Searcharxiv⌕ Search

arXiv subjects

Dezhi Luo

Publications and source records attributed to Dezhi Luo.

15 recordsLinked to original sources
arXiv2026

VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning

Native visual reasoning treats visual generation as the medium of reasoning itself: visual states (i.e. images and videos) are not merely inputs to be understood or outputs to be rendered, but first-class substrates for problem solving beyond language. Yet progress remains bottlenecked by the lack of scalable training tasks, reliable feedback, and controlled comparisons across generative substrates. In this work, we introduce VBVR-Pro, a closed-loop testbed that makes native visual reasoning through generation trainable, verifiable, optimizable, and experimentally controllable. 1) Task scaling. VBVR-Pro turns visual reasoning into a controlled task space of 300 procedurally generated tasks. Models trained on VBVR-Pro show strong transfer beyond the proposed suite across seven external visual reasoning benchmarks such as RISE-Video, MME-CoF-Pro, and BabyVision. 2) Verifiable rewards. VBVR-Pro provides verifiable reward scorers for task-grounded evaluation. Through a systematic study of leading MLLMs as judges, we identify recurring failure modes of the prevalent VLM-as-a-judge paradigm. In contrast, the proposed scorers are grounded in deterministic, task-specific rules, achieve fine-grained alignment with human judgments. Importantly, they serve as reliable reward signals for large-scale multi-task reinforcement learning and demonstrate stronger post-RL performance across visual reasoning tasks. 3) Mechanism study. VBVR-Pro enables controlled modality studies across more than 30 image, video, and interleaved generators. Our analysis shows that video generation remains strongest for tasks requiring persistent spatiotemporal state tracking, while interleaved generation provides a compute-efficient alternative. Critically, ablations and probing suggest the presence of vision-native trajectories that are crucial to visual reasoning. We release all data, models, scorers, and code.

cs.CV↗
arXiv2026

A Very Big Video Reasoning Suite

Rapid progress in video models has largely focused on visual quality, leaving their reasoning capabilities underexplored. Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture, enabling intuitive reasoning over spatiotemporal structure such as continuity, interaction, and causality. However, systematically studying video reasoning and its scaling behavior is hindered by the lack of large-scale training data. To address this gap, we introduce the Very Big Video Reasoning (VBVR) Dataset, an unprecedentedly large-scale resource spanning 200 curated reasoning tasks following a principled taxonomy and over one million video clips, approximately three orders of magnitude larger than existing datasets. We further present VBVR-Bench, a verifiable evaluation framework that moves beyond model-based judging by incorporating rule-based, human-aligned scorers, enabling reproducible and interpretable diagnosis of video reasoning capabilities. Leveraging the VBVR suite, we conduct one of the first large-scale scaling studies of video reasoning and observe early signs of emergent generalization to unseen reasoning tasks. Together, VBVR lays a foundation for the next stage of research in generalizable video reasoning. The data, benchmark toolkit, and models are publicly available at https://video-reason.com/?v=vbvr .

cs.CV↗
arXiv2026

Egocentric Bias in Vision-Language Models

Visual perspective taking--inferring how the world appears from another's viewpoint--is foundational to social cognition. We introduce FlipSet, a diagnostic benchmark for Level-2 visual perspective taking (L2 VPT) in vision-language models. The task requires simulating 180-degree rotations of 2D character strings from another agent's perspective, isolating spatial transformation from 3D scene complexity. Evaluating 103 VLMs reveals systematic egocentric bias: the vast majority perform below chance, with roughly three-quarters of errors reproducing the camera viewpoint. Control experiments expose a compositional deficit--models achieve high theory-of-mind accuracy and above-chance mental rotation in isolation, yet fail catastrophically when integration is required. This dissociation indicates that current VLMs lack the mechanisms needed to bind social awareness to spatial operations, suggesting fundamental limitations in model-based spatial reasoning. FlipSet provides a cognitively grounded testbed for diagnosing perspective-taking capabilities in multimodal systems.

cs.CV↗
arXiv2026

Vision-Language Models Mistake Head Orientation for Gaze Direction: Nonverbal Conversation Cues

Where someone looks is a nonverbal communication cue that children and adults readily use. How well can Vision-Language Models (VLMs) infer gaze targets? To construct evaluation stimuli, we captured 1,360 real-world photos of scenes in which a person gazes at one of several objects on a table. Importantly, we also controlled the gazer's head orientation: sometimes it was directed toward the gaze target, sometimes toward a distractor object, and sometimes left unconstrained. We found a substantial performance gap between VLMs and humans, ruled out alternative explanations such as resolution and object-naming skills, and identified the main reason for the gap as VLMs inferring gaze direction using head orientation rather than eye appearance. Such a bias is likely due to data rather than architecture, as suggested by a proof-of-concept experiment finetuning a transformer-based vision model. Future work should investigate whether these findings hold broadly across various deep learning methods trained on existing data, and whether better data mitigates this problem for all architectures. Pinpointing the reason sets the stage for technologies that can interpret gaze targets to have more efficient interactions with humans.

