventional approaches on MERIT, with strong generalization capabilities validated across 8 established retrieval benchmarks. Collectively, our contributions—a novel dataset, identification of critical limitations in existing approaches, and an innovative fine-tuning framework—establish a foundation for future research in interleaved multi-condition semantic retrieval. **Data & Code**: [MERIT-2025.github.io](https://merit-2025.github.io)

The rapid development of Multimodal Large Language Models (MLLMs) poses increasing demands on the diversity and complexity of multimodal datasets. Yet manual annotation pipelines can no longer keep pace. Existing augmentation methods often follow fixed rules and lack verifiable control over sample diversity and reasoning complexity. To address this, we introduce Scalable COunterfactual Program Evolution (SCOPE), a framework that uses symbolic Visual Programming to guide program evolution via counterfactual reasoning. SCOPE performs the three steps of counterfactual inference: (1) Abduction, by generating verifiable programs to model reasoning associations; (2) Action, by intervening on program structure along three axes—reasoning path, visual context, and cross-instance composition; and (3) Prediction, by categorizing evolved instances by difficulty, structure, and input multiplicity. Based on this process, we build SCOPE-Train and SCOPE-Test, evolving benchmarks with expert validation. To support training, we propose MAP, a curriculum learning strategy that aligns model capacity with sample difficulty. Experiments show that SCOPE improves reasoning performance, exposes model blind spots, and enhances visual dialog capabilities.

Recent endeavors in Multimodal Large Language Models (MLLMs) aim to unify visual comprehension and generation. However, these two capabilities remain largely independent, as if they are two separate functions encapsulated within the same model. Consequently, visual comprehension does not enhance visual generation, and the reasoning mechanisms of LLMs have not been fully integrated to revolutionize image generation. In this paper, we propose to enable the collaborative co-evolution of visual comprehension and generation, advancing image generation into an iterative introspective process. We introduce a two-stage training approach: supervised fine-tuning teaches the MLLM with the foundational ability to generate genuine CoT for visual generation, while reinforcement learning activates its full potential via an exploration-exploitation trade-off. Ultimately, we unlock the Aha moment in visual generation, advancing MLLMs from text-to-image tasks to unified image generation. Extensive experiments demonstrate that our model not only excels in text-to-image generation and image editing, but also functions as a superior image semantic evaluator with enhanced visual comprehension capabilities. Project Page: \url{https://janus-pro-r1.github.io}.

Recent advances in reinforcement learning (RL) methods such as Grouped Relative Policy Optimization (GRPO) have strengthened the reasoning capabilities of Large Vision-Language Models (LVLMs). However, due to the inherent entanglement between visual and textual modalities, applying GRPO to LVLMs often leads to reward convergence across different responses to the same sample as training progresses, hindering effective gradient updates and causing the enhancement of chain-of-thought reasoning to stagnate or even collapse. To address this issue, we propose a progressive instruction evolution framework, EvolvedGRPO, to gradually generate more complex questions via editing instructions in an adversarial way, progressively aligned with the model’s evolving capabilities. Specifically, we design two instruction editing strategies across modalities, incorporating incrementally increasing editing instructions and RL-based adversarial data augmentation to improve the effectiveness of model training. To address GRPO's limitations on overly difficult problems, we first train on basic subproblem versions of complex multi-modal questions in both the visual and textual modalities, progressively increasing difficulty to enable prefix-style process rewards, effectively combining the strengths of both process rewards and group-wise relative rewards. Finally, EvolvedGRPO achieves state-of-the-art performance among open-source RL models on multi-modal reasoning tasks, even approaching the closed-source GPT-4o in reasoning capabilities, and demonstrates better performance on unseen LVLM general benchmarks. The Code for EvolvedGRPO is available at https://github.com/SHENZHEBEI/EvolvedGRPO.

Many studies have aimed to improve Transformer model efficiency using low-rank-based methods that compress sequence length with predetermined or learned compression matrices. However, these methods fix compression coefficients for tokens in the same position during inference, ignoring sequence-specific variations. They also overlook the impact of hidden state dimensions on efficiency gains. To address these limitations, we propose dynamic bilinear low-rank attention (DBA), an efficient and effective attention mechanism that compresses sequence length using input-sensitive dynamic compression matrices. DBA achieves linear time and space complexity by jointly optimizing sequence length and hidden state dimension while maintaining state-of-the-art performance. Specifically, we demonstrate through experiments and the properties of low-rank matrices that sequence length can be compressed with compression coefficients dynamically determined by the input sequence. In addition, we illustrate that the hidden state dimension can be approximated by extending the Johnson-Lindenstrauss lemma, thereby introducing only a small amount of error. DBA optimizes the attention mechanism through bilinear forms that consider both the sequence length and hidden state dimension. Moreover, the theoretical analysis substantiates that DBA excels at capturing high-order relationships in cross-attention problems. Experimental results across different tasks with varied sequence length conditions demonstrate that DBA achieves state-of-the-art performance compared to several robust baselines. DBA also maintains higher processing speed and lower memory usage, highlighting its efficiency and effectiveness across diverse applications.

