Searcharxiv⌕ Search

arXiv subjects

Zhen Dong

Publications and source records attributed to Zhen Dong.

At least 37 records · Page 2Linked to original sources
arXiv2026

HorizonDrive: Self-Corrective Autoregressive World Model for Long-horizon Driving Simulation

Closed-loop driving simulation requires real-time interaction beyond short offline clips, pushing current driving world models toward autoregressive (AR) rollout. Existing AR distillation approaches typically rely on frame sinks or student-side degradation training. The former transfers poorly to driving due to fast ego-motion and rapid scene changes, while the latter remains bounded by the teacher's single-pass output length and thus provides only a limited supervision horizon. A natural question is: can the teacher itself be extended via AR rollout to provide unbounded-horizon supervision at bounded memory cost? The key difficulty is that a standard teacher drifts under its own predictions, contaminating the supervision it provides. Our key insight is to make the teacher rollout-capable, ensuring reliable supervision from its own AR rollouts. This is instantiated as HorizonDrive, an anti-drifting training-and-distillation framework for AR driving simulation. First, scheduled rollout recovery (SRR) trains the base model to reconstruct ground-truth future clips from prediction-corrupted histories, yielding a teacher that remains stable across long AR rollouts. Second, the rollout-capable teacher is extended via AR rollout, providing long-horizon distribution-matching supervision under bounded memory, while a short-window student aligns to it with teacher rollout DMD (TRD) for efficient real-time deployment. HorizonDrive natively supports minute-scale AR rollout under bounded memory; on nuScenes, HorizonDrive reduces FID by 52% and FVD by 37%, and lowers ARE and DTW by 21% and 9% relative to the strongest long-horizon streaming baselines, while remaining competitive with single-pass driving video generators.

cs.CV↗
arXiv2026

OSAQ: Outlier Self-Absorption for Accurate Low-bit LLM Quantization

Large Language Models (LLMs) have demonstrated remarkable capabilities. However, their massive parameter scale leads to significant resource consumption and latency during inference. Post-training weight-only quantization offers a promising solution by reducing model size and accelerating token generation through alleviating the memory-bound issue. Nevertheless, the presence of inherent systematic outliers in weights continues to be a major obstacle. While existing methods, such as scaling and rotation, attempt to address this issue, the performance remains unsatisfactory. In this paper, we propose Outlier Self-Absorption Quantization (OSAQ), which performs additive weight suppression guided by the second-order low-rank property for low-bit weight-only quantization of LLMs. Specifically, we observe that the Hessian exhibits low-rank consistency across different inputs, with certain directions consistently showing vanishing curvature. Leveraging this property, we identify a stable null space of the Hessian and then construct an additive weight transformation by linearly combining the vectors within this null space, thereby suppressing weight outliers without affecting the task loss. This additive transformation can be absorbed into the weights offline, requiring no inter-layer transformations and introducing no inference overhead. Moreover, the construction is efficiently achieved by a closed-form solution, without resource-intensive training or iterative procedures. Extensive experiments demonstrate that OSAQ effectively suppresses outliers and enhances low-bit quantization performance. For instance, in 2-bit quantization, OSAQ, when integrated with GPTQ, achieves over 40% lower perplexity compared to vanilla GPTQ.

cs.LG↗
arXiv2026

DPG-CD: Depth-Prior-Guided Cross-Modal Joint 2D-3D Change Detection

Urban spatial evolution is manifested not only through horizontal expansion but also through vertical structural changes. Consequently, jointly capturing 2D semantic changes and 3D height changes is essential for urban morphology analysis and emergency management. In practical scenarios, collecting 3D observations is often constrained by high acquisition costs and the inability to support frequent updates. The multi-temporal cross-modal input consisting of pre-event Digital Surface Model (DSM) and post-event imagery provides a practical solution for 3D change detection in high-frequency urban monitoring, disaster assessment, and emergency response scenarios. However, this setting remains challenging as imagery and DSM data exhibit significant spectral-geometric representation gaps. Moreover, modality differences may be confused with actual changes, and robust change detection requires effective fusion of semantic and geometric features from multi-temporal data. In this paper, we propose DPG-CD, a depth-prior-guided multi-temporal cross-modal fusion framework for joint 2D semantic and 3D height change detection. Specifically, an estimated depth prior is introduced into the imagery to mitigate the modality gap with DSM. A gated fusion mechanism then selectively injects geometric cues from depth prior while preserving discriminative spectral representations. Subsequently, a multi-stage cross-temporal cross-modal feature fusion architecture is employed to extract change-aware features. Finally, a multi-task decoder jointly predicts 2D semantic changes and 3D height changes, complemented by an auxiliary DSM prediction task to improve structural consistency and height estimation accuracy. Experiments on two public datasets, Hi-BCD and 3DCD, and a new dataset, NYC-MMCD, demonstrate that DPG-CD outperforms state-of-the-art methods on both 2D and 3D change detection tasks.