cs.CV↗
arXiv2026

Vision Language Models Cannot Reason About Physical Transformation

Understanding physical transformations is fundamental for reasoning in dynamic environments. While Vision Language Models (VLMs) show promise in embodied applications, whether they genuinely understand physical transformations remains unclear. We introduce ConservationBench evaluating conservation -- whether physical quantities remain invariant under transformations. Spanning four properties with paired conserving/non-conserving scenarios, we generate and evaluate 23,040 questions across 112 VLMs. Results reveal systematic failure: performance remains near chance with improvements on conservation tasks accompanied by drops on controls. Control experiments show strong textual priors favoring invariance, yet models perform worse with actual visual content when performance is balanced across conserving and non-conserving scenarios. Neither temporal resolution, prompting, nor curated sampling helps. These findings show that current VLMs fail to maintain transformation-invariant representations of physical properties across dynamic scenes.

cs.AI↗
arXiv2026

Increasing Computation Resolves Conflicts in Vision Language Models

Cognitive control, the ability to coordinate competing information sources in pursuit of goals, is fundamental to intelligent behavior. We systematically investigate whether Vision Language Models (VLMs) exhibit cognitive control and how computational resources modulate conflict resolution. We construct a benchmark of 4,410 tasks across seven conflict paradigms (Stroop, Flanker, and five realistic variants) spanning multiple difficulty levels and visual complexities, testing 47 VLMs with rigorous experimental control. We find that VLMs exhibit robust congruency effects across all tasks, with larger models systematically resolving conflicts more effectively than smaller models. Critically, VLMs reproduce the fine-grained demand-resource relationship observed in human temporal dynamics: larger models drop below chance on incongruent high-conflict trials while smaller models fail to meaningfully engage and perform at chance, mirroring human behavior at short processing times and establishing parameter count as a proxy for conflict resolution capacity. These findings demonstrate that human-like cognitive control emerges from optimization dynamics in large-scale neural networks, suggesting that adaptive flexibility under conflict may naturally arise through scaling.

cs.NE↗
arXiv2026

Probing Perceptual Constancy in Large Vision-Language Models

Perceptual constancy is the ability to maintain stable perceptions of objects despite changes in sensory input, such as variations in distance, angle, or lighting. This ability is crucial for visual understanding in a dynamic world. Here, we explored such ability in current Vision Language Models (VLMs). In this study, we evaluated 155 VLMs using 236 experiments across three domains: color, size, and shape constancy. The experiments included single-image and video adaptations of classic cognitive tasks, along with novel tasks in in-the-wild conditions. We found significant variability in VLM performance across these domains, with model performance in shape constancy clearly dissociated from that of color and size constancy.

cs.CV↗
arXiv2025

Rethinking the Simulation vs. Rendering Dichotomy: No Free Lunch in Spatial World Modelling

Spatial world models, representations that support flexible reasoning about spatial relations, are central to developing computational models that could operate in the physical world, but their precise mechanistic underpinnings are nuanced by the borrowing of underspecified or misguided accounts of human cognition. This paper revisits the simulation versus rendering dichotomy and draws on evidence from aphantasia to argue that fine-grained perceptual content is critical for model-based spatial reasoning. Drawing on recent research into the neural basis of visual awareness, we propose that spatial simulation and perceptual experience depend on shared representational geometries captured by higher-order indices of perceptual relations. We argue that recent developments in embodied AI support this claim, where rich perceptual details improve performance on physics-based world engagements. To this end, we call for the development of architectures capable of maintaining structured perceptual representations as a step toward spatial world modelling in AI.

q-bio.NC↗
arXiv2025

The Philosophical Foundations of Growing AI Like A Child

Despite excelling in high-level reasoning, current language models lack robustness in real-world scenarios and perform poorly on fundamental problem-solving tasks that are intuitive to humans. This paper argues that both challenges stem from a core discrepancy between human and machine cognitive development. While both systems rely on increasing representational power, the absence of core knowledge, foundational cognitive structures in humans, prevents language models from developing robust, generalizable abilities, where complex skills are grounded in simpler ones within their respective domains. It explores empirical evidence of core knowledge in humans, analyzes why language models fail to acquire it, and argues that this limitation is not an inherent architectural constraint. Finally, it outlines a workable proposal for systematically integrating core knowledge into future multi-modal language models through the large-scale generation of synthetic training data using a cognitive prototyping strategy.

