Open-world semantic segmentation presently relies significantly on extensive image-text pair datasets, which often suffer from a lack of fine-grained pixel annotations on sufficient categories. The acquisition of such data is rendered economically prohibitive due to the substantial investments of both human labor and time. In light of the formidable image generation capabilities of diffusion models, we introduce a novel diffusion models-driven pipeline for automatically generating datasets tailored to the needs of open-world semantic segmentation, named “MagicSeg”. Our MagicSeg initiates from class labels and proceeds to generate high-fidelity textual descriptions, which in turn, serve as guidance for the diffusion model to generate images. Rather than only generating positive samples for each label, our process encompasses the simultaneous generation of corresponding negative images, designed to serve as paired counterfactual samples for contrastive training. Then to provide a self-supervised signal for open-world segmentation pretraining, our MagicSeg integrates an open-vocabulary detection model and interactive segmentation model to extract object masks as precise segmentation labels from images based on the provided category labels. By applying our dataset to the contrastive language-image pretraining model with the pseudo mask supervision and the auxiliary counterfactual contrastive training, the downstream model obtains strong performance on open-world semantic segmentation. We evaluate our model on PASCAL VOC, PASCAL Context, and COCO, achieving SOTA with performance of 62.9%, 26.7%, and 40.2% respectively, demonstrating our dataset effectiveness in enhancing open-world semantic segmentation capabilities.

Recent multimodal large language models (MLLMs) have made remarkable progress on fine-grained perception tasks under the "Thinking with Images" (TwI) paradigm by iteratively performing various visual tool operations. However, this paradigm relies heavily on frequent external tool calls and repeated image re-encoding, which leads to substantial computational overhead and inference latency. To address these issues, we propose Beyond the Eye (BEE), a novel implicit visual tool paradigm centered on self-regulated capability. BEE directly incorporates visual tool invocation behaviors into the training objective and encourages the model to develop a self-regulated invocation mechanism. This design enables the model to adaptively balance internal knowledge and implicit tools, avoiding redundant tool usage while substantially reducing inference latency. Specifically, BEE involves a two-stage training process: (1) Formalized Chain-of-Thought (CoT) Supervised Fine-tuning (SFT). We construct CoT trajectories with structured tool slots and mixed invocation states. This stage activates the model's implicit tool representations and adaptive switching capability. (2) Self-regulated Reward-Driven Alignment. To address redundant tool usage caused by ambiguous cognitive boundaries, we first introduce the Net Tool Gain (NTG) metric to quantify this phenomenon. Based on this observation, we further propose a self-regulated reward mechanism. This mechanism penalizes ineffective tool dependency and encourages the model to perform knowledge routing, ensuring that implicit tools are invoked only when the model's internal knowledge is insufficient. BEE achieves state-of-the-art performance in fine-grained visual perception while remaining competitive in general reasoning tasks and achieving substantial gains in inference efficiency.

In this paper, we introduce SemHiTok, a unified image Tokenizer via Semantic-Guided Hierarchical codebook (SGHC) that provides consistent discrete representations for multimodal understanding and generation. Recently, unified image tokenizers have sparked exploration within the research community, which is designed to capture high-level semantic features for understanding and retaining low-level pixel features for generation. Previous works attempt to train a unified image tokenizer by combining loss for semantic distillation and pixel reconstruction. However, due to the differing levels of features prioritized by multimodal understanding and generation, joint training methods face significant challenges in achieving a good trade-off. SemHiTok addresses this challenge through a novel semantic-guided hierarchical codebook, which builds pixel sub-codebooks on a pretrained semantic codebook. This design decouples semantic and pixel in both terms of structure and training strategy, enabling the tokenizer to capture pixel features while retaining its ability to comprehend high-level semantic information. Our experiments demonstrate that SemHiTok achieves SOTA performance in image reconstruction and multimodal understanding under the LLaVA-v1.5 setting. Further, we develop a unified MLLM with SemHiTok, which exhibits superior performance across multimodal understanding and generation tasks. For understanding, SemHiTok achieves impressive performance on most benchmarks. For generation, our model achieves SOTA performance on MJHQ30K in unified MLLMs. Our code and models will be open source.