cs.CV↗
arXiv2026

Arena as Offline Reward: Efficient Fine-Grained Preference Optimization for Diffusion Models

Reinforcement learning from human feedback (RLHF) effectively promotes preference alignment of text-to-image (T2I) diffusion models. To improve computational efficiency, direct preference optimization (DPO), which avoids explicit reward modeling, has been widely studied. However, its reliance on binary feedback limits it to coarse-grained modeling on chosen-rejected pairs, resulting in suboptimal optimization. In this paper, we propose ArenaPO, which leverages Arena scores as offline rewards to provide refined feedback, thus achieving efficient and fine-grained optimization without a reward model. This enables ArenaPO to benefit from both the rich rewards of traditional RLHF and the efficiency of DPO. Specifically, we first construct a model Arena in which each model's capability is represented as a Gaussian distribution, and infer these capabilities by traversing the annotated pairwise preferences. Each output image is treated as a sample from the corresponding capability distribution. Then, for a image pair, conditioned on the two capability distributions and the observed pairwise preference, the absolute quality gap is estimated using latent-variable inference based on truncated normal distribution, which serves as fine-grained feedback during training. It does not require a reward model and can be computed offline, thus introducing no additional training overhead. We conduct ArenaPO training on Pick-a-Pic v2 and HPD v3 datasets, showing that ArenaPO consistently outperforms existing baselines.

cs.CV↗
arXiv2026

TOL: Textual Localization with OpenStreetMap

Natural language provides an intuitive way to express spatial intent in geospatial applications. While existing localization methods often rely on dense point cloud maps or high-resolution imagery, OpenStreetMap (OSM) offers a compact and freely available map representation that encodes rich semantic and structural information, making it well-suited for large-scale localization. However, text-to-OSM (T2O) localization remains largely unexplored. In this paper, we formulate the T2O localization task, which aims to estimate accurate 2D positions in urban environments from textual scene descriptions without relying on geometric observations or GNSS-based initial location. To support the proposed task, we introduce TOL, a large-scale benchmark spanning multiple continents and diverse urban environments. TOL contains approximately 121K textual queries paired with OSM map tiles and covers about 316 km of road trajectories across Boston, Karlsruhe, and Singapore. We further propose TOLoc, a coarse-to-fine localization framework that explicitly models the semantics of surrounding objects and their directional information. In the coarse stage, direction-aware features are extracted from both textual descriptions and OSM tiles to construct global descriptors, which are used to retrieve candidate locations for the query. In the fine stage, the query text and top-1 retrieved tile are jointly processed, where a dedicated alignment module fuses the textual descriptor and local map features to regress the 2-DoF pose. Experimental results demonstrate that TOLoc achieves strong localization performance, outperforming the best existing method by 6.53\%, 9.93\%, and 8.32\% at 5 m, 10 m, and 25 m thresholds, respectively, and shows strong generalization to unseen environments. Dataset, code and models will be publicly available at: https://github.com/WHU-USI3DV/TOL.

