CVPR 2026 Papers

We introduce NitroGen, a vision-action foundation model for generalist gaming agents that is trained on 40,000 hours of gameplay videos across more than 1,000 games. We scale embodied agents through three key ingredients: 1) an internet-scale video-action dataset constructed by automatically extracting player actions from publicly available gameplay videos, 2) a multi-game benchmark environment that can measure cross-game generalization, and 3) a unified vision-action model trained with large-scale behavior cloning. NitroGen exhibits competence across diverse domains, including combat encounters in 3D action games, high-precision control in 2D platformers, and exploration in procedurally generated worlds. When fine-tuned, pre-training transfers effectively to unseen games, achieving up to 52% relative improvement in task success rates over models trained from scratch. We release the dataset, evaluation suite, and model weights to advance research on generalist embodied agents.

Embodied navigation for long-horizon tasks, guided by complex natural language instructions, remains a formidable challenge in artificial intelligence. Existing agents often struggle with robust long-term planning about unseen environments, leading to high failure rates. To address these limitations, we introduce NavForesee, a novel Vision-Language Model (VLM) that unifies high-level language planning and predictive world model imagination within a single, unified framework.Our approach empowers a single VLM to concurrently perform planning and predictive foresight. Conditioned on the full instruction and historical observations, the model is trained to understand the navigation instructions by decomposing the task, tracking its progress, and formulating the subsequent sub-goal. Simultaneously, it functions as a generative world model, providing crucial foresight by predicting short-term environmental dynamics and long-term navigation milestones. The VLM's structured plan guides its targeted prediction, while the imagined future provides rich context to inform the navigation actions, creating a powerful internal feedback loop of perception-planning/prediction-action. We demonstrate through extensive experiments on the R2R-CE and RxR-CE benchmark that NavForesee achieves highly competitive performance in complex scenarios. Our work highlights the immense potential of fusing explicit language planning with implicit spatiotemporal prediction, paving the way for more intelligent and capable embodied agents.

Short-form videos have become a primary medium for digital advertising, requiring scalable and efficient content creation. However, current workflows and AI tools remain disjoint and modality-specific, leading to high production costs and low overall efficiency. To address this issue, we propose AutoCut, an end-to-end advertisement video editing framework based on multimodal discretization and controllable editing. AutoCut employs dedicated encoders to extract video and audio features, then applies residual vector quantization to discretize them into unified tokens aligned with textual representations, constructing a shared video-audio-text token space. Built upon a foundation model, we further develop a multimodal large language model for video editing through combined multimodal alignment and supervised fine-tuning, supporting tasks covering video selection and ordering, script generation, and background music selection within a unified editing framework. Finally, a complete production pipeline converts the predicted token sequences into deployable long video outputs. Experiments on real-world advertisement datasets show that AutoCut reduces production cost and iteration time while substantially improving consistency and controllability, paving the way for scalable video creation. Code and data are available at https://github.com/AdAutoCut/Autocut

The rapid evolution of video generative models has shifted their focus from producing visually plausible outputs to tackling tasks requiring physical plausibility and logical consistency. However, despite recent breakthroughs such as Veo 3's chain-of-frames reasoning, it remains unclear whether these models can exhibit reasoning capabilities similar to large language models (LLMs). Existing benchmarks predominantly evaluate visual fidelity and temporal coherence, failing to capture higher-order reasoning abilities. To bridge this gap, we propose **TiViBench**, a hierarchical benchmark specifically designed to evaluate the reasoning capabilities of image-to-video (I2V) generation models. TiViBench systematically assesses reasoning across four dimensions: i) **Structural Reasoning & Search**, ii) **Spatial & Visual Pattern Reasoning**, iii) **Symbolic & Logical Reasoning**, and iv) **Action Planning & Task Execution**, spanning 24 diverse task scenarios across 3 difficulty levels. Through extensive evaluations, we show that commercial models (*e.g.*, Sora 2, Veo 3.1) demonstrate stronger reasoning potential, while open-source models reveal untapped potential that remains hindered by limited training scale and data diversity. To further unlock this potential, we introduce **VideoTPO**, a simple yet effective test-time strategy inspired by preference optimization. By performing LLM self-analysis on generated candidates to identify strengths and weaknesses, VideoTPO significantly enhances reasoning performance without requiring additional training, data, or reward models. Together, TiViBench and VideoTPO pave the way for evaluating and advancing reasoning in video generation models, setting a foundation for future research in this emerging field.

