CVPR 2026 Papers

Post-training alignment of video generation models with human preferences is a critical goal. Developing effective Reward Models (RMs) for this process faces significant methodological hurdles. Current data collection paradigms, reliant on in-prompt pairwise annotations, suffer from labeling noise. Concurrently, the architectural design of VLM-based RMs, particularly their output mechanisms, remains underexplored. Furthermore, RM is susceptible to reward hacking in post-training. To mitigate these limitations, we propose SoliReward, a systematic framework for video RM training. Our framework first sources high-quality, cost-efficient data via single-item binary annotations, then constructs preference pairs using a cross-prompt pairing strategy. Architecturally, we employ a Hierarchical Progressive Query Attention mechanism to enhance feature aggregation. Finally, we introduce a modified BT loss that explicitly accommodates win-tie scenarios. This approach regularizes the RM's score distribution for positive samples, providing more nuanced preference signals to alleviate over-focus on a small number of top-scoring samples. Our approach is validated on benchmarks evaluating physical plausibility, subject deformity, and semantic alignment, demonstrating improvements in direct RM evaluation metrics and in the efficacy of post-training on video generation models. Code and benchmark are available at https://github.com/lian700/SoliReward.

Spatial understanding remains a weakness of Large Vision-Language Models (LVLMs). Existing supervised fine-tuning (SFT) and recent reinforcement learning with verifiable rewards (RLVR) pipelines depend on costly supervision, specialized tools, or constrained environments that limit scale. We introduce Spatial-SSRL, a self-supervised RL paradigm that derives verifiable signals directly from ordinary RGB or RGB-D images. Spatial-SSRL automatically formulates five pretext tasks that capture 2D and 3D spatial structure: shuffled patch reordering, flipped patch recognition, cropped patch inpainting, regional depth ordering, and relative 3D position prediction. These tasks provide ground-truth answers that are easy to verify and require no human or LVLM annotation. Training on our tasks substantially improves spatial reasoning while preserving general visual capabilities. On seven spatial understanding benchmarks in both image and video settings, Spatial-SSRL delivers average accuracy gains of 4.63% (3B) and 3.89% (7B) over the Qwen2.5-VL baselines. Our results show that simple, intrinsic supervision enables RLVR at scale and provides a practical route to stronger spatial intelligence in LVLMs.

3D modeling is shifting from static visual representations toward physical, articulated assets that can be directly used in simulation and interaction. However, most existing 3D generation methods overlook key physical and articulation properties, thereby limiting their utility in embodied AI. To bridge this gap, we introduce PhysX-Anything, the first simulation-ready physical 3D generative framework that, given a single in-the-wild image, produces high-quality sim-ready 3D assets with explicit geometry, articulation, and physical attributes. Specifically, we propose the first VLM-based physical 3D generative model, along with a new 3D representation that efficiently tokenizes geometry. It reduces the number of tokens by 193x, enabling explicit geometry learning within standard VLM token budgets without introducing any special tokens during fine-tuning and significantly improving generative quality. In addition, to overcome the limited diversity of existing physical 3D datasets, we construct a new dataset, PhysX-Mobility, which expands the object categories in prior physical 3D datasets by over 2xand includes more than 2K common real-world objects with rich physical annotations. Extensive experiments on PhysX-Mobility and in-the-wild images demonstrate that PhysX-Anything delivers strong generative performance and robust generalization. Furthermore, simulation-based experiments in a MuJoCo-style environment validate that our sim-ready assets can be directly used for contact-rich robotic policy learning. We believe PhysX-Anything can substantially empower a broad range of downstream applications, especially in embodied AI and physics-based simulation.

Foundational vision models have become the de facto standard for many vision tasks due to their strong performance. However, they are notoriously opaque and remain hard to interpret. We present ALOE (ALign Once to Explain), a one-time, label-free feature alignment based approach that efficiently converts foundational vision models into inherently interpretable B-cos variants. Once aligned, the B-cos backbone is used as a drop-in replacement across several downstream tasks--amortizing the cost of interpretability. ALOE is robust across pre-training paradigms (supervised, self-supervised, vision-language) and is 100-1000x more data-efficient than training from scratch. On classification, it outperforms fully-supervised B-cos models (e.g., +9.2 p.p. top-1 on ImageNet for ViT-B/16), retains strong linear probing, k-NN, and zero-shot transfer performance competitive with foundational backbones (DINOv3, SigLIP2) across diverse datasets. It also preserves spatially structured features useful for dense prediction, while yielding well-localized and highly human-interpretable explanations by design. Code link https://github.com/rmaser/aloe.