cs.AI↗
arXiv2025

Probing Mechanical Reasoning in Large Vision Language Models

Mechanical reasoning is a hallmark of human intelligence, defined by its ubiquitous yet irreplaceable role in human activities ranging from routine tasks to civil engineering. Embedding machines with mechanical reasoning is therefore an important step towards building human-level artificial intelligence. Here, we leveraged 155 cognitive experiments to test the understanding of system stability, gears and pulley systems, leverage principle, inertia and motion, and fluid mechanics in 26 Vision Language Models (VLMs). Results indicate that VLMs consistently perform worse than humans on all domains, while demonstrate significant difficulty in reasoning about gear systems and fluid mechanics. Notably, their performance on these tasks do not improve as number of parameters increase, suggesting that current attention-based architecture may fail to grasp certain underlying mechanisms required for mechanical reasoning, particularly those pertaining to mental simulations.

cs.AI↗
arXiv2025

Vision Language Models Know Law of Conservation without Understanding More-or-Less

Understanding law of conservation is a critical milestone in human cognitive development considered to be supported by the apprehension of quantitative concepts and the reversibility of operations. To assess whether this critical component of human intelligence has emerged in Vision Language Models, we have curated the ConserveBench, a battery of 365 cognitive experiments across four dimensions of physical quantities: volume, solid quantity, length, and number. The former two involve transformational tasks which require reversibility understanding. The latter two involve non-transformational tasks which assess quantity understanding. Surprisingly, we find that while Vision Language Models are generally good at transformational tasks, they tend to fail at non-transformational tasks. There is a dissociation between understanding the reversibility of operations and understanding the concept of quantity, which both are believed to be the cornerstones of understanding law of conservation in humans.

cs.AI↗
arXiv2025

Vision Language Models See What You Want but not What You See

Knowing others' intentions and taking others' perspectives are two core components of human intelligence that are considered to be instantiations of theory-of-mind. Infiltrating machines with these abilities is an important step towards building human-level artificial intelligence. Here, to investigate intentionality understanding and level-2 perspective-taking in Vision Language Models (VLMs), we constructed the IntentBench and PerspectBench, which together contains over 300 cognitive experiments grounded in real-world scenarios and classic cognitive tasks. We found VLMs achieving high performance on intentionality understanding but low performance on level-2 perspective-taking. This suggests a potential dissociation between simulation-based and theory-based theory-of-mind abilities in VLMs, highlighting the concern that they are not capable of using model-based reasoning to infer others' mental states.

cs.AI↗
arXiv2025

Vision-Language Models Are Not Pragmatically Competent in Referring Expression Generation

Referring Expression Generation (REG) is a core task for evaluating the pragmatic competence of vision-language systems, requiring not only accurate semantic grounding but also adherence to principles of cooperative communication (Grice, 1975). However, current evaluations of vision-language models (VLMs) often overlook the pragmatic dimension, reducing REG to a region-based captioning task and neglecting Gricean maxims. In this work, we revisit REG from a pragmatic perspective, introducing a new dataset (RefOI) of 1.5k images annotated with both written and spoken referring expressions. Through a systematic evaluation of state-of-the-art VLMs, we identify three key failures of pragmatic competence: (1) failure to uniquely identify the referent, (2) inclusion of excessive or irrelevant information, and (3) misalignment with human pragmatic preference, such as the underuse of minimal spatial cues. We also show that standard automatic evaluations fail to capture these pragmatic violations, reinforcing superficial cues rather than genuine referential success. Our findings call for a renewed focus on pragmatically informed models and evaluation frameworks that align with real human communication.

cs.CL↗
arXiv2025

Core Knowledge Deficits in Multi-Modal Language Models

While Multi-modal Large Language Models (MLLMs) demonstrate impressive abilities over high-level perception and reasoning, their robustness in the wild remains limited, often falling short on tasks that are intuitive and effortless for humans. We examine the hypothesis that these deficiencies stem from the absence of core knowledge--rudimentary cognitive abilities innate to humans from early childhood. To explore the core knowledge representation in MLLMs, we introduce CoreCognition, a large-scale benchmark encompassing 12 core knowledge concepts grounded in developmental cognitive science. We evaluate 230 models with 11 different prompts, leading to a total of 2,530 data points for analysis. Our experiments uncover four key findings, collectively demonstrating core knowledge deficits in MLLMs: they consistently underperform and show reduced, or even absent, scalability on low-level abilities relative to high-level ones. Finally, we propose Concept Hacking, a novel controlled evaluation method that reveals MLLMs fail to progress toward genuine core knowledge understanding, but instead rely on shortcut learning as they scale.

cs.CL↗