Referring Image Segmentation (RIS) has been extensively studied over the past decade, leading to the development of advanced algorithms. However, there has been a lack of research investigating how existing algorithms should be benchmarked with complex language queries, which include more informative descriptions of surrounding objects and backgrounds (e.g., the black car vs. the black car is parking on the road and beside the bus ). Given the significant improvement in the semantic understanding capability of large pre-trained models, it is crucial to take a step further in RIS by incorporating complex language that resembles real-world applications. To close this gap, building upon the existing RefCOCO and Visual Genome datasets, we propose a new RIS benchmark with complex queries, namely RIS-CQ . The RIS-CQ dataset is of high quality and large scale, which challenges the existing RIS with enriched, specific, and informative queries, and enables a more realistic scenario of RIS research. Besides, we present a niche targeting method to better task the RIS-CQ, called Dual-Modality Graph Alignment ( DuMoGa ) model, which outperforms a series of RIS methods. To provide a valuable foundation for future advancements in the field of RIS with complex queries, we release the datasets, pre-processing and synthetic scripts, and the algorithm implementations at https://github.com/lili0415/DuMoGa .

In image question answering, due to the abundant and sometimes redundant information, precisely matching and integrating the information from both text and images is a challenge. In this paper, we propose the Decomposition-Integration Enhancing Multimodal Insight (DIEM) which initially decomposes the given question and image into multiple subquestions and several sub-images aiming to isolate specific elements for more focused analysis. We then in-tegrate these sub-elements by matching each subquestion with its relevant sub-images, while also retaining the original image, to construct a comprehensive answer to the original question without losing sight of the overall context. This strategy mirrors the human cognitive process of simplifying complex problems into smaller components for individual analysis, followed by an integration of these insights. We implement DIEM on the LLaVA-v1.5 model, and evaluate its performance on ScienceQA and MM-Vet. Ex-perimental results indicate that our method boosts accu-racy in most question classes of the ScienceQA (+2.03% in average), especially in the image modality (+3.40%). On MM-Vet, our method achieves an improvement in MM-Vet scores, increasing from 31.1 to 32.4. These findings high-light DIEM's effectiveness in harmonizing the complexities of multimodal data, demonstrating its ability to enhance accuracy and depth in image question answering through its decomposition-integration process.

Multi-modal Large Language Models (MLLMs) tuned on machine-generated instruction-following data have demonstrated remarkable performance in various multi-modal understanding and generation tasks. However, the hallucinations inherent in machine-generated data, which could lead to hallucinatory outputs in MLLMs, remain under-explored. This work aims to investigate various hallucinations (i.e., object, relation, attribute hallucinations) and mitigate those hallucinatory toxicities in large-scale machine-generated visual instruction datasets. Drawing on the human ability to identify factual errors, we present a novel hallucination detection and elimination framework, HalluciDoctor, based on the cross-checking paradigm. We use our framework to identify and eliminate hallucinations in the training data automatically. Interestingly, HalluciDoctor also indicates that spurious correlations arising from long-tail object cooccurrences contribute to hallucinations. Based on that, we execute counterfactual visual instruction expansion to balance data distribution, thereby enhancing MLLMs' resistance to hallucinations. Comprehensive experiments on hallucination evaluation benchmarks show that our method successfully mitigates 44.6% hallucinations relatively and maintains competitive performance compared to LLaVA. The data and code for this paper are publicly available.11https://github.com/Yuqifan1117/HalluciDoctor

Recent advancements in Multimodal Large Language Models (MLLMs) have been utilizing Visual Prompt Generators (VPGs) to convert visual features into tokens that LLMs can recognize. This is achieved by training the VPGs on millions of image-caption pairs, where the VPG-generated tokens of images are fed into a frozen LLM to generate the corresponding captions. However, this image-captioning based training objective inherently biases the VPG to concentrate solely on the primary visual contents sufficient for caption generation, often neglecting other visual details. This shortcoming results in MLLMs’ underperformance in comprehending demonstrative instructions consisting of multiple, interleaved, and multimodal instructions that demonstrate the required context to complete a task. To address this issue, we introduce a generic and lightweight Visual Prompt Generator Complete module (VPG-C), which can infer and complete the missing details essential for comprehending demonstrative instructions. Further, we propose a synthetic discriminative training strategy to fine-tune VPG-C, eliminating the need for supervised demonstrative instructions. As for evaluation, we build DEMON, a comprehensive benchmark for demonstrative instruction understanding. Synthetically trained with the proposed strategy, VPG-C achieves significantly stronger zero-shot performance across all tasks of DEMON. Further evaluation on the MME and OwlEval benchmarks also demonstrate the superiority of VPG-C. The code and models are available at https://github.com/DCDmllm/Cheetah.

