Large language model agents increasingly act through terminals, yet existing surveys disperse terminal-mediated behavior across software engineering, tool use, and computer-use research. We regard terminal agents as systems whose dominant progress-bearing action–observation loop is mediated by terminal command execution, textual feedback, and stateful environment interaction. Using terminal-mediated execution as an organizing lens, this survey establishes workload-level boundaries and connects system architecture, competence acquisition, and evaluation through a seven-dimensional terminal competence profile. Our synthesis shows that realized behavior is jointly shaped by the model, interface, harness, runtime, and environment. Executable trajectories ground learning in action consequences, verification, and recovery, whereas prevailing evaluations emphasize final outcomes and expose process quality, recovery, and governance unevenly. Bounded fixed-condition diagnostics illustrate two implications: benchmark families expose different process signals, and matched system comparisons reveal benchmark-dependent performance and limits of component attribution. These findings motivate explicit reporting of system and runtime conditions, supported by replayable traces and process-level evidence. The framework provides a unified basis for studying terminal-mediated agency across software engineering and emerging application domains.

Personalized outfit recommendation poses a significant challenge in e-commerce and social media platforms, requiring systems that balance user preferences with aesthetic compatibility. Collaborative filtering (CF) provides a traditional solution for this, but it struggles with data-sparse scenarios and complex user-item-outfit relationships. Meanwhile, existing template-based approaches are constrained by rigid pre-designed structures. To bridge these research gaps, we introduce CFALR (Collaborative Filtering-Augmented Large Language Model for Recommendation), a novel framework that synergizes collaborative filtering with large language models for personalized outfit recommendation. Specifically, CFALR describes user-outfit interactions in natural language and leverages LLMs to capture fashion semantics while employing CF-enhanced embeddings to bridge the semantic space and the collaborative interaction spaces. Our technical contributions include: (1) the first LLM-based architecture specifically designed for personalized outfit recommendation, (2) a CF-augmented generative mechanism that efficiently navigates the extensive combination space of outfit items, and (3) trainable projection layers that optimally integrate relational and content features. Experiments on Polyvore and IQON benchmarks demonstrate CFALR's superior performance over both traditional CF-based and LLM-based methods in personalized fill-in-the-blank and personalized outfit generation tasks.

Large vision-language models (LVLMs) have achieved significant progress in video understanding, yet understanding long videos remains challenging due to the large number of visual tokens and limited context windows. Visual sampling provides a practical solution by selecting an informative subset of frames. However, existing methods typically either rely on relevance-aware sampling, leading to redundant frame selection and insufficient temporal coverage, or adopt a fixed sampling strategy regardless of query type. In this paper, we propose VisualRouter, a training-free and plug-and-play framework for query-grounded visual sampling. VisualRouter first classifies each query as either global or local and then applies the corresponding sampling strategy. For global queries, it employs a relevance-coverage hybrid strategy that preserves temporal coverage while retaining query-relevant visual evidence. For local queries, it adopts an event-aware frame selection strategy that performs event partitioning, segment-level frame allocation, and intra-event frame selection, jointly balancing relevance, coverage, and diversity with a limited number of input frames. Experiments show that VisualRouter consistently improves multiple LVLMs over uniform sampling, achieving gains of 5.2

Fashion is a knowledge-intensive domain in which effective decision-making depends on integrating multiple types of knowledge. Although Large Language Models (LLMs) have transformed many areas, their application in fashion remains limited by hallucinations and weak domain specialization. Knowledge Graph (KG)-based Retrieval-Augmented Generation (RAG) offers a promising way to add structured knowledge to LLMs. However, existing fashion KGs are typically restricted to product-level attributes or item relations, and fail to capture the broader fashion ecosystem. To bridge these gaps, we propose FashionEcoKG, a comprehensive, domain-wide knowledge graph built with expert-level precision and professionalism. It is constructed through a three-stage agentic pipeline that extracts high-fidelity knowledge cores from authoritative textbooks and strengthens structural connectivity through cross-domain augmentation and generative expansion. To leverage this resource, we further develop PG-RAG (Pruning-Grounding RAG), a training-free framework designed to handle the conceptual density and linguistic noise of fashion queries. Specifically, we introduce a Dual-Granularity Path Re-Ranking (DGPR) module of two stages. The Pruning-based Semantic Ranking (PSR) module distills each query into a skeleton form to improve retrieval recall, while the Grounding-based Agentic Ranking (GAR) performs point-wise scrutiny of candidate paths against the original full query to ensure global relevance. Experiments on a curated fashion QA dataset show that PG-RAG effectively leverages FashionEcoKG to improve retrieval and answer accuracy, outperforming both non-RAG and existing KG-RAG baselines.