Early children's developmental trajectories set up a natural goal for sample-efficient pretraining of vision foundation models. We introduce BabyVLM-V2, a developmentally grounded framework for infant-inspired vision-language modeling that extensively improves upon BabyVLM-V1 through a longitudinal, multifaceted pretraining set, a versatile model, and, most importantly, DevCV Toolbox for cognitive evaluation. The pretraining set maximizes coverage while minimizing curation of a longitudinal, infant-centric audiovisual corpus, yielding video-utterance, image-utterance, and multi-turn conversational data that mirror infant experiences. DevCV Toolbox adapts all vision-related measures of the recently released NIH Baby Toolbox into a benchmark suite of ten multimodal tasks, covering spatial reasoning, memory, and vocabulary understanding aligned with early children's capabilities. Experimental results show that a compact model pretrained from scratch can achieve competitive performance on DevCV Toolbox, outperforming GPT-4o on some tasks. We hope the principled, unified BabyVLM-V2 framework will accelerate research in developmentally plausible pretraining of vision foundation models.

Large vision-language models (LVLMs) achieve strong performance on visual reasoning tasks but remain highly susceptible to hallucination. Existing detection methods predominantly rely on coarse, whole-image measures of how an object token relates to the input image. This global strategy is limited: hallucinated tokens may exhibit weak but widely scattered correlations across many local regions, which aggregate into deceptively high overall relevance, thus evading the current global hallucination detectors. We begin with a simple yet critical observation: a faithful object token must be strongly grounded in a specific image region. Building on this insight, we introduce a patch-level hallucination detection framework that examines fine-grained token-level interactions across model layers. Our analysis uncovers two characteristic signatures of hallucinated tokens: (i) they yield diffuse, non-localized attention patterns, in contrast to the compact, well-focused attention and (ii) they fail to exhibit meaningful semantic alignment with any visual region. Guided by these findings, we develop a lightweight and interpretable detection method that leverages patch-level statistical features, combined with hidden-layer representations. Our approach achieves up to 90% accuracy in token-level hallucination detection, demonstrating the superiority of fine-grained structural analysis for detecting hallucinations.

The paradigm of Multimodal Large Language Models (MLLMs) offers a promising blueprint for advancing the electromagnetic (EM) domain. However, prevailing approaches often deviate from the native MLLM paradigm, instead using task-specific or pipelined architectures that lead to fundamental limitations in model performance and generalization. Fully realizing the MLLM potential in EM domain requires overcoming three main challenges: (1) Data. The scarcity of high-quality datasets with paired EM signals and descriptive text annotations used for MLLMs pre-training; (2) Benchmark. The absence of comprehensive benchmarks to systematically evaluate and compare the performance of models on EM signal-to-text tasks; (3) Model. A critical fragility in low Signal-to-Noise Ratio (SNR) environments, where critical signal features can be obscured, leading to significant performance degradation. To address these challenges, we introduce a tripartite contribution to establish a foundation for MLLMs in the EM domain. First, to overcome data scarcity, we construct and release EM-100k, a large-scale dataset comprising over 100,000 EM signal-text pairs. Second, to enable rigorous and standardized evaluation, we propose EM-Bench, the most comprehensive benchmark featuring diverse downstream tasks spanning from perception to reasoning. Finally, to tackle the core modeling challenge, we present MERLIN, a novel training framework designed not only to align low-level signal representations with high-level semantic text, but also to explicitly enhance model robustness and performance in challenging low-SNR environments. Comprehensive experiments validate our method, showing that MERLIN is state-of-the-art in the EM-Bench and exhibits remarkable robustness in low-SNR settings.

Decoding visual stimuli from electroencephalography (EEG) signals is a crucial step toward practical brain-computer interfaces (BCIs). However, this task requires large-scale and high-quality EEG-image paired datasets. Compared with abundant image data, the limited EEG recordings restrict the decoding models' performance. To address this challenge, we propose EEGiT, a framework that converts sequential EEG signals into image-like EEG patches and enables the direct use of a pretrained Vision Transformer (ViT) as the EEG encoder. To preserve the spatial topology of brain regions and minimize distributional differences across channels, we group EEG electrodes according to anatomical structures and apply linear interpolation along the spatial dimension. We then resample the EEG signals to align the structure of EEG patches with that of image patches in ViT. This design encourages effective transfer of visual priors learned from large-scale image datasets to EEG representation learning. Experiments on the THINGS-EEG and EEG-3D datasets show that fine-tuning pretrained ViTs improves EEG-to-image retrieval and EEG-based visual classification, while maintaining robustness and strong cross-subject generalization. These results demonstrate a promising direction for leveraging powerful vision models to mitigate data scarcity in EEG decoding.