We introduce InstructVid2Vid, an end-to-end diffusion-based methodology for video editing guided by human language instructions. Our approach empowers video manipulation guided by natural language directives, eliminating the need for per-example fine-tuning or inversion. The proposed InstructVid2Vid model modifies a pretrained image generation model, Stable Diffusion, to generate a time-dependent sequence of video frames. By harnessing the collective intelligence of disparate models, we engineer a training dataset rich in video-instruction triplets, which is a more cost-efficient alternative to collecting data in real-world scenarios. To enhance the coherence between successive frames within the generated videos, we propose the Inter-Frames Consistency Loss and incorporate it during the training process. With multimodal classifier-free guidance during the inference stage, the generated videos is able to resonate with both the input video and the accompanying instructions. Experimental results demonstrate that InstructVid2Vid is capable of generating high-quality, temporally coherent videos and performing diverse edits, including attribute editing, background changes, and style transfer. These results underscore the versatility and effectiveness of our proposed method.

For multimodal LLMs, the synergy of visual comprehension (textual output) and generation (visual output) presents an ongoing challenge. This is due to a conflicting objective: for comprehension, an MLLM needs to abstract the visuals; for generation, it needs to preserve the visuals as much as possible. Thus, the objective is a dilemma for visual-tokens. To resolve the conflict, we propose encoding images into morph-tokens to serve a dual purpose: for comprehension, they act as visual prompts instructing MLLM to generate texts; for generation, they take on a different, non-conflicting role as complete visual-tokens for image reconstruction, where the missing visual cues are recovered by the MLLM. Extensive experiments show that morph-tokens can achieve a new SOTA for multimodal comprehension and generation simultaneously. Our project is available at https://github.com/DCDmllm/MorphTokens.

Large Language Models (LLMs) demonstrate remarkable proficiency in comprehending and handling text-based tasks. Many efforts are being made to transfer these attributes to video modality, which are termed Video-LLMs. However, existing Video-LLMs can only capture the coarse-grained semantics and are unable to effectively handle tasks related to comprehension or localization of specific video segments. In light of these challenges, we propose Momentor, a Video-LLM capable of accomplishing fine-grained temporal understanding tasks. To support the training of Momentor, we design an automatic data generation engine to construct Moment-10M, a large-scale video instruction dataset with segment-level instruction data. We train Momentor on Moment-10M, enabling it to perform segment-level reasoning and localization. Zero-shot evaluations on several tasks demonstrate that Momentor excels in fine-grained temporally grounded comprehension and localization.

Visual programming, a modular paradigm, integrates different modules and Python operators to solve various vision-language tasks. Unlike end-to-end models that need task-specific data, it performs visual processing and inference in an unsupervised manner. Current visual programming methods generate programs in a single pass where the ability to evaluate and optimize based on feedback, unfortunately, is lacking, which consequentially limits their effectiveness for complex, multi-step problems. Inspired by benders decomposition, we introduce De-fine, a training-free framework that automatically decomposes complex tasks into simpler subtasks and refines programs through auto-feedback. This model-agnostic approach can improve logical reasoning performance by integrating the strengths of multiple models. Our experiments across various visual tasks show that De-fine creates more robust programs. Moreover, viewing each feedback module as an independent agent will yield fresh prospects for the field of agent research.

The remarkable performance of Multimodal Large Language Models (MLLMs) has demonstrated their proficient understanding capabilities in handling various visual tasks. Nevertheless, the opaque nature of black-box reasoning processes persists as an enigma, rendering them uninterpretable and struggling with hallucination. Their ability to execute intricate reasoning tasks is also constrained, culminating in stagnation of progression. In this work, we introduce Fact, a novel paradigm designed to generate multimodal rationales that are faithful, concise, and transferable for teaching MLLMs. This paradigm utilizes verifiable visual programming to generate executable code guaranteeing faithfulness. Through a series of operations including pruning, merging, and bridging, the rationale enhances its conciseness. Furthermore, we filter rationales that can be transferred to end-to-end paradigms from programming paradigms to guarantee transferability. Empirical evidence from experiments demonstrates the superiority of Fact across models of varying parameter sizes, significantly enhancing their compositional reasoning and generalization ability and reducing hallucinations owing to its high correlation between images and text.