cs.CV↗
arXiv2026

TS-Attn: Temporal-wise Separable Attention for Multi-Event Video Generation

Generating high-quality videos from complex temporal descriptions that contain multiple sequential actions is a key unsolved problem. Existing methods are constrained by an inherent trade-off: using multiple short prompts fed sequentially into the model improves action fidelity but compromises temporal consistency, while a single complex prompt preserves consistency at the cost of prompt-following capability. We attribute this problem to two primary causes: 1) temporal misalignment between video content and the prompt, and 2) conflicting attention coupling between motion-related visual objects and their associated text conditions. To address these challenges, we propose a novel, training-free attention mechanism, Temporal-wise Separable Attention (TS-Attn), which dynamically rearranges attention distribution to ensure temporal awareness and global coherence in multi-event scenarios. TS-Attn can be seamlessly integrated into various pre-trained text-to-video models, boosting StoryEval-Bench scores by 33.5% and 16.4% on Wan2.1-T2V-14B and Wan2.2-T2V-A14B with only a 2% increase in inference time. It also supports plug-and-play usage across models for multi-event image-to-video generation. The source code and project page are available at https://github.com/Hong-yu-Zhang/TS-Attn.

cs.CV↗
arXiv2026

Style-Decoupled Adaptive Routing Network for Underwater Image Enhancement

Underwater Image Enhancement (UIE) is essential for robust visual perception in marine applications. However, existing methods predominantly rely on uniform mapping tailored to average dataset distributions, leading to over-processing mildly degraded images or insufficient recovery for severe ones. To address this challenge, we propose a novel adaptive enhancement framework, SDAR-Net. Unlike existing uniform paradigms, it first decouples specific degradation styles from the input and subsequently modulates the enhancement process adaptively. Specifically, since underwater degradation primarily shifts the appearance while keeping the scene structure, SDAR-Net formulates image features into dynamic degradation style embeddings and static scene structural representations through a carefully designed training framework. Subsequently, we introduce an adaptive routing mechanism. By evaluating style features and adaptively predicting soft weights at different enhancement states, it guides the weighted fusion of the corresponding image representations, accurately satisfying the adaptive restoration demands of each image. Extensive experiments show that SDAR-Net achieves a new state-of-the-art (SOTA) performance with a PSNR of 25.72 dB on real-world benchmark, and demonstrates its utility in downstream vision tasks. Our code is available at https://github.com/WHU-USI3DV/SDAR-Net.

cs.CV↗
arXiv2026

AWARE: Adaptive Whole-body Active Rotating Control for Enhanced LiDAR-Inertial Odometry under Human-in-the-Loop Interaction

Human-in-the-loop (HITL) UAV operation is essential in complex and safety-critical aerial surveying environments, where human operators provide navigation intent while onboard autonomy must maintain accurate and robust state estimation. A key challenge in this setting is that resource-constrained UAV platforms are often limited to narrow-field-of-view LiDAR sensors. In geometrically degenerate or feature-sparse scenes, limited sensing coverage often weakens LiDAR Inertial Odometry (LIO)'s observability, causing drift accumulation, degraded geometric accuracy, and unstable state estimation, which directly compromise safe and effective HITL operation and the reliability of downstream surveying products. To overcome this limitation, we present AWARE, a bio-inspired whole-body active yawing framework that exploits the UAV's own rotational agility to extend the effective sensor horizon and improve LIO's observability without additional mechanical actuation. The core of AWARE is a differentiable Model Predictive Control (MPC) framework embedded in a Reinforcement Learning (RL) loop. It first identifies the viewing direction that maximizes information gain across the full yaw space, and a lightweight RL agent then adjusts the MPC cost weights online according to the current environmental context, enabling an adaptive balance between estimation accuracy and flight stability. A Safe Flight Corridor mechanism further ensures operational safety within this HITL paradigm by decoupling the operator's navigational intent from autonomous yaw optimization to enable safe and efficient cooperative control. We validate AWARE through extensive experiments in diverse simulated and real-world environments.

cs.RO↗
arXiv2026

Denoise and Align: Towards Source-Free UDA for Robust Panoramic Semantic Segmentation