Interleaved multimodal generation enables capabilities beyond unimodal generation models, such as step-by-step instructional guides, visual planning, and generating visual drafts for reasoning. However, the quality of existing interleaved generation models under general instructions remains limited by insufficient training data and base model capacity. We present DuoGen, an interleaved generation framework that systematically addresses data curation, architecture design, and evaluation. On the data side, we build a large-scale, high-quality instruction-tuning dataset by combining multimodal conversations rewritten from curated raw websites, and diverse synthetic examples covering everyday scenarios. Architecturally, DuoGen leverages the strong visual understanding of a pretrained multimodal LLM and the visual generation capabilities of a diffusion transformer (DiT) pretrained on video generation, avoiding costly unimodal pretraining and enabling flexible base model selection. A two-stage decoupled strategy first instruction-tunes the MLLM, then aligns DiT with it using curated interleaved image-text sequences. Across public and newly proposed benchmarks, DuoGen outperforms prior open-source models in text quality, image fidelity, and image-context alignment, and also achieves state-of-the-art performance on text-to-image and image editing among unified generation models. Data and code are released at https://research.nvidia.com/labs/dir/duogen/

Vision foundation models (VFMs) have achieved strong performance across various vision tasks. However, it still remains challenging to apply VFMs for cross-domain few-shot segmentation (CD-FSS), which segments objects of novel classes under domain shifts using only a few labeled exemplars. The challenge is mainly driven by two factors: (1) limited labeled exemplars per novel class relative to the scale of VFM pre-training, making the model prone to overfitting during retraining, and (2) target-domain shifts underrepresented during pre-training, inducing cross-domain inconsistency and layerwise sensitivity. To address these issues, we propose Hierarchical Exemplar Representation Adaptation (HERA), a three-stage select-regularize-calibrate VFM-based segmentation framework that learns effectively from limited labels and adapts to novel domains without source-data retraining. We first design Hierarchical Layer Selection (HLS) to adaptively identify the most informative VFM layer using a data-dependent Exemplar Transfer Risk (ETR) computed for each candidate layer. Then, Prior-Guided Regularization (PGR) regularizes interactions on the selected representation, yielding well-structured local signals for the subsequent stage. Furthermore, Pixelwise Adaptive Calibration (PAC) combines the selected representation with the refined interaction maps to calibrate pixelwise predictions, producing consistent masks. Together, these stages form a hierarchical select-regularize-calibrate pipeline that guides frozen VFM features in new domains while fine-tuning less than 2.7% of parameters at test time. Extensive experiments show that HERA surpasses the state-of-the-art by more than 4.1 mIoU across multiple CD-FSS benchmarks.

Vision-Language Models (VLMs) have advanced rapidly within the unified Transformer architecture, yet their deployment on resource-constrained devices remains challenging due to high computational complexity. While pruning has emerged as an effective technique for compressing VLMs, existing approaches predominantly focus on a single mode by pruning either parameters or tokens, neglecting fully exploring the inherent redundancy in each mode, which leads to substantial performance degradation at high pruning ratios. To address the above limitations, we propose Collaborative Multi-Mode Pruning (CoMP), a novel framework tailored for VLMs by performing joint parameter and token pruning. Specifically, we first design a Collaborative Importance Metric (CIM) that investigates the mutual interference between the coupled parameters and tokens. It incorporates distinct significance of tokens into the computation of parameter importance scores, while simultaneously mitigating the affect of pruned parameters on token importance scores. Moreover, we develop a Multi-Mode Pruning Strategy (MPS) that decomposes the overall pruning process into a sequence of pruning stages, while in each stage we estimate the priory of different pruning modes based on their pruning cost and adaptively shift to the optimal one. Additionally, MPS integrates the historical cost and random exploration, in order to achieve a stable pruning process and avoid local optimum. Extensive experiments across various vision-language tasks and models demonstrate that our method effectively promotes the performance under high pruning ratios by comparing to the state-of-the-art approaches. The source code is available at https://github.com/Wuzimeng/CoMP.git.