We introduce the DEMON Challenge, defined as a benchmark for demonstrative instruction following, to ACM Multimedia 2024. The DEMON Challenge aims to assess the ability of models and systems to comprehend demonstrative instructions consisting of multiple, interleaved, and multimodal context that demonstrate the required information to complete a task. These instructions are curated from a diverse range of multi-modal datasets, spanning various fields and scenarios, to ensure comprehensive coverage and challenge diversity. The challenge details and participation information are available on the https://dcdmllm.github.io/DEMON-challenge/.

World models are progressively being employed across diverse fields, extending from basic environment simulation to complex scenario construction. However, existing models are mainly trained on domain-specific states and actions, and confined to single-modality state representations. In this paper, We introduce WorldGPT, a generalist world model built upon Multimodal Large Language Model (MLLM). WorldGPT acquires an understanding of world dynamics through analyzing millions of videos across various domains. To further enhance WorldGPT's capability in specialized scenarios and long-term tasks, we have integrated it with a novel cognitive architecture that combines memory offloading, knowledge retrieval, and context reflection. As for evaluation, we build WorldNet, a multimodal state transition prediction benchmark encompassing varied real-life scenarios. Conducting evaluations on WorldNet directly demonstrates WorldGPT's capability to accurately model state transition patterns, affirming its effectiveness in understanding and predicting the dynamics of complex scenarios. We further explore WorldGPT's emerging potential in serving as a world simulator, helping multimodal agents generalize to unfamiliar domains through efficiently synthesising multimodal instruction instances which are proved to be as reliable as authentic data for fine-tuning purposes. The code and dataset are available on the https://github.com/DCDmllm/WorldGPT

In recent times, Vision-Language Models (VLMs) have been trained under two predominant paradigms. Generative training has enabled Multimodal Large Language Models (MLLMs) to tackle various complex tasks, yet issues such as hallucinations and weak object discrimination persist. Discriminative training, exemplified by models like CLIP, excels in zero-shot image-text classification and retrieval, yet struggles with complex scenarios requiring fine-grained semantic differentiation. This paper addresses these challenges by proposing a unified approach that integrates the strengths of both paradigms. Considering interleaved image-text sequences as the general format of input samples, we introduce a structure-induced training strategy that imposes semantic relationships between input samples and the MLLM’s hidden state. This approach enhances the MLLM’s ability to capture global semantics and distinguish fine-grained semantics. By leveraging dynamic sequence alignment within the Dynamic Time Warping framework and integrating a novel kernel for fine-grained semantic differentiation, our method effectively balances generative and discriminative tasks. Extensive experiments demonstrate the effectiveness of our approach, achieving state-of-the-art results in multiple generative tasks, especially those requiring cognitive and discrimination abilities. Additionally, our method surpasses discriminative benchmarks in interleaved and fine-grained retrieval tasks. By employing a retrieval-augmented generation strategy, our approach further enhances performance in some generative tasks within one model, offering a promising direction for future research in vision-language modeling.

The swift advancement in Multimodal LLMs (MLLMs) also presents significant challenges for effective knowledge editing. Current methods, including intrinsic knowledge editing and external knowledge resorting, each possess strengths and weaknesses, struggling to balance the desired properties of reliability, generality, and locality when applied to MLLMs. In this paper, we propose \textbf{UniKE}, a novel multimodal editing method that establishes a unified perspective and paradigm for intrinsic knowledge editing and external knowledge resorting. Both types of knowledge are conceptualized as vectorized key-value memories, with the corresponding editing processes resembling the assimilation and accommodation phases of human cognition, conducted at the same semantic levels. Within such a unified framework, we further promote knowledge collaboration by disentangling the knowledge representations into the semantic and truthfulness spaces. Extensive experiments validate the effectiveness of our method, which ensures that the post-edit MLLM simultaneously maintains excellent reliability, generality, and locality. The code for UniKE is available at https://github.com/beepkh/UniKE.