Open-Vocabulary Segmentation (OVS) aims to segment image regions beyond predefined category sets by leveraging semantic descriptions. While CLIP based approaches excel in semantic generalization, they frequently lack the fine-grained spatial awareness required for dense prediction. Recent efforts have incorporated Vision Foundation Models (VFMs) like DINO to alleviate these limitations. However, these methods still struggle with the precise edge perception necessary for high fidelity segmentation. In this paper, we analyze internal representations of DINO and discover that its inherent boundary awareness is not absent but rather undergoes progressive attenuation as features transition into deeper transformer blocks. To address this, we propose OVS-DINO, a novel framework that revitalizes latent edge-sensitivity of DINO through structural alignment with the Segment Anything Model (SAM). Specifically, we introduce a Structure-Aware Encoder (SAE) and a Structure-Modulated Decoder (SMD) to effectively activate boundary features of DINO using SAM's structural priors, complemented by a supervision strategy utilizing SAM generated pseudo-masks. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple weakly-supervised OVS benchmarks, improving the average score by 2.1

Developing intelligent agents capable of operating a wide range of Graphical User Interfaces (GUIs) with human-level proficiency is a key milestone on the path toward Artificial General Intelligence. While most existing datasets and benchmarks for training and evaluating GUI agents are static and idealized, failing to reflect the complexity and unpredictability of real-world environments, particularly the presence of anomalies. To bridge this research gap, we propose D-GARA, a dynamic benchmarking framework, to evaluate Android GUI agent robustness in real-world anomalies. D-GARA introduces a diverse set of real-world anomalies that GUI agents commonly face in practice, including interruptions such as permission dialogs, battery warnings, and update prompts. Based on D-GARA framework, we construct and annotate a benchmark featuring commonly used Android applications with embedded anomalies to support broader community research. Comprehensive experiments and results demonstrate substantial performance degradation in state-of-the-art GUI agents when exposed to anomaly-rich environments, highlighting the need for robustness-aware learning. D-GARA is modular and extensible, supporting the seamless integration of new tasks, anomaly types, and interaction scenarios to meet specific evaluation goals.

Recent work on domain-specific reasoning with large language models (LLMs) often relies on training-intensive approaches that require parameter updates. While activation steering has emerged as a parameter efficient alternative, existing methods apply static, manual interventions that fail to adapt to the dynamic nature of complex reasoning. To address this limitation, we propose RISER (Router-based Intervention for Steerable Enhancement of Reasoning), a plug-and-play intervention framework that adaptively steers LLM reasoning in activation space. RISER constructs a library of reusable reasoning vectors and employs a lightweight Router to dynamically compose them for each input. The Router is optimized via reinforcement learning under task-level rewards, activating latent cognitive primitives in an emergent and compositional manner. Across seven diverse benchmarks, RISER yields 3.4-6.5

Existing weight-activation quantization methods for Large Language Models (LLMs) primarily address channel-wise outliers but often neglect token-wise outliers, which limits the accuracy of quantized models. In this work, we propose Pre fixQuant, a novel quantization method that achieves state-of-the art performance across various precision levels (W4A4KV4 and W4A8KV4) and granularities (dynamic and static quantization) by effectively isolating token-wise outliers. First, PrefixQuant eliminates token-wise outliers by prefixing outlier tokens in the KVcache, a process that is training-free and highly efficient (e.g., 1 minutes for Llama-3-70B). Second, PrefixQuant introduces new trainable parameters for block-wise training to compensate for quantization error. Our experiments show that PrefixQuant significantly outperforms existing dynamic quantization methods, even under coarser static quantization settings. For instance, PrefixQuant achieves an average accuracy improvement of +3.08 and +2.85 points over SpinQuant (dynamic quantization) on five zero-shot reasoning tasks under dynamic and static quantization settings, respectively, on W4A4KV4 Llama-3-8B. Additionally, we demonstrate up to 2.74× prefilling speedup and 2.16× decoding speedup for LLMs using W4A4 PrefixQuant. Our code is available at https://github.com/ChenMnZ/PrefixQuant.