We contend that embodied learning is fundamentally a lifecycle problem rather than a single-stage optimization. Systems that optimize only one link (data collection, simulation, learning, or deployment) rarely sustain improvement or generalize beyond narrow settings. We introduce Arcadia, a closed-loop framework that operationalizes embodied lifelong learning by tightly coupling four stages: (1) Self-evolving exploration and grounding for autonomous data acquisition in physical environments, (2) Generative scene reconstruction and augmentation for realistic and extensible scene creation, (3) a Shared embodied representation architecture that unifies navigation and manipulation within a single multimodal backbone, and (4) Sim-from-real evaluation and evolution that closes the feedback loop through simulation-based adaptation. This coupling is non-decomposable: removing any stage breaks the improvement loop and reverts to one-shot training. Arcadia delivers consistent gains on navigation and manipulation benchmarks and transfers robustly to physical robots, indicating that a tightly coupled lifecycle: continuous real-world data acquisition, generative simulation update, and shared-representation learning, supports lifelong improvement and end-to-end generalization. We release standardized interfaces enabling reproducible evaluation and cross-model comparison in reusable environments, positioning Arcadia as a scalable foundation for general-purpose embodied agents.

The rapid rise of highly realistic AI-generated images necessitates reliable and generalizable detection methods. However, existing methods are constrained by their discriminative nature: by learning a single static decision boundary, they tend to memorize generator-specific artifacts and consequently fail to generalize to the unseen distributions of new generative models. To overcome this limitation, we propose PPM-CLIP, a new framework that shifts from static classification to conditional generative modeling based on the CLIP vision-language model. Instead of learning a fixed decision boundary, a Probabilistic Prompt Modeling (PPM) module is used as a generator that produces an adaptive distribution of prompts according to the input image. This allows the model to flexibly capture novel artifacts, rather than matching them against fixed templates. In addition, to enhance the visual encoder's sensitivity to subtle artifacts, a Patch-Wise Contrastive Learning (PWCL) strategy is introduced. Extensive experiments on Ojha, GenImage, and DRCT benchmarks demonstrate that our generative paradigm significantly outperforms state-of-the-art methods, especially in cross-domain detection.

Recent advances in diffusion-based language models (DLMs) have shown remarkable potential for de novo protein design. However, enabling controllable protein generation requires integrating diverse biological conditions, such as structure, functions, and chemical interactions, each represented in distinct modalities. Existing approaches often either support a single condition or treat multiple conditions through separate modality-specific encoders. This isolation limits cross-modal interaction, reduces generation quality, and complicates the incorporation of new conditions without retraining or redesigning the backbone. To address these limitations, we introduce MMCP-GEN, a DLM for Multi-Modal, Multi-Condition Protein sequence GENeration. MMCP-GEN establishes a new paradigm for controllable protein generation under complex multimodal constraints. Its core is a modality-composable and extensible conditioning mechanism that fuses heterogeneous biological conditions via learnable queries and modality-indicator heads, enabling disentangled, extensible, and cross-modal condition integration without retraining the backbone. A joint generation-and-scoring objective further aligns sequence recovery with structural fidelity. Empirically, MMCP-GEN achieves state-of-the-art performance across structure-, function-, and ligand-conditioned tasks, improving sequence recovery by up to 5% and outperforming attentive baselines in diverse functional annotation tasks. These results establish MMCP-GEN as a general and high-fidelity framework for controllable protein generation. The source code is publicly available at https://github.com/WanyuGroup.

Story visualization aims to generate coherent image sequences that faithfully represent a narrative and match given character references. Despite progress in generative models, existing benchmarks remain narrow in scope, often limited to short prompts, lacking character references, or single-image cases, failing to reflect real-world narrative complexity and obscuring true model performance. We introduce ViStoryBench, a comprehensive benchmark designed to evaluate story visualization models across varied narrative structures, visual styles, and character settings. It features richly annotated multi-shot scripts derived from curated stories spanning literature, film, and folklore. Large language models assist in story summarization and script generation, with all outputs verified by humans for coherence and fidelity. Character references are carefully curated to maintain consistency across different artistic styles. ViStoryBench proposes a suite of multi-dimensional automated metrics to evaluate character consistency, style similarity, prompt alignment, aesthetic quality, and artifacts like copy-paste behavior. These metrics are validated through human studies and used to assess a broad range of open-source and commercial models, enabling systematic analysis and encouraging advances in visual storytelling.