Recent studies indicate that dense retrieval models struggle to perform well on a wide variety of retrieval tasks that lack dedicated training data, as different retrieval tasks often entail distinct search intents. To address this challenge, in this work we leverage instructions to flexibly describe retrieval intents and introduce I3, a unified retrieval system that performs Intent-Introspective retrieval across various tasks, conditioned on Instructions without any task-specific training. I3 innovatively incorporates a pluggable introspector in a parameter-isolated manner to comprehend specific retrieval intents by jointly reasoning over the input query and instruction, and seamlessly integrates the introspected intent into the original retrieval model for intent-aware retrieval. Furthermore, we propose progressively-pruned intent learning. It utilizes extensive LLM-generated data to train I3 phase-by-phase, embodying two key designs: progressive structure pruning and drawback extrapolation-based data refinement. Extensive experiments show that in the BEIR benchmark, I3 significantly outperforms baseline methods designed with task-specific retrievers, achieving state-of-the-art zero-shot performance without any task-specific tuning.

Dynamic graph-based data are ubiquitous in the real world, such as social networks, finance systems, and traffic flow. Fast and accurately detecting anomalies in these dynamic graphs is of vital importance. However, despite promising results the current anomaly detection methods have achieved, there are two major limitations when coping with dynamic graphs. The first limitation is that the topological structures and the temporal dynamics have been modeled separately, resulting in less expressive features for detection. The second limitation is that the models have been trained by unreliable noisy labels generated by random negative sampling, rendering it severely vulnerable to subtle perturbations. To overcome the above limitations, we propose RustGraph, a robust anomaly detection framework by jointly learning structural-temporal dependency in dynamic graphs. To this end, we design a variational graph auto-encoder with informative prior that simultaneously encodes both graph structural and temporal information. Then we introduce a fine-grained contrastive learning method to learn better node representations by utilizing the temporal consistency between two snapshots. Furthermore, we formulate the noisy label learning problem for anomaly detection in dynamic graph, and then propose a robust anomaly detector to improve the model performance by leveraging learned graph structure signal. Our extensive experiments on six real-world datasets demonstrate the proposed RustGraph method achieves state-of-the-art performance with an average of 3.64% improvement on AUC-ROC metric compared with all baselines. The codes are available athttps://github.com/aubreygjh/RustGraph.

Recent research on video moment retrieval has mostly focused on enhancing the performance of accuracy, efficiency, and robustness, all of which largely rely on the abundance of high-quality annotations. While the precise frame-level annotations are time-consuming and cost-expensive, few attentions have been paid to the labeling process. In this work, we explore a new interactive manner to stimulate the process of human-in-the-loop annotation in video moment retrieval task. The key challenge is to select “ambiguous” frames and videos for binary annotations to facilitate the network training. To be specific, we propose a new hierarchical uncertainty-based modeling that explicitly considers modeling the uncertainty of each frame within the entire video sequence corresponding to the query description, and selecting the frame with the highest uncertainty. Only selected frame will be annotated by the human experts, which can largely reduce the workload. After obtaining a small number of labels provided by the expert, we show that it is sufficient to learn a competitive video moment retrieval model in such a harsh environment. Moreover, we treat the uncertainty score of frames in a video as a whole, and estimate the difficulty of each video, which can further relieve the burden of video selection. In general, our active learning strategy for video moment retrieval works not only at the frame level but also at the sequence level. Experiments on two public datasets validate the effectiveness of our proposed method. Our code is released at https://github.com/renjie-liang/HUAL.

Prompt tuning, a recently emerging paradigm, enables the powerful vision-language pre-training models to adapt to downstream tasks in a parameter- and data- efficient way, by learning the "soft prompts" to condition frozen pretraining models. Though effective, it is particularly problematic in the few-shot scenario, where prompt tuning performance is sensitive to the initialization and requires a time-consuming process to find a good initialization, thus restricting the fast adaptation ability of the pre-training models. In addition, prompt tuning could undermine the generalizability of the pre-training models, because the learnable prompt tokens are easy to overfit to the limited training samples. To address these issues, we introduce a novel Gradient-RegulAted Meta-prompt learning (GRAM) framework that jointly meta-learns an efficient soft prompt initialization for better adaptation and a lightweight gradient regulating function for strong cross-domain generalizability in a meta-learning paradigm using only the unlabeled image-text pre-training data. Rather than designing a specific prompt tuning method, our GRAM can be easily incorporated into various prompt tuning methods in a model-agnostic way, and comprehensive experiments show that GRAM brings about consistent improvement for them in several settings (i.e., few-shot learning, cross-domain generalization, cross-dataset generalization, etc.) over 11 datasets. Further, experiments show that GRAM enables the orthogonal methods of textual and visual prompt tuning to work in a mutually-enhanced way, offering better generalizability beyond the uni-modal prompt tuning methods.