Panoramic semantic segmentation is pivotal for comprehensive 360° scene understanding in critical applications like autonomous driving and virtual reality. However, progress in this domain is constrained by two key challenges: the severe geometric distortions inherent in panoramic projections and the prohibitive cost of dense annotation. While Unsupervised Domain Adaptation (UDA) from label-rich pinhole-camera datasets offers a viable alternative, many real-world tasks impose a stricter source-free (SFUDA) constraint where source data is inaccessible for privacy or proprietary reasons. This constraint significantly amplifies the core problems of domain shift, leading to unreliable pseudo-labels and dramatic performance degradation, particularly for minority classes. To overcome these limitations, we propose the DAPASS framework. DAPASS introduces two synergistic modules to robustly transfer knowledge without source data. First, our Panoramic Confidence-Guided Denoising (PCGD) module generates high-fidelity, class-balanced pseudo-labels by enforcing perturbation consistency and incorporating neighborhood-level confidence to filter noise. Second, a Contextual Resolution Adversarial Module (CRAM) explicitly addresses scale variance and distortion by adversarially aligning fine-grained details from high-resolution crops with global semantics from low-resolution contexts. DAPASS achieves state-of-the-art performances on outdoor (Cityscapes-to-DensePASS) and indoor (Stanford2D3D) benchmarks, yielding 55.04% (+2.05%) and 70.38% (+1.54%) mIoU, respectively.

cs.CV↗
arXiv2026

QFT: Quantized Full-parameter Tuning of LLMs with Affordable Resources

Large Language Models (LLMs) have showcased remarkable impacts across a wide spectrum of natural language processing tasks. Fine-tuning these pretrained models on downstream datasets provides further significant performance gains; however, this process typically requires a large number of expensive, high-end GPUs. Although there have been efforts focused on parameter-efficient fine-tuning, they cannot fully unlock the powerful potential of full-parameter fine-tuning. In this paper, we propose QFT, a Quantized Full-parameter Tuning framework for LLMs that quantizes and stores all training states, including weights, gradients, and optimizer states, in INT8 format to reduce training memory, thereby enabling full-parameter fine-tuning on existing GPUs at an affordable cost. To ensure training performance, we make two key efforts: i) for quantized gradients and optimizer states, we theoretically prove that the Lion optimizer, with its property of consistent update magnitudes, is highly robust to quantization; ii) and for quantized weights, we employ the hybrid feature quantizer, which identifies and protects a small subset of sparse critical features while quantizing the remaining dense features, thus ensuring accurate weight updates without FP32 backups. Moreover, to support backpropagation in the integer context, we develop a stack-based gradient flow scheme with O(1) complexity, forming a unified integer training pipeline. As a result, QFT reduces the model state memory to 21% of the standard solution while achieving comparable performance, e.g., tuning a LLaMA-7B model requires only <30GB of memory, making it feasible on a single A6000 GPU.

cs.CL↗
arXiv2026

ButterflyQuant: Ultra-low-bit LLM Quantization through Learnable Orthogonal Butterfly Transforms

Large language models require massive memory footprints, severely limiting deployment on consumer hardware. Quantization reduces memory through lower numerical precision, but extreme 2-bit quantization suffers from catastrophic performance loss due to outliers in activations. Rotation-based methods such as QuIP and QuaRot apply orthogonal transforms to eliminate outliers before quantization, using computational invariance: $\mathbf{y} = \mathbf{Wx} = (\mathbf{WQ}^T)(\mathbf{Qx})$ for orthogonal $\mathbf{Q}$. However, these methods use fixed transforms--Hadamard matrices achieving optimal worst-case coherence $μ= 1/\sqrt{n}$--that cannot adapt to specific weight distributions. We identify that different transformer layers exhibit distinct outlier patterns, motivating layer-adaptive rotations rather than one-size-fits-all approaches. In this work, we propose ButterflyQuant, which replaces Hadamard rotations with learnable butterfly transforms parameterized by continuous Givens rotation angles. Unlike Hadamard's discrete $\{+1, -1\}$ entries that are non-differentiable and thus prohibit gradient-based learning, butterfly transforms' continuous parameterization enables smooth optimization while guaranteeing orthogonality by construction. This orthogonal constraint ensures theoretical guarantees in outlier suppression while achieving $O(n \log n)$ computational complexity with only $\frac{n \log n}{2}$ learnable parameters. We further introduce a uniformity regularization on post-transformation activations to promote smoother distributions amenable to quantization. Learning requires only 128 calibration samples and converges in minutes on a single GPU.

cs.LG↗
arXiv2026

K-Sort Eval: Efficient Preference Evaluation for Visual Generation via Corrected VLM-as-a-Judge