Recent advances in Visual-Language-Action (VLA) models have shown promising potential for robotic manipulation tasks. However, real-world robotic tasks often involve long-horizon, multi-step problem-solving and require generalization for continual skill acquisition, extending beyond single actions or skills. These challenges present significant barriers for existing VLA models, which use monolithic action decoders trained on aggregated data, resulting in poor scalability. To address these challenges, we propose AtomicVLA, a unified planning-and-execution framework that jointly generates task-level plans, atomic skill abstractions, and fine-grained actions. AtomicVLA constructs a scalable atomic skill library through a Skill-Guided Mixture-of-Experts (SG-MoE), where each expert specializes in mastering generic yet precise atomic skills. Furthermore, we introduce a flexible routing encoder that automatically assigns dedicated atomic experts to new skills, enabling continual learning. We validate our approach through extensive experiments. In simulation, AtomicVLA outperforms $π_{0}$ by 2.4\% on LIBERO, 10\% on LIBERO-LONG, and outperforms $π_{0}$ and $π_{0.5}$ by 0.22 and 0.25 in average task length on CALVIN. Additionally, our AtomicVLA consistently surpasses baselines by 18.3\% and 21\% in real-world long-horizon tasks and continual learning. These results highlight the effectiveness of atomic skill abstraction and dynamic expert composition for long-horizon and lifelong robotic tasks. The project page is \href{https://zhanglk9.github.io/atomicvla-web/}{here}.

End-to-end driving requires a coherent understanding of future scenes, yet existing methods model these scenes using task-specific heads and output formats, with limited scalability. Can video generation instead provide a shared predictor? We introduce SUV, a unified end-to-end driving framework that casts future Scene Understanding as Video generation using a pretrained video foundation model. SUV models future appearance, semantics, relative depth, and instance-level dynamics as video streams with a shared video expert, without stream-specific visual prediction heads. Through joint video-action attention, the action expert attends to the latent representations of all future streams and generates the ego trajectory. Experiments show that SUV directly predicts all four future streams, while controlled ablations show that structured future supervision and direct future-stream access yield higher trajectory planning scores. With only a single front camera and no candidate-trajectory selection, SUV outperforms a broad set of recent state-of-the-art methods on both NAVSIM-v2 splits, achieving 91.0 EPDMS on navtest and 36.9 on navhard. On the long-tail WOD-E2E benchmark, SUV achieves a competitive RFS of 7.94.

Vision-Language-Action (VLA) models have emerged as a prominent paradigm for end-to-end autonomous driving; however, their efficient deployment is severely constrained by high computational latency and exposure bias arising from sequential autoregressive decoding. Conversely, while specialized diffusion policies enable low-latency, parallel execution, training them from scratch typically yields narrow, single-task architectures that lack holistic visual-linguistic reasoning. Successfully transforming pre-trained autoregressive generalists into parallel diffusion models could combine multi-task cognitive intelligence with execution efficiency, yet this transition presents a formidable architectural challenge due to mismatched attention patterns (causal versus bidirectional) and divergent optimization objectives. To bridge this divide, we introduce WAM-Diff2, a multi-task discrete diffusion VLA framework powered by a three-stage hierarchical distillation strategy. By structuring the architectural shift through progressive block-wise adaptation, block-wise distillation, and model-wise cross-scale distillation, WAM-Diff2 preserves the underlying semantic foundations of the base model while accelerating inference. Extensive evaluations across driving understanding, perception, and planning benchmarks demonstrate that WAM-Diff2 effectively mitigates exposure bias and achieves performance parity with autoregressive baselines. Crucially, the autoregressive-to-diffusion transition yields a 2.8x decoding speedup, which scales to an ultimate 15.1x acceleration when combined with system-level optimizations including FlashInfer and CUDA Graphs.