Scene Graph Generation (SGG) aims to extractrelationships in images for vision understanding. Although recent works have made steady progress on SGG, they still suffer long-tail distribution issues that tail-predicates are more costly to train and hard to distinguish due to a small amount of annotated data compared to frequent predicates. Existing re-balancing strategies try to handle it via prior rules but are still confined to pre-defined conditions, which are not scalable for various models and datasets. In this paper, we propose a Cross-modal prediCate boosting (CaCao) framework, where a visually-prompted language model is learned to generate diverse fine-grained predicates in a low-resource way. The proposed CaCao can be applied in a plug-and-play fashion and automatically strengthen existing SGG to tackle the long-tailed problem. Based on that, we further introduce a novel Entangled cross-modal prompt approach for open-world predicate scene graph generation (Epic), where models can generalize to unseen predicates in a zero-shot manner. Comprehensive experiments on three benchmark datasets show that CaCao consistently boosts the performance of multiple scene graph generation models in a model-agnostic way. Moreover, our Epic achieves competitive performance on open-world predicate prediction. The data and code for this paper are publicly available.1

This paper addresses the Unsupervised Domain Adaptation (UDA) for the dense frame prediction task - Video Object Grounding (VOG). This investigation springs from the recognition of the limited generalization capabilities of data-driven approaches when confronted with unseen test scenarios. We set the goal of enhancing the adaptability of the source-dominated model from a labeled domain to the unlabeled target domain through re-training on pseudo-labels (i.e., predicted boxes of language-described objects). Given the potential for source-domain biases in the pseudo-label generation, we decompose the labeling refinement as two cascaded debiasing subroutines: (1) we develop a discarded training strategy to correct the Biased Proposal Selection by filtering out the examples with uncertain proposals selected from the proposal (candidate box) set. The identifier of these uncertain examples is the discordance between the predictions of the source-dominated model and those of a target-domain clustered classifier, which remains free from the source-domain bias. (2) With the refined proposals as a foundation, we measure Grounding Coordinate Offset based on the semantic distance of the model's prediction across domains, based on which we alleviate source-domain bias in the target model through adversarial learning. To verify the superiority of the proposed method, we collected two UDA-VOG datasets called I2O-VOG and R2M-VOG by manually dividing and combining the well-known VOG datasets. The extensive experiments on them show our model significantly outperforms SOTA methods by a large margin.

Temporal grounding is the task of locating a specific segment from an untrimmed video according to a query sentence. This task has achieved significant momentum in the computer vision community as it enables activity grounding beyond pre-defined activity classes by utilizing the semantic diversity of natural language descriptions. The semantic diversity is rooted in the principle of compositionality in linguistics, where novel semantics can be systematically described by combining known words in novel ways (compositional generalization). However, existing temporal grounding datasets are not carefully designed to evaluate the compositional generalizability. To systematically benchmark the compositional generalizability of temporal grounding models, we introduce a new Compositional Temporal Grounding task and construct two new dataset splits, i.e., Charades-CG and ActivityNet-CG. We empirically find that they fail to generalize to queries with novel combinations of seen words. We argue that the inherent compositional structure (i.e., composition constituents and their relationships) inside the videos and language is the crucial factor to achieve compositional generalization. Based on this insight, we propose a variational cross-graph reasoning framework that explicitly decomposes video and language into hierarchical semantic graphs, respectively, and learns fine-grained semantic correspondence between the two graphs. Meanwhile, we introduce a novel adaptive structured semantics learning approach to derive the structure-informed and domain-generalizable graph representations, which facilitate the fine-grained semantic correspondence reasoning between the two graphs. To further evaluate the understanding of the compositional structure, we also introduce a more challenging setting, where one of the components in the novel composition is unseen. This requires more sophisticated understanding of the compositional structure to infer the potential semantics of the unseen word based on the other learned composition constituents appearing in both the video and language context, and their relationships. Extensive experiments validate the superior compositional generalizability of our approach, demonstrating its ability to handle queries with novel combinations of seen words as well as novel words in the testing composition.

Text-based image captioning (TextCap) requires simultaneous comprehension of visual content and reading the text of images to generate a natural language description. Although a task can teach machines to understand the complex human environment further given that text is omnipresent in our daily surroundings, it poses additional challenges in normal captioning. A text-based image intuitively contains abundant and complex multimodal relational content, that is, image details can be described diversely from multiview rather than a single caption. Certainly, we can introduce additional paired training data to show the diversity of images' descriptions, this process is labor-intensive and time-consuming for TextCap pair annotations with extra texts. Based on the insight mentioned above, we investigate how to generate diverse captions that focus on different image parts using an unpaired training paradigm. We propose the Multimodal relAtional Graph adversarIal InferenCe (MAGIC) framework for diverse and unpaired TextCap. This framework can adaptively construct multiple multimodal relational graphs of images and model complex relationships among graphs to represent descriptive diversity. Moreover, a cascaded generative adversarial network is developed from modeled graphs to infer the unpaired caption generation in image–sentence feature alignment and linguistic coherence levels. We validate the effectiveness of MAGIC in generating diverse captions from different relational information items of an image. Experimental results show that MAGIC can generate very promising outcomes without using any image–caption training pairs.