The rapid development of visual generative models raises the need for more scalable and human-aligned evaluation methods. While the crowdsourced Arena platforms offer human preference assessments by collecting human votes, they are costly and time-consuming, inherently limiting their scalability. Leveraging vision-language model (VLMs) as substitutes for manual judgments presents a promising solution. However, the inherent hallucinations and biases of VLMs hinder alignment with human preferences, thus compromising evaluation reliability. Additionally, the static evaluation approach lead to low efficiency. In this paper, we propose K-Sort Eval, a reliable and efficient VLM-based evaluation framework that integrates posterior correction and dynamic matching. Specifically, we curate a high-quality dataset from thousands of human votes in K-Sort Arena, with each instance containing the outputs and rankings of K models. When evaluating a new model, it undergoes (K+1)-wise free-for-all comparisons with existing models, and the VLM provide the rankings. To enhance alignment and reliability, we propose a posterior correction method, which adaptively corrects the posterior probability in Bayesian updating based on the consistency between the VLM prediction and human supervision. Moreover, we propose a dynamic matching strategy, which balances uncertainty and diversity to maximize the expected benefit of each comparison, thus ensuring more efficient evaluation. Extensive experiments show that K-Sort Eval delivers evaluation results consistent with K-Sort Arena, typically requiring fewer than 90 model runs, demonstrating both its efficiency and reliability.

cs.CV↗
arXiv2026

AnchoredDream: Zero-Shot 360° Indoor Scene Generation from a Single View via Geometric Grounding

Single-view indoor scene generation plays a crucial role in a range of real-world applications. However, generating a complete 360° scene from a single image remains a highly ill-posed and challenging problem. Recent approaches have made progress by leveraging diffusion models and depth estimation networks, yet they still struggle to maintain appearance consistency and geometric plausibility under large viewpoint changes, limiting their effectiveness in full-scene generation. To address this, we propose AnchoredDream, a novel zero-shot pipeline that anchors 360° scene generation on high-fidelity geometry via an appearance-geometry mutual boosting mechanism. Given a single-view image, our method first performs appearance-guided geometry generation to construct a reliable 3D scene layout. Then, we progressively generate the complete scene through a series of modules: warp-and-inpaint, warp-and-refine, post-optimization, and a novel Grouting Block, which ensures seamless transitions between the input view and generated regions. Extensive experiments demonstrate that AnchoredDream outperforms existing methods by a large margin in both appearance consistency and geometric plausibility--all in a zero-shot manner. Our results highlight the potential of geometric grounding for high-quality, zero-shot single-view scene generation.

cs.CV↗
arXiv2026

Expert Knowledge-Guided Decision Calibration for Accurate Fine-Grained Tree Species Classification

Accurate fine-grained tree species classification is critical for forest inventory and biodiversity monitoring. Existing methods predominantly focus on designing complex architectures to fit local data distributions. However, they often overlook the long-tailed distributions and high inter-class similarity inherent in limited data, thereby struggling to distinguish between few-shot or confusing categories. In the process of knowledge dissemination in the human world, individuals will actively seek expert assistance to transcend the limitations of local thinking. Inspired by this, we introduce an external "Domain Expert" and propose an Expert Knowledge-Guided Classification Decision Calibration Network (EKDC-Net) to overcome these challenges. Our framework addresses two core issues: expert knowledge extraction and utilization. Specifically, we first develop a Local Prior Guided Knowledge Extraction Module (LPKEM). By leveraging Class Activation Map (CAM) analysis, LPKEM guides the domain expert to focus exclusively on discriminative features essential for classification. Subsequently, to effectively integrate this knowledge, we design an Uncertainty-Guided Decision Calibration Module (UDCM). This module dynamically corrects the local model's decisions by considering both overall category uncertainty and instance-level prediction uncertainty. Furthermore, we present a large-scale classification dataset covering 102 tree species, named CU-Tree102 to address the issue of scarce diversity in current benchmarks. Experiments on three benchmark datasets demonstrate that our approach achieves state-of-the-art performance. Crucially, as a lightweight plug-and-play module, EKDC-Net improves backbone accuracy by 6.42% and precision by 11.46% using only 0.08M additional learnable parameters. The dataset, code, and pre-trained models are available at https://github.com/WHU-USI3DV/TreeCLS.