Test-time adaptation (TTA) has emerged as a promising solution to address real world domain shifts in medical image segmentation. Current approaches adapt by updating or regularizing a pre-trained source model. However, they face two major issues: (i) the source models on which they rely are prone to overfitting under domain shifts; (ii) in dynamic continual testing scenarios, error accumulation and class forgetting are further exacerbated. To overcome these limitations, we propose TanGo, a novel framework that combines Training to adapt with Foundation Guidance and continual style calibration. During training, TanGo learns generalization priors from vision foundation models (VFMs) through distribution-wise consistency learning. We incorporate stable low-frequency representations from a frozen encoder of VFMs as priors to guide the source model, constraining its output feature distribution to yield a more generalizable feature space. At test time, we propose an instance-wise style calibration method that uses a learnable data decorator to shift dynamic test images back toward an enhanced source-like distribution. Subsequently, a series of constraints is applied to draw the decorated test samples closer to the enhanced source distribution while maintaining their semantic integrity, thus boosting the continual adaptability of the source model. Extensive experiments on multiple medical image segmentation tasks demonstrate that TanGo achieves state-of-the-art performance.

Video-based commonsense captioning aims to generate captions for the video content while providing multiple commonsense about the underlying events. Existing approaches rely on constructing a "video -> content caption -> commonsense" reasoning chain, which generates visually ungrounded commonsense and neglects inter-category commonsense correlations. Firstly, the existing reasoning chain induces the model's excessive reliance on content caption when generating commonsense, resulting in generic outputs with limited visual relevance. Secondly, the reasoning chain adopts multiple isolated decoders for commonsense generation, which fails to leverage the correlations between different categories of commonsense. To address these limitations, we introduce a novel self-critical distillation network (SCD-Net), which optimizes the reasoning chain by enhancing visual reasoning and establishing inter-category commonsense correlations. Specifically, on the one hand, we introduce self-critical learning and design a reward function to allow the model to refine its output. This mechanism incentivizes the model to maximize the utilization of visual information, thus improving the model's capacity for visual comprehension. On the other hand, we propose a joint reasoning distillation framework that fosters mutual inference among diverse commonsense categories. In this framework, we incorporate the cascaded decoder and knowledge distillation strategy to facilitate inter-category commonsense knowledge transfer while maintaining the fairness of the testing. Our experiments on the large-scale Video-to-Commonsense dataset demonstrate that our approach performs favorably against state-of-the-art methods. The code will be released soon.

While Vision-Language Models (VLMs) have achieved remarkable performance, their Euclidean embeddings remain limited in capturing hierarchical relationships such as part-to-whole or parent-child structures, and often face challenges in multi-object compositional scenarios. Hyperbolic VLMs mitigate this issue by better preserving hierarchical structures and modeling part-whole relations (i.e., whole scene and its part images) through entailment. However, existing approaches do not model that each part has a different level of semantic representativeness to the whole. We propose UNcertainty-guided Compositional Hyperbolic Alignment (UNCHA) for enhancing hyperbolic VLMs. UNCHA models part-to-whole semantic representativeness with hyperbolic uncertainty, by assigning lower uncertainty to more representative parts and higher uncertainty to less representative ones for the whole scene. This representativeness is then incorporated into the contrastive objective with uncertainty-guided weights. Finally, the uncertainty is further calibrated with an entailment loss regularized with entropy-based term. With the proposed losses, UNCHA learns hyperbolic embeddings with more accurate part-whole ordering, capturing the underlying compositional structure in an image and improving its understanding of complex multi-object scenes. UNCHA achieves state-of-the-art performance on zero-shot classification, retrieval, and multi-label classification benchmarks. Our code and models are available at: https://github.com/jeeit17/UNCHA.git.