Temporal grounding in videos aims to localize one target video segment that semantically corresponds to a given query sentence. Thanks to the semantic diversity of natural language descriptions, temporal grounding allows activity grounding beyond pre-defined classes and has received increasing attention in recent years. The semantic diversity is rooted in the principle of compositionality in linguistics, where novel semantics can be systematically described by combining known words in novel ways (compositional generalization). However, current temporal grounding datasets do not specifically test for the compositional generalizability. To systematically measure the compositional generalizability of temporal grounding models, we introduce a new Compositional Temporal Grounding task and construct two new dataset splits, i.e., Charades-CG and ActivityNet-CG. Evaluating the state-of-the-art methods on our new dataset splits, we empirically find that they fail to generalize to queries with novel combinations of seen words. To tackle this challenge, we propose a variational cross-graph reasoning framework that explicitly decomposes video and language into multiple structured hierarchies and learns fine-grained semantic correspondence among them. Experiments illustrate the superior compositional generalizability of our approach. The repository of this work is at ht tps: / / gi thub. com/YYJMJC/ Composi tional- Temporal-Grounding.

Understanding human emotions is a crucial ability for intelligent robots to provide better human-robot interactions. The existing works are limited to trimmed video-level emotion classification, failing to locate the temporal window corresponding to the emotion. In this paper, we introduce a new task, named Temporal Emotion Localization in videos (TEL), which aims to detect human emotions and localize their corresponding temporal boundaries in untrimmed videos with aligned subtitles. TEL presents three unique challenges compared to temporal action localization: 1) The emotions have extremely varied temporal dynamics; 2) The emotion cues are embedded in both appearances and complex plots; 3) The fine-grained temporal annotations are complicated and labor-intensive. To address the first two challenges, we propose a novel dilated context integrated network with a coarse-fine two-stream architecture. The coarse stream captures varied temporal dynamics by modeling multi-granularity temporal contexts. The fine stream achieves complex plots understanding by reasoning the dependency between the multi-granularity temporal contexts from the coarse stream and adaptively integrates them into fine-grained video segment features. To address the third challenge, we introduce a cross-modal consensus learning paradigm, which leverages the inherent semantic consensus between the aligned video and subtitle to achieve weakly-supervised learning. We contribute a new testing set with 3,000 manually-annotated temporal boundaries so that future research on the TEL problem can be quantitatively evaluated. Extensive experiments show the effectiveness of our approach on temporal emotion localization. The repository of this work is at https://github.com/YYJMJC/TemporalEmotion-Localization-in-Videos.

Fingerspelling detection, which aims to localize and recognize fingerspelling gestures in raw, untrimmed videos, is a nascent but important research area that could help bridge the communication gap between deaf people and others. Many existing works tend to exploit additional knowledge, such as pose annotations, and newly datasets for performance improvement. However, in real-world applications, additional data collection and annotation require tremendous human efforts that are not always affordable. In this paper, we propose the Weakly-supervised Disentanglement Network, namely WED, that requires no additional knowledge, and better exploits the video-sentence weak supervisions. Specifically, WED incorporates two critical components: 1) Masked Disentanglement Module, which employs a Variational Autoencoder for signed letters disentanglement. Each latent factor in the VAE corresponds to a particular signed letter, and we mask latent factors corresponding to letters that do not appear in the video during decoding. Compared to the vanilla VAE, the masked reconstruction leverages the video-sentence weak supervision, leading to a better sign language oriented disentanglement; and 2) the Dynamic Memory Network module, which leverages the disentangled sign knowledge as prior knowledge and reference for sign-related frame identification and gesture recognition through a carefully designed memory reading component. We conduct extensive experiments on the benchmark ChicagoFSWild and ChicagoFSWild+ datasets. Empirical studies validate that the WED network achieves effective sign gesture disentanglement, contributing to the state-of-the-art performance for fingerspelling detection and recognition.