cs.CV↗
arXiv2026

R-KV: Redundancy-aware KV Cache Compression for Reasoning Models

Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 16% of the KV cache. This KV-cache reduction also leads to a 90% memory saving and a 6.6X throughput over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.

cs.CL↗
arXiv2026

SVII-3D: Advancing Roadside Infrastructure Inventory with Decimeter-level 3D Localization and Comprehension from Sparse Street Imagery

The automated creation of digital twins and precise asset inventories is a critical task in smart city construction and facility lifecycle management. However, utilizing cost-effective sparse imagery remains challenging due to limited robustness, inaccurate localization, and a lack of fine-grained state understanding. To address these limitations, SVII-3D, a unified framework for holistic asset digitization, is proposed. First, LoRA fine-tuned open-set detection is fused with a spatial-attention matching network to robustly associate observations across sparse views. Second, a geometry-guided refinement mechanism is introduced to resolve structural errors, achieving precise decimeter-level 3D localization. Third, transcending static geometric mapping, a Vision-Language Model agent leveraging multi-modal prompting is incorporated to automatically diagnose fine-grained operational states. Experiments demonstrate that SVII-3D significantly improves identification accuracy and minimizes localization errors. Consequently, this framework offers a scalable, cost-effective solution for high-fidelity infrastructure digitization, effectively bridging the gap between sparse perception and automated intelligent maintenance.

cs.CV↗
arXiv2026

Unleashing the Capabilities of Large Vision-Language Models for Intelligent Perception of Roadside Infrastructure

Automated perception of urban roadside infrastructure is crucial for smart city management, yet general-purpose models often struggle to capture the necessary fine-grained attributes and domain rules. While Large Vision Language Models (VLMs) excel at open-world recognition, they often struggle to accurately interpret complex facility states in compliance with engineering standards, leading to unreliable performance in real-world applications. To address this, we propose a domain-adapted framework that transforms VLMs into specialized agents for intelligent infrastructure analysis. Our approach integrates a data-efficient fine-tuning strategy with a knowledge-grounded reasoning mechanism. Specifically, we leverage open-vocabulary fine-tuning on Grounding DINO to robustly localize diverse assets with minimal supervision, followed by LoRA-based adaptation on Qwen-VL for deep semantic attribute reasoning. To mitigate hallucinations and enforce professional compliance, we introduce a dual-modality Retrieval-Augmented Generation (RAG) module that dynamically retrieves authoritative industry standards and visual exemplars during inference. Evaluated on a comprehensive new dataset of urban roadside scenes, our framework achieves a detection performance of 58.9 mAP and an attribute recognition accuracy of 95.5%, demonstrating a robust solution for intelligent infrastructure monitoring.

cs.CV↗
arXiv2025

Can World Models Benefit VLMs for World Dynamics?

Trained on internet-scale video data, generative world models are increasingly recognized as powerful world simulators that can generate consistent and plausible dynamics over structure, motion, and physics. This raises a natural question: with the advent of strong video foundational models, might they supplant conventional vision encoder paradigms for general-purpose multimodal understanding? While recent studies have begun to explore the potential of world models on common vision tasks, these explorations typically lack a systematic investigation of generic, multimodal tasks. In this work, we strive to investigate the capabilities when world model priors are transferred into Vision-Language Models: we re-purpose a video diffusion model as a generative encoder to perform a single denoising step and treat the resulting latents as a set of visual embedding. We empirically investigate this class of models, which we refer to as World-Language Models (WorldLMs), and we find that generative encoders can capture latents useful for downstream understanding that show distinctions from conventional encoders. Naming our best-performing variant Dynamic Vision Aligner (DyVA), we further discover that this method significantly enhances spatial reasoning abilities and enables single-image models to perform multi-frame reasoning. Through the curation of a suite of visual reasoning tasks, we find DyVA to surpass both open-source and proprietary baselines, achieving state-of-the-art or comparable performance. We attribute these gains to WorldLM's inherited motion-consistency internalization from video pre-training. Finally, we systematically explore extensive model designs to highlight promising directions for future work. We hope our study can pave the way for a new family of VLMs that leverage priors from world models and are on a promising path towards generalist vision learners.

cs.CV↗