Recent advances in multimodal large language models (MLLMs) have shown impressive reasoning capabilities. However, further enhancing existing MLLMs necessitates high-quality vision-language datasets with carefully curated task complexities, which are both costly and challenging to scale. Although recent self-improving models that iteratively refine themselves offer a feasible solution, they still suffer from two core challenges: (i) most existing methods augment visual or textual data separately, resulting in discrepancies in data complexity (e.g., over-simplified diagrams paired with redundant textual descriptions); and (ii) the evolution of data and models is also separated, leading to scenarios where models are exposed to tasks with mismatched difficulty levels. To address these issues, we propose C2-Evo, an automatic, closed-loop self-improving framework that jointly evolves both training data and model capabilities. Specifically, given a base dataset and a base model, C2-Evo enhances them by a cross-modal data evolution loop and a data-model evolution loop. The former loop expands the base dataset by generating complex multimodal problems that combine structured textual sub-problems with iteratively specified geometric diagrams, while the latter loop adaptively selects the generated problems based on the performance of the base model, to conduct supervised fine-tuning and reinforcement learning alternately. Consequently, our method continuously refines its model and training data, and consistently obtains considerable performance gains across multiple mathematical reasoning benchmarks. Our code, models, and datasets will be released.

Recent studies have revealed the potential of training open-source Large Language Models (LLMs) to unleash LLMs' reasoning ability for enhancing vision-language navigation (VLN) performance, and simultaneously mitigate the domain gap between LLMs' training corpus and the VLN task. However, these approaches predominantly adopt straightforward input-output mapping paradigms, causing the mapping learning difficult and the navigational decisions unexplainable. Chain-of-Thought (CoT) training is a promising way to improve both navigational decision accuracy and interpretability, while the complexity of the navigation task makes the perfect CoT labels unavailable and may lead to overfitting through pure CoT supervised fine-tuning. To address these issues, we propose EvolveNav, a novel sElf-improving embodied reasoning paradigm that realizes adaptable and generalizable navigational reasoning for boosting LLM-based vision-language Navigation. Specifically, EvolveNav involves a two-stage training process: (1) Formalized CoT Supervised Fine-Tuning, where we train the model with curated formalized CoT labels to first activate the model's navigational reasoning capabilities, and simultaneously increase the reasoning speed; (2) Self-Reflective Post-Training, where the model is iteratively trained with its own reasoning outputs as self-enriched CoT labels to enhance the supervision diversity. A self-reflective auxiliary task is also designed to encourage the model to learn correct reasoning patterns by contrasting with wrong ones. Experimental results under both task-specific and cross-task training paradigms demonstrate the consistent superiority of EvolveNav over previous LLM-based VLN approaches on various popular benchmarks, including R2R, REVERIE, CVDN, and SOON. EvolveNav open avenues for exploring effective self-improving reasoning paradigms, enabling building agents capable of self-evolving for promoting LLM-based embodied AI research.

Recent Vision-Language-Action (VLA) models have advanced end-to-end autonomous driving by incorporating reasoning for better interpretability and planning quality. However, most existing approaches directly generate the final trajectory without explicitly examining its future consequences, which limits their reliability in complex and dynamic environments. To address this limitation, we propose IRR-Drive (Intend, Reflect, Refine), an adaptive multimodal reflection framework for autonomous driving. Specifically, to tightly couple high-level reasoning with physical constraints, IRR-Drive first generates a preliminary textual intention and anticipates potential interactions by predicting future semantic bird's-eye view (BEV) representations. This dual-modality (Text + BEV) reflection space explicitly models anticipated scene evolution, enabling the model to rigorously self-correct and refine its initial intent before generating the final trajectory. Furthermore, to balance planning performance and computational efficiency, we construct reflection-oriented training data and design an adaptive reflection reward, enabling the model to adaptively select its reasoning mode according to scene complexity. Instead of using reasoning primarily as an auxiliary interpretation, IRR-Drive directly integrates an adaptive reflection mechanism into the planning framework, enabling grounded, decision-aware trajectory correction that is driven by scene complexity. Our method achieves state-of-the-art performance on the NAVSIM benchmark in both PDMS and EPDMS. Extensive experiments demonstrate the effectiveness of our multimodal reflection framework and validate the efficacy of the proposed adaptive reflection strategy.