LLM-based agents increasingly interact with external environments through terminal command execution, yet existing surveys have rarely treated the terminal itself as a primary analytical object. This survey examines terminal agents, namely systems whose task progress depends on iterative command execution, textual feedback, and stateful terminal command interaction, and clarifies their boundaries with adjacent categories such as software-engineering agents, GUIor browser-based computer-use agents, and CLI-packaged assistants. Through a substratecentered lens, we systematize the literature around architectures and outer-loop design patterns, competence acquisition through executable environments, command-observation trajectories, and post-training, and evaluation protocols for terminal-mediated capabilities. Across systems, acquisition pipelines, and benchmarks, the synthesis shows that outer-loop design is not an implementation detail but a first-class variable that materially shapes measured performance. The evidence further indicates that terminal competence is multi-dimensional, spanning how agents formulate actions, interpret feedback, manage runtimes, track state and context, verify progress, recover from failures, and control side effects, rather than reducible to a single capability ranking. Current evidence remains concentrated in software engineering, while cross-domain transfer, model-versus-scaffold attribution, reliable recovery in mutable environments, and process-level evaluation remain underdeveloped. The survey provides an evidence-calibrated map of established findings, emerging practices, and unresolved challenges for terminal agents.

Vision-language models increasingly succeed on multimodal reasoning benchmarks, yet their visual evidence often becomes unstable once it enters the language stack, weakening evidence-grounded reasoning. To understand this fragility, we examine the internal dynamics of VLMs through a mechanistic lens and uncover a stable three-stage redistribution of multimodal attention focus across depth: an early question-conditioned organization, a critical middle visual-dominant relay, and a late return to answer formation. We operationalize the middle phase as the Visual Relay Window (VRW), and show that its geometry varies with task demand, is causally tied to grounded generation, and distinguishes unsupported answers from stronger reasoning trajectories. Guided by this internal rhythm, we propose TRACE, a task-adaptive inference-time control framework with lightweight trained modules. It reshapes relay allocation during prefill and preserves assembled visual support after handoff during decoding. Across four open-weight VLM backbones and seven benchmarks, TRACE delivers large gains on grounding-sensitive settings, improving them by 4.33 points on average and by up to 6.6 points, while also improving reasoning-heavy tasks. These results show that explicitly controlling multimodal focus across depth offers a unified and effective mechanism for strengthening evidence-grounded multimodal reasoning.

The technical evolution of Multimodal Large Language Models (MLLMs) has profoundly influenced the AI community, reshaping how intelligent systems understand, reason, and interact in both the digital and physical worlds. Through large-scale multimodal pretraining, MLLMs lay the foundation for general perception and alignment, while it remains unclear how such models can be transformed into human-grounded behaviors. Multimodal Post-Training, which further refines pretrained MLLMs to align their behaviors with human intent and real-world task demands, has become an important paradigm for advancing multimodal intelligence. Accordingly, both academia and industry have devoted increasing efforts to this direction, pushing the boundaries of multimodal general intelligence at an unprecedented pace. In this survey, we systematically review existing MMPoT research from a behavior-shaping perspective. Guided by this view, we introduce a unified framework that organizes existing MLLMs post-training methods into five major families: instruction following, preference calibration, reasoning enhancement, domain adaptation, and scalable training. Moreover, we revisit benchmarks and evaluation protocols to critically assess the capabilities and limitations of post-trained MLLMs. To conclude, we provide a synthesized outlook on broader open questions, and highlight future research directions toward more general and reliable multimodal capabilities. An updated paper list is available at: https://github.com/zchoi/Awesome-post-training-for-MLLMs.

Recent attempts to transfer features from 2D Vision–Language Models (VLMs) to 3D semantic segmentation expose a persistent trade-off. Directly projecting 2D features into 3D yields noisy and fragmented predictions, whereas enforcing geometric coherence necessitates costly training pipelines and large-scale, annotated 3D data. We argue that this limitation stems from the dominant \textit{segmentation-and-matching} paradigm, which fails to reconcile 2D semantics with 3D geometric structure. The geometric cues are not eliminated during the 2D-to-3D transfer but remain latent within the noisy and view-aggregated features. To exploit this property, we propose \textbf{GeoPurify} that applies a small Student Affinity Network to purify 2D VLM-generated 3D point features using geometric priors distilled from a 3D self-supervised teacher model. During inference, we devise a Geometry-Guided Pooling module to further denoise the point cloud and ensure the semantic and structural consistency. Benefiting from latent geometric information and the learned affinity network, GeoPurify effectively mitigates the trade-off and achieves superior data efficiency. Extensive experiments on major 3D benchmarks demonstrate that GeoPurify achieves or surpasses state-of-the-art performance while utilizing only \textbf{$\sim$1.5\%} of the training data.