Aligning egocentric video with wearable sensors has shown promise for human action recognition, but faces practical limitations in user discomfort, privacy concerns, and scalability. We explore exocentric video with ambient sensors as a non-intrusive, scalable alternative. However, the Global Alignment approach prevalent in egocentric-wearable settings fails in this new setting due to two problems: (P1) inability to capture local details such as subtle motions, and (P2) over-reliance on modality-invariant temporal features that distort negative relationships. To resolve these problems, we propose DETACH, a decomposed spatio-temporal alignment framework. By decomposing both modalities into spatial-temporal components, we preserve subtle temporal cues in videos from spatial features, and convert implicit sensor channel activations into explicit spatial prototypes via online clustering, thereby establishing cross-modal spatial grounding. To avoid over-reliance on temporal features, a spatial-temporal weighted contrastive loss leverages this grounding for fine-grained temporal alignment, prioritizing hard negatives and suppressing false negatives. Comprehensive experiments with downstream tasks on Opportunity++ and HWU-USP datasets demonstrate improvements of up to 30% in F1 and 50% in mAP over adapted egocentric-wearable baselines.

Video Large Language Models (Video-LLMs) are improving rapidly, yet current Video Question Answering (VideoQA) benchmarks often admit single-cue shortcuts, under-testing reasoning that must integrate evidence across time. We introduce HERBench, a benchmark designed to make multi-evidence integration unavoidable: each question requires at least three non-overlapping cues drawn from distinct video segments. HERBench contains 26,806 five-way multiple-choice questions across 12 compositional tasks. To make evidential demand measurable, we introduce the Minimum Required Frame-Set (MRFS), the smallest number of frames a model must fuse to answer correctly, and show that HERBench imposes higher evidential demand than prior benchmarks. Evaluating 13 state-of-the-art Video-LLMs yields only 31-42% accuracy, only modestly above the 20% random-guess baseline. We disentangle this failure into two critical bottlenecks: (1) a retrieval deficit, where frame selectors overlook key evidence, and (2) a fusion deficit, where models fail to integrate information even when all necessary evidence is provided. HERBench thus provides a principled benchmark for studying robust multi-evidence video understanding.

Multimodal person Re-IDentification (ReID) aims to reliably associate specific individuals by utilizing complementary information from heterogeneous modalities. In contrast, low-frequency radio frequency (RF) signals, with their superior penetration capability and illumination invariance, provide ideal complementary information. However, a key challenge in fusing these heterogeneous modalities lies in the conventional approach that relies heavily on cross-modal distribution matching, which often over-regularizes and weakens the discriminative capacity within each modality. Rather than enforcing direct distribution alignment, canonical correlation analysis (CCA) constructs a shared subspace that maximizes cross-modal correlation, inherently balancing modality specificity and shared semantics. Inspired by this, we reformulate cross-modal alignment as a correlation maximization problem, avoiding direct constraints on feature distributions and guiding the model to harmonize intra-modal discriminative learning with cross-modal alignment. Specifically, VRCLIP first refines CLIP's visual encoder with illumination-disentangling objectives, then aligns RGB and RF embeddings in a canonical correlation subspace, and finally employs an RF-anchored reliability gate for adaptive fusion. To advance the area, we will release VRR, the first large-scale vision-radio ReID dataset with over 650K paired image-radar samples and position annotations for 31 participants. Extensive experiments show state-of-the-art 93.9% mAP and robust generalization across diverse lighting and occlusion conditions.

Recent large vision-language models (LVLMs) have demonstrated impressive reasoning ability by generating long chain-of-thought (CoT) responses. However, CoT reasoning in multimodal contexts is highly vulnerable to visual hallucination propagation: once an intermediate reasoning step becomes inconsistent with the visual evidence, subsequent steps-even if logically valid-can still lead to incorrect final answers. Existing solutions attempt to mitigate this issue by training models to "think with images" via reinforcement learning (RL). While effective, these methods are costly, model-specific, and difficult to generalize across architectures. Differently, we present a lightweight method that bypasses RL training and provides an iterative, training-free, plug-and-play framework for visually-grounded multimodal reasoning. Our key idea is to supervise each reasoning step at test time with visual evidence, ensuring that every decoded token is justified by corresponding visual cues. Concretely, we construct a textual visual-evidence pool that guides the model's reasoning generation. When existing evidence is insufficient, a visual decider module dynamically extracts additional relevant evidence from the image based on the ongoing reasoning context, expanding the pool until the model achieves sufficient visual certainty to terminate reasoning and produce the final answer. Extensive experiments on multiple LVLM backbones and benchmarks demonstrate the effectiveness of our approach. Our method achieves 16.5%-29.5% improvements on TreeBench and 13.7% RH-AUC gains on RH-Bench, substantially reducing hallucination rates while improving reasoning accuracy without additional training.