In this paper, we address the challenging task of multimodal reasoning by incorporating the notion of "slow thinking" into multimodal large language models (MLLMs). Our core idea is that models can learn to adaptively use different levels of reasoning to tackle questions of varying complexity. We propose a novel paradigm of Self-structured Chain of Thought (SCoT), which consists of minimal semantic atomic steps. Unlike existing methods that rely on structured templates or free-form paradigms, our method not only generates flexible CoT structures for various complex tasks but also mitigates the phenomenon of overthinking for easier tasks. To introduce structured reasoning into visual cognition, we design a novel AtomThink framework with four key modules: (i) a data engine to generate high-quality multimodal reasoning paths; (ii) a supervised fine-tuning (SFT) process with serialized inference data; (iii) a policy-guided multi-turn inference method; and (iv) an atomic capability metric to evaluate the single-step utilization rate. Extensive experiments demonstrate that the proposed AtomThink significantly improves the performance of baseline MLLMs, achieving more than 10% average accuracy gains on MathVista and MathVerse. Compared to state-of-the-art structured CoT approaches, our method not only achieves higher accuracy but also improves data utilization by 5 × and boosts inference efficiency by 85.3%.

The integration of visual evidence has significantly enhanced the capabilities of large multimodal models. However, this integration predominantly relies on generating discrete outputs (etc., code or box coordinates) to invoke external tools, a process that introduces rigid dependencies and substantial latency. To overcome these limitations, we propose EVA (LatEnt Visual StAtes), a novel framework that natively generates continuous latent visual representations. These internal representations manifest as an adaptive sequence of Latent_slot tokens, serving as intermediate visual thoughts during the reasoning process. These Latent_slot tokens are then trained end-to-end with the discrete text tokens. This co-optimization, notably, causes extreme policy deviation in the 'transition window' following the Latent_slot tokens. We develop D-GSPO (Decouple-GSPO) to target this root cause by decoupling the optimization of latent and discrete components. To support SFT, we construct EVA-230K, a high-quality text-image interleaved CoT dataset encompassing a diverse range of real-world scenes, documents, charts and OCR tasks. Extensive experiments across multiple benchmarks confirm that EVA achieves significant performance gains while enhancing inference efficiency.

The convergence of embodied intelligence and world models has catalyzed growing interest in integrating physical laws into AI systems. While prior surveys have examined world models and embodied intelligence separately, we focus on the progression that connects these capabilities as a unified developmental pathway from passive observation to active physical comprehension. This survey provides a systematic framework revealing how physical AI advances through four interconnected stages: perception transforms sensory data into structured physical representations, reasoning derives explanations from observed phenomena, modeling enables predictive simulation grounded in physical principles, and embodied interaction closes the loop through physical manipulation and environmental feedback. Each stage enables and enhances the next: perceptual grounding supports causal reasoning, reasoning unlocks predictive capabilities, and robust models drive genuine physical interaction. Through analysis of developments spanning architectural innovations, training methodologies, causal inference, and embodied systems, we synthesize how physical understanding emerges through cumulative integration across this progression. Our framework reveals the evolution from isolated, task-specific solutions toward integrated architectures that advance from pattern recognition toward causal reasoning and counterfactual prediction. This perspective provides foundations for next-generation physical AI systems with direct implications for safe, generalizable, and interpretable deployment across robotics, scientific discovery, and autonomous systems. We maintain a continuously updated taxonomy repository at https://github.com/AI4Phys/Awesome-AI-for-Physics.

Using reinforcement learning to post-train joint video-audio generation models requires a reward signal. Existing methods construct this reward by combining metrics for individual quality dimensions, including audio quality, visual fidelity, and synchronization. However, these metrics evaluate perceptual dimensions separately and fail to capture the overall semantic and temporal coherence among the text prompt, video, and audio that shapes human preferences. Optimizing models against these metrics encourages reward hacking, generating video-audio content that achieves high scores on these metrics yet appears incoherent or unfaithful to human viewers. To address this problem, we first construct a large-scale human-preference dataset VAPref-10K for joint video-audio generation, comprising 9K prompts and 10.3K fine-grained paired comparisons from open-source generation models. We also introduce the VA-Judger-Bench benchmark with both in-domain and out-of-domain model comparisons to evaluate whether reward models truly align with human preferences. We further propose VA-Judger, a chain-of-thought omni-reward model for joint video-audio generation. In particular, VA-Judger first learns from pairs with clear quality gaps to establish structured output and coarse preference discrimination, then distills reliable preference explanations for harder near-quality comparisons via rejection sampling verified against human annotations, and finally performs dimension-wise reinforcement learning that decomposes human feedback into individual quality dimensions for denser reward signals than a single binary preference label. Experiments show that VA-Judger outperforms metric baselines in predicting human preferences on both in-domain and out-of-domain evaluations. Using its human-aligned rewards for post-training audio-video generation model also yields significant improvements in generation quality.

Recent progress in spatial reasoning with Multimodal Large Language Models (MLLMs) increasingly leverages geometric priors from 3D encoders. However, most existing integration strategies remain passive: geometry is exposed as a global stream and fused in an indiscriminate manner, which often induces semantic-geometry misalignment and redundant signals. We propose GeoThinker, a framework that shifts the paradigm from passive fusion to active perception. Instead of feature mixing, GeoThinker enables the model to selectively retrieve geometric evidence conditioned on its internal reasoning demands. GeoThinker achieves this through Spatial-Grounded Fusion applied at carefully selected VLM layers, where semantic visual priors selectively query and integrate task-relevant geometry via frame-strict cross-attention, further calibrated by Importance Gating that biases per-frame attention toward task-relevant structures. Comprehensive evaluation results show that GeoThinker sets a new state-of-the-art in spatial intelligence, achieving a peak score of 72.6 on the VSI-Bench. Furthermore, GeoThinker demonstrates robust generalization and significantly improved spatial perception across complex downstream scenarios, including embodied referring and autonomous driving. Our results indicate that the ability to actively integrate spatial structures is essential for next-generation spatial intelligence.

Emerging World-Action Models (WAMs) have demonstrated promising performance in autonomous driving by jointly modeling future driving scene evolution and trajectory planning. However, existing WAMs are typically trained with video data, which is only 2D projections of the underlying 4D driving scene. Consequently, WAMs fail to understand and capture the structure of 4D scenes and thus generate visually plausible yet 4D inconsistent future predictions that mislead downstream planning. To alleviate this issue, we present 4D-WAM, a model that leverages geometric foundation models for training-time supervision to enable 4D consistent world modeling. Specifically, we feed WAM-predicted future frames into a geometric foundation model, and use 4D-aware responses to define a 4D consistency loss. This loss encourages the model to understand, represent, and predict physically consistent 4D scenes during training, without additional inference cost. Moreover, we identify an early-decision phenomenon in WAMs and propose a decision-oriented timestep sampling strategy that emphasizes supervision at early, high-noise stages, where driving decisions are primarily formed. By propagating 4D supervision to this critical decision-formation phase, the proposed strategy further improves trajectory planning. Extensive experiments demonstrate that 4D-WAM effectively models 4D consistent scene evolution and achieves state-of-the-art performance on challenging NAVSIM-v1 and NAVSIM-v2 benchmarks.

Graphical User Interface (GUI) Agents, benefiting from recent advances in multimodal large language models (MLLM), have achieved significant development. However, due to the frequent updates of GUI applications, adapting to new tasks without forgetting old tasks in GUI continual learning remains an open problem. Existing works are generally trained on a fixed set of tasks and adapt to new tasks either through supervised fine-tuning (SFT) or reinforcement learning (RL), suffering from catastrophic forgetting and slow adaptation. In this work, we propose a \textbf{C}ontinual \textbf{G}UI \textbf{L}earning (CGL) framework that dynamically balances adaptation efficiency and skill retention by enhancing the synergy between SFT and RL. Specifically, we introduce an SFT proportion adjustment mechanism guided by policy entropy to dynamically control the weight allocation between the SFT and RL training phases. Additionally, we propose gradient surgery and entropy-regulated tuning strategies to enable GUI agents to continuously evolve while maintaining competence across previously learned domains. On top of that, we propose a AndroidControl-CL benchmark, which divides GUI applications into distinct task groups to effectively simulate and evaluate the performance of GUI continual learning. Experimental results demonstrate the effectiveness of our proposed CGL framework under the continual learning setting. The benchmark, code and model will be made publicly available.

Pre-trained Multi-modal Large Language Models (MLLMs) provide a knowledge-rich foundation for post-training by leveraging their inherent perception and reasoning capabilities to solve complex tasks. However, the lack of an efficient evaluation framework impedes the diagnosis of their performance bottlenecks. Current evaluation primarily relies on testing after supervised fine-tuning, which introduces laborious additional training and autoregressive decoding costs. Meanwhile, common pre-training metrics cannot quantify a model's perception and reasoning abilities in a disentangled manner. Furthermore, existing evaluation benchmarks are typically limited in scale or misaligned with pre-training objectives. Thus, we propose RADAR, an efficient ability-centric evaluation framework for Revealing Asymmetric Development of Abilities in MLLM pRe-training. RADAR involves two key components: (1) Soft Discrimination Score, a novel metric for robustly tracking ability development without fine-tuning, based on quantifying nuanced gradations of the model preference for the correct answer over distractors; and (2) Multi-Modal Mixture Benchmark, a new 15K+ sample benchmark for comprehensively evaluating pre-trained MLLMs' perception and reasoning abilities in a 0-shot manner, where we unify authoritative benchmark datasets and carefully collect new datasets, extending the evaluation scope and addressing the critical gaps in current benchmarks. With RADAR, we comprehensively reveal the asymmetric development of perceptual and reasoning capabilities in pretrained MLLMs across diverse factors, including data volume, model size, and pretraining strategy. Our RADAR underscores the need for a decomposed perspective on pre-training ability bottlenecks, informing targeted interventions to advance MLLMs efficiently. Our code is publicly available at https://github.com/Nieysh/RADAR.

Unified vision-language models have made significant progress in multimodal understanding and generation, yet they largely fall short in producing multimodal interleaved outputs, which is a crucial capability for tasks like visual storytelling and step-by-step visual reasoning. In this work, we propose a reinforcement learning-based post-training strategy to unlock this capability in existing unified models, without relying on large-scale multimodal interleaved datasets. We begin with a warm-up stage using a hybrid dataset comprising curated interleaved sequences and limited data for multimodal understanding and text-to-image generation, which exposes the model to interleaved generation patterns while preserving its pretrained capabilities. To further refine interleaved generation, we propose a unified policy optimization framework that extends Group Relative Policy Optimization (GRPO) to the multimodal setting. Our approach jointly models text and image generation within a single decoding trajectory and optimizes it with our novel hybrid rewards covering textual relevance, visual-text alignment, and structural fidelity. Additionally, we incorporate process-level rewards to provide step-wise guidance, enhancing training efficiency in complex multimodal tasks. Experiments on MMIE and InterleavedBench demonstrate that our approach significantly enhances the quality and coherence of multimodal interleaved generation.