Event forecasting is inherently influenced by multifaceted considerations, including international relations, regional historical dynamics, and cultural contexts. However, existing LLM-based approaches employ single-model architectures that generate predictions along a singular explicit trajectory, constraining their ability to capture diverse geopolitical nuances across complex regional contexts. To address this limitation, we introduce ThinkTank-ME, a novel Think Tank framework for Middle East event forecasting that emulates collaborative expert analysis in real-world strategic decision-making. To facilitate expert specialization and rigorous evaluation, we construct POLECAT-FOR-ME, a Middle East-focused event forecasting benchmark. Experimental results demonstrate the superiority of multi-expert collaboration in handling complex temporal geopolitical forecasting tasks. The code is available at https://github.com/LuminosityX/ThinkTank-ME.

Unsupervised medical anomaly detection learns normal anatomical patterns from healthy training images and identifies deviations at test time. Reconstruction-based and diffusion-based methods commonly use the difference between an input image and its reconstruction as anomaly evidence. However, this residual can be ambiguous. Expressive models may preserve pathological structures, while benign anatomical variation, imaging noise, and acquisition differences may also produce large reconstruction errors. We propose discriminative mask-guided diffusion (DMD), a medical anomaly detection framework that complements residual-based localization with reconstruction-shift discrimination. DMD first learns a compact quantized latent representation of normal images. Localized masks then perturb selected latent regions, and a latent diffusion model reconstructs the perturbed representations. The resulting reconstructions are paired with their original normal images to define a self-supervised classification task. At inference, the classifier provides a learned image-level anomaly score, while the residual between the input and its diffusion-based reconstruction yields a pixel-level anomaly map. Experiments on five datasets spanning brain MRI, breast ultrasound, and chest radiography show that DMD achieves the best overall performance among the state-of-the-art baseline methods.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across multimodal tasks such as visual perception and reasoning, leading to good performance on various multimodal evaluation benchmarks. However, these benchmarks keep a static nature and overlap with the pre-training data, resulting in fixed complexity constraints and data contamination issues. This raises the concern regarding the validity of the evaluation. To address these two challenges, we introduce a dynamic multimodal evaluation protocol called Vision-Language Bootstrapping (VLB). VLB provides a robust and comprehensive assessment for LVLMs with reduced data contamination and flexible complexity. To this end, VLB dynamically generates new visual question-answering samples through a multimodal bootstrapping module that modifies both images and language, while ensuring that newly generated samples remain consistent with the original ones by a judge module. By composing various bootstrapping strategies, VLB offers dynamic variants of existing benchmarks with diverse complexities, enabling the evaluation to co-evolve with the ever-evolving capabilities of LVLMs. Extensive experimental results across multiple benchmarks, including SEEDBench, MMBench, and MME, show that VLB significantly reduces data contamination and exposes performance limitations of LVLMs.

Recent breakthroughs in generative Artificial Intelligence (AI) have ignited a revolutionary wave across information retrieval and recommender systems. This workshop serves as a premier interdisciplinary platform to explore how generative models, particularly Large Language Models (LLMs) and Large Multimodal Models (LMMs), are transforming multimodal search and recommendation paradigms[3, 6, 9, 10, 12-14]. We aim to convene researchers and practitioners to discuss innovative architectures, methodologies, and evaluation strategies spanning generative document retrieval[5, 8], generative image retrieval[7, 16], grounded answer generation[17], generative recommendation[2, 4, 11], and related tasks involving multiple modalities[1, 15]. The workshop will facilitate discussions on improving algorithms, generating personalized content, evolving user-system interactions, enhancing trustworthiness, and refining evaluation methodologies for these cutting-edge systems. This timely workshop seeks to identify promising research directions, address key challenges, and foster collaborations towards the development of next-generation intelligent systems.