Large-scale vision-language pre-training has shown impressive advances in a wide range of downstream tasks. Existing methods mainly model the cross-modal alignment by the similarity of the global representations of images and text, or advanced cross-modal attention upon image and text features. However, they fail to explicitly learn the fine-grained semantic alignment between visual regions and textual phrases, as only global image-text alignment information is available. In this paper, we introduce LOUPE, a fine-grained semantically aLigned visiOn-langUage PrE-training framework, which learns fine-grained semantic alignment from the novel perspective of game-theoretic interactions. To efficiently estimate the game-theoretic interactions, we further propose an uncertainty-aware neural Shapley interaction learning module. Experiments show that LOUPE achieves state-of-the-art performance on a variety of vision-language tasks. Without any object-level human annotations and fine-tuning, LOUPE achieves competitive performance on object detection and visual grounding. More importantly, LOUPE opens a new promising direction of learning fine-grained semantics from large-scale raw image-text pairs.

Video-and-Language Inference is a recently proposed task for joint video-and-language understanding. This new task requires a model to draw inference on whether a natural language statement entails or contradicts a given video clip. In this paper, we study how to address three critical challenges for this task: judging the global correctness of the statement involved multiple semantic meanings, joint reasoning over video and subtitles, and modeling long-range relationships and complex social interactions. First, we propose an adaptive hierarchical graph network that achieves in-depth understanding of the video over complex interactions. Specifically, it performs joint reasoning over video and subtitles in three hierarchies, where the graph structure is adaptively adjusted according to the semantic structures of the statement. Secondly, we introduce semantic coherence learning to explicitly encourage the semantic coherence of the adaptive hierarchical graph network from three hierarchies. The semantic coherence learning can further improve the alignment between vision and linguistics, and the coherence across a sequence of video segments. Experimental results show that our method significantly outperforms the baseline by a large margin.

Visual navigation is a task of training an embodied agent by intelligently navigating to a target object (e.g., television) using only visual observations. A key challenge for current deep reinforcement learning models lies in the requirements for a large amount of training data. It is exceedingly expensive to construct sufficient 3D synthetic environments annotated with the target object information. In this paper, we focus on visual navigation in the low-resource setting, where we have only a few training environments annotated with object information. We propose a novel unsupervised reinforcement learning approach to learn transferable meta-skills (e.g., bypass obstacles, go straight) from unannotated environments without any supervisory signals. The agent can then fast adapt to visual navigation through learning a high-level master policy to combine these meta-skills, when the visual-navigation-specified reward is provided. Experimental results show that our method significantly outperforms the baseline by 53.34% relatively on SPL, and further qualitative analysis demonstrates that our method learns transferable motor primitives for visual navigation.

Visual Storytelling~(VIST) is a task to tell a narrative story about a certain topic according to the given photo stream. The existing studies focus on designing complex models, which rely on a huge amount of human-annotated data. However, the annotation of VIST is extremely costly and many topics cannot be covered in the training dataset due to the long-tail topic distribution. In this paper, we focus on enhancing the generalization ability of the VIST model by considering the few-shot setting. Inspired by the way humans tell a story, we propose a topic adaptive storyteller to model the ability of inter-topic generalization. In practice, we apply the gradient-based meta-learning algorithm on multi-modal seq2seq models to endow the model the ability to adapt quickly from topic to topic. Besides, We further propose a prototype encoding structure to model the ability of intra-topic derivation. Specifically, we encode and restore the few training story text to serve as a reference to guide the generation at inference time. Experimental results show that topic adaptation and prototype encoding structure mutually bring benefit to the few-shot model on BLEU and METEOR metric. The further case study shows that the stories generated after few-shot adaptation are more relative and expressive.

The EmbodiedQA is a task of training an embodied agent by intelligently navigating in a simulated environment and gathering visual information to answer questions. Existing approaches fail to explicitly model the mental imagery function of the agent, while the mental imagery is crucial to embodied cognition, and has a close relation to many high-level meta-skills such as generalization and interpretation. In this paper, we propose a novel Mental Imagery eNhanceD (MIND) module for the embodied agent, as well as a relevant deep reinforcement framework for training. The MIND module can not only model the dynamics of the environment (e.g. 'what might happen if the agent passes through a door') but also help the agent to create a better understanding of the environment (e.g. 'The refrigerator is usually in the kitchen'). Such knowledge makes the agent a faster and better learner in locating a feasible policy with only a few trails. Furthermore, the MIND module can generate mental images that are treated as short-term subgoals by our proposed deep reinforcement framework. These mental images facilitate policy learning since short-term subgoals are easy to achieve and reusable. This yields better planning efficiency than other algorithms that learn a policy directly from primitive actions. Finally, the mental images visualize the agent's intentions in a way that human can understand, and this endows our agent's actions with more interpretability. The experimental results and further analysis prove that the agent with the MIND module is superior to its counterparts not only in EQA performance but in many other aspects such as route planning, behavioral interpretation, and the ability to generalize from a few examples.