Learning general-purpose representations of geographic locations has become essential to geospatial tasks such as population estimation and environmental monitoring. To obtain such representations, multimodal geo-foundation models often use contrastive learning (CL) to align satellite imagery with geo-coordinates, implicitly assuming that cross-modal (shared) information suffices for downstream tasks. However, given the breadth of tasks, task-relevant information may lie beyond the shared space, so retaining modality-specific (unique) features can improve task performance. Prior methods retain unique information through extra training objectives or external models, increasing training complexity. Motivated by the conventional wisdom that earlier layers capture general input features while later layers become task-specific, we hypothesize that intermediate layers in CL models retain more modality-specific structure than the alignment-optimized final layer. Through a trifecta layerwise analysis of modality gap, representation similarity, and mutual information, we validate this trend and find that fusing intermediate (more unique) and final (more shared) representations yields consistent gains on diverse geospatial tasks. Our findings reveal underutilized information diversity in CL models and show that simple layerwise fusion is an efficient path to richer geo-embeddings.

Streaming video reasoning requires models to operate in a setting where history grows without bound while meaningful evidence remains scarce. In such a landscape, relevant signal is like an oasis -- small, critical, and easily lost in a desert of redundancy. Enlarging memory only widens the desert; aggressive compression dries up the oasis. The real difficulty lies in discovering where to look, not how much to remember. We therefore introduce OASIS, a novel framework for streaming video reasoning that tackles this challenge through structured, on-demand retrieval. It organizes streaming history into hierarchical events and performs reasoning as controlled refinement -- short-context inference first, followed by semantically grounded retrieval only when uncertainty arises. As the retrieval is driven by high-level intent rather than embedding similarity, the retrieved memory is substantially more accurate and less noisy. Additionally, the mechanism is plug-and-play, training-free, and readily attaches to different streaming MLLM backbones. Experiments across multiple benchmarks and backbones show that OASIS achieves strong gains in long-horizon accuracy and compositional reasoning with bounded token cost and low request delay.

We present MetaSpectra+, a compact multifunctional camera that supports two operating modes: (1) snapshot HDR + hyperspectral or (2) snapshot polarization + hyperspectral imaging. It utilizes a novel metasurface-refractive assembly that splits the incident beam into multiple channels and independently controls each channel's dispersion, exposure, and polarization. Unlike prior multifunctional metasurface imagers restricted to narrow (10--100 nm) bands, MetaSpectra+ operates over nearly the entire visible spectrum (250 nm). Relative to snapshot hyperspectral imagers, it achieves the shortest total track length and the highest reconstruction accuracy on benchmark datasets. The demonstrated prototype reconstructs high-quality hyperspectral datacubes and either an HDR image or two orthogonal polarization channels from a snapshot measurement.

Single-image reflection removal (SIRR) is a highly ill-posed and computationally demanding problem. Existing CNN or Transformer-based methods often rely on large receptive fields and heavy computation, limiting their deployment on resource-constrained devices. To address this, we propose LightRR, a lightweight yet effective reflection removal network that unifies a wavelet-based mechanism and State Space Modeling (SSM).Specifically, we introduce an Asymmetric Frequency Mamba Block (AFM), which leverages the Discrete Wavelet Transform (DWT) to decompose features into low- and high-frequency components. This allows for targeted modeling of frequency-specific dependencies via Mamba-based state space dynamics.This design not only captures long-range context efficiently but also reduces spatial resolution and computation while preserving critical details.Furthermore, a knowledge distillation-enhanced encoder allows the network to inherit the representational power of large pre-trained models during training, enabling lightweight inference.Extensive experiments on multiple real-world benchmarks demonstrate that LightRR achieves performance comparable to state-of-the-art methods, while using only 3.01% of the parameters and 5.22% of the FLOPs (vs. RDNet), highlighting its superior balance between accuracy and efficiency.

Open-vocabulary semantic segmentation extends pixel-level recognition to arbitrary text-described categories. Despite strong global semantic understanding, vision-language models such as CLIP exhibit limited spatial precision and semantic ambiguity across large vocabularies, constraining their effectiveness for dense prediction. We present S2C2Seg, a training-free framework that integrates with existing methods through Category Subset Selection (CSS) and Consistent Semantic Guidance (CSG). CSS employs three complementary scoring functions to filter category candidates: CLIP-based global semantic similarity, spatial presence from dense prediction models, and multi-view consistency via alignment and conditional entropy. This joint exploitation of semantic, spatial, and consistency cues reduces category redundancy and semantic ambiguity. CSG adaptively fuses CLIP global features with local spatial predictions through category-specific confidence weighting, applying stronger regularization to high-similarity categories for correcting prediction biases while preserving spatial precision for low-confidence categories. Extensive experiments across eight benchmarks demonstrate broad applicability: when integrated with SCLIP, ProxyCLIP, and CorrCLIP, S2C2Seg achieves consistent improvements of 3.4 to 9.7 percentage points in mIoU, establishing a new state-of-the-art of 51.2% average mIoU.

We study concept-level forgetting in pretrained vision models: removing an entire semantic category so the system no longer recognizes that object in unseen images and contexts, rather than merely forgetting specific training examples. Prior work either applies blunt global projections or fine-tunes parameters, which can introduce collateral damage to unrelated features, add compute, and become unstable as forgetting strength increases. We introduce Contrastive Subnet Erasure (CSE), a training-free, encoder-centric edit that targets a compact set of channels most responsible for the class and attenuates them in a calibrated manner. The modification is algebraically folded into the subsequent layer, yielding no inference-time overhead and leaving task heads unchanged. To evaluate whether forgetting generalizes beyond the data used to specify the class, we introduce a cross dataset protocol in which the class is defined on a source dataset and performance is measured on a disjoint target dataset drawn from a different distribution with no shared images. This setup tests whether the model still fails to recognize the object when it looks different or appears in new scenes, and it helps avoid overfitting to patterns in the source dataset. Across CIFAR 10, CIFAR 100, and ImageNet under this protocol, CSE achieves stronger forgetting of the target class while better preserving non target utility than existing baselines in both single class and multi class settings. Overall, CSE provides a simple, stable, and deployment-ready mechanism for class-level unlearning in vision.

Vision-Language Models (VLMs) often struggle with robust 3D spatial reasoning. Prevailing methods that rely on fine-tuning with 3D visual question-answering (VQA) datasets may overfit dataset-specific biases, while integrating specialized 3D visual encoders is often inflexible and cumbersome. In this paper, we argue that genuine spatial understanding should emerge from learning fundamental geometric priors, not only from high-level VQA supervision. We propose GASP (Geometric-Aware Spatial Priors), a framework that injects these priors directly into the LLM's transformer layers. GASP employs a small correspondence head, applied as a deep supervision signal across all layers, and is trained with a dual objective leveraging ground-truth geometry from large-scale video scenes: a contrastive loss on ground-truth point correspondences enforces 2D view-invariance, while a depth consistency supervision resolves 3D geometric ambiguities. Our analysis first provides a diagnostic showing that standard VLMs' internal correspondence matching accuracy is very low (often below 5%). We then demonstrate that our training substantially improves this behavior, boosting peak layer-wise correspondence to over 70% and maintaining over 85% temporal robustness while baselines remain below 5%. These internal improvements translate to significant gains on downstream spatial benchmarks---including +18.2% on All-Angles Bench and +29.0% on VSI-Bench---all without training on any 3D VQA data. Our findings indicate that learning from fundamental geometric priors is a promising and generalizable pathway towards VLMs with more reliable 3D spatial reasoning.

Visible-Infrared Person Re-Identification (VI-ReID) is a challenging retrieval task due to the substantial modality gap between visible and infrared images. While existing methods attempt to bridge this gap by learning modality-invariant features within a shared embedding space, they often overlook the complex and implicit correlations between modalities. This limitation becomes more severe under distribution shifts, where infrared samples are often far fewer than visible ones. To address these challenges, we propose a novel network termed Bi-directional Interaction Transformation (BIT). Instead of relying on rigid feature alignment, BIT adopts a matching-based strategy that explicitly models the interaction between visible and infrared image pairs. Specifically, BIT employs an encoder-decoder architecture where the encoder extracts preliminary feature representations, and the decoder performs bi-directional feature integration and query aware scoring to enhance cross-modality correspondence. To our best knowledge, BIT is the first to introduce such pairwise matching-driven interaction in VI-ReID. Extensive experiments on several benchmarks demonstrate that our BIT achieves state-of-the-art performance, highlighting its effectiveness in the VI-ReID task.

Medical vision-language pretraining (VLP) models have recently been investigated for their generalization to diverse downstream tasks. However, current medical VLP methods typically force the model to learn simple and complex concepts simultaneously. This anti-cognitive process leads to suboptimal feature representations, especially under distribution shift. To address this limitation, we propose a Knowledge-driven Cognitive Orchestration for Medical VLP (MedKCO) that involves both the ordering of the pretraining data and the learning objective of vision-language contrast. Specifically, we design a two level curriculum by incorporating diagnostic sensitivity and intra-class sample representativeness for the ordering of the pretraining data. Moreover, considering the inter-class similarity of medical images, we introduce a self-paced asymmetric contrastive loss to dynamically adjust the participation of the pretraining objective. We evaluate the proposed pretraining method on three medical imaging scenarios in multiple vision-language downstream tasks, and compare it with several curriculum learning methods. Extensive experiments show that our method significantly surpasses all baselines. https://github.com/Mr-Talon/MedKCO.

Recent advances in Vision-Language Models (VLMs) have benefited from Reinforcement Learning (RL) for enhanced reasoning. However, existing methods still face critical limitations, including the lack of low-level visual information and effective visual feedback. To address these problems, this paper proposes a unified multimodal interleaved reasoning framework ForeSight, which enables VLMs to See Further with low-level visual cues and Think Deeper with effective visual feedback. First, it introduces a set of low-level visual tools to integrate essential visual information into the reasoning chain, mitigating the neglect of fine-grained visual features. Second, a mask-based visual feedback mechanism is elaborated to incorporate visual reflection into the thinking process, enabling the model to dynamically re-examine and update its answers. Driven by RL, ForeSight learns to autonomously decide on tool invocation and answer verification, with the final answer accuracy as the reward signal. To evaluate the performance of the proposed framework, we construct a new dataset, Character and Grounding SalBench (CG-SalBench), based on the SalBench dataset. Experimental results demonstrate that the ForeSight-7B model significantly outperforms other models with the same parameter scale, and even surpasses the current SOTA closed-source models on certain metrics.

Generalist Anomaly Detection (GAD) seeks to overcome the domain-specific limitations of traditional anomaly detection by training a unified model that can generalize to unseen classes. A promising GAD strategy involves using residual features to create a class-invariant space. However, existing methods that directly model the distribution of residuals face unpredictable risks: there is inconsistency between residual and instance features, i.e., subtle defects may yield small residuals (false negatives), or normal feature residuals could be large due to the diversity of normality (false positives). To address these limitations, we propose a novel residual-based learning framework that re-purposes residuals as a guide to learn instance-level normality, rather than modeling their distribution directly. Our framework features two new attention-based modules: Residual Feature Learning (RFL), which uses learnable proxies to capture diverse patterns from the residual features, and Normality Learning from Support (NLS), which leverages these residual proxies to aggregate query-related normality proxies from the support instance features. These dynamically generated normality proxies are then used to hunt for normality within the query patch features, enabling accurate anomaly localization. Extensive experiments on GAD benchmarks demonstrate the effectiveness of our method. Code is available at HNQ-GAD.

Systems such as video chatbots and navigation robots often depend on streaming image captioning to interpret visual inputs. Existing approaches typically employ large multimodal language models (MLLMs) for this purpose, but their substantial computational cost hinders practical application.This limitation motivates our development of a lightweight captioning model. Our investigation begins by replacing the large-scale language component in MLLMs with a compact 125M-parameter model.Surprisingly, this compact model, despite a 93x reduction in size, achieves comparable performance to MLLMs, suggesting that factual image captioning does not significantly require the complex reasoning abilities of LLMs. Despite this promising result, our lightweight model still lacks reliability. To address this, we draw inspiration from the human visual process: perceiving a global and coarse understanding of the scene before attending to finer details. Accordingly, we propose a multimodal self-refinement framework that guides the model to utilize features from salient regions, identified by referencing the previous coarse caption, and to produce a refined description. Experimental results demonstrate the superiority of our model in both single-sentence and detailed captioning, extending even to long-range video QA tasks.