To equip artificial intelligence with a comprehensive understanding towards a temporal world, video and 4D panoptic scene graph generation abstracts visual data into nodes to represent entities and edges to capture temporal relations. Existing methods encode entity masks tracked across temporal dimensions (mask tubes), then predict their relations with temporal pooling operation, which does not fully utilize the motion indicative of the entities' relation. To overcome this limitation, we introduce a contrastive representation learning framework that focuses on motion pattern for temporal scene graph generation. Firstly, our framework encourages the model to learn close representations for mask tubes of similar subject-relation-object triplets. Secondly, we seek to push apart mask tubes from their temporally shuffled versions. Moreover, we also learn distant representations for mask tubes belonging to the same video but different triplets. Extensive experiments show that our motion-aware contrastive framework significantly improves state-of-the-art methods on both video and 4D datasets.

Cross-modal retrieval (CMR) is a fundamental task in multimedia research, focused on retrieving semantically relevant targets across different modalities. While traditional CMR methods match text and image via embedding-based similarity calculations, recent advancements in pre-trained generative models have established generative retrieval as a promising alternative. This paradigm assigns each target a unique identifier and leverages a generative model to directly predict identifiers corresponding to input queries without explicit indexing. Despite its great potential, current generative CMR approaches still face semantic information insufficiency in both identifier construction and generation processes. To address these limitations, we propose a novel unified Semantic-enhanced generative Cross-mOdal REtrieval framework (SemCORE), designed to unleash the semantic understanding capabilities in generative cross-modal retrieval task. Specifically, we first construct a Structured natural language IDentifier (SID) that effectively aligns target identifiers with generative models optimized for natural language comprehension and generation. Furthermore, we introduce a Generative Semantic Verification (GSV) strategy enabling fine-grained target discrimination. Additionally, to the best of our knowledge, SemCORE is the first framework to simultaneously consider both text-to-image and image-to-text retrieval tasks within generative cross-modal retrieval. Extensive experiments demonstrate that our framework outperforms state-of-the-art generative cross-modal retrieval methods. Notably, SemCORE achieves substantial improvements across benchmark datasets, with an average increase of 8.65 points in Recall@1 for text-to-image retrieval.

Talking face video generation with arbitrary speech audio is a significant challenge within the realm of digital human technology. The previous studies have emphasized the significance of audio-lip synchronization and visual quality. Currently, limited attention has been given to the learning of visual uncertainty, which creates several issues in existing systems, including inconsistent visual quality and unreliable performance across different input conditions. To address the problem, we propose a Joint Uncertainty Learning Network (JULNet) for high-quality talking face video generation, which incorporates a representation of uncertainty that is directly related to visual error. Specifically, we first design an uncertainty module to individually predict the error map and uncertainty map after obtaining the generated image. The error map represents the difference between the generated image and the ground truth image, while the uncertainty map is used to predict the probability of incorrect estimates. Furthermore, to match the uncertainty distribution with the error distribution through a KL divergence term, we introduce a histogram technique to approximate the distributions. By jointly optimizing error and uncertainty, the performance and robustness of our model can be enhanced. Extensive experiments demonstrate that our method achieves superior high-fidelity and audio-lip synchronization in talking face video generation compared to previous methods.

The accumulation of adversarial perturbations in the feature space makes it impossible for Deep Neural Networks (DNNs) to know what features are robust and reliable, and thus DNNs can be fooled by relying on a single contaminated feature. Numerous defense strategies attempt to improve their robustness by denoising, deactivating, or recalibrating non-robust features. Despite their effectiveness, we still argue that these methods are under-explored in terms of determining how trustworthy the features are. To address this issue, we propose a novel Evidence-based Multi-Feature Fusion (termed EMFF) for adversarial robustness. Specifically, our EMFF approach introduces evidential deep learning to help DNNs quantify the belief mass and uncertainty of the contaminated features. Subsequently, a novel multi-feature evidential fusion mechanism based on Dempster's rule is proposed to fuse the trusted features of multiple blocks within an architecture, which further helps DNNs avoid the induction of a single manipulated feature and thus improve their robustness. Comprehensive experiments confirm that compared with existing defense techniques, our novel EMFF method has obvious advantages and effectiveness in both scenarios of white-box and black-box attacks, and also prove that by integrating into several adversarial training strategies, we can improve the robustness of across distinct architectures, including traditional CNNs and recent vision Transformers with a few extra parameters and almost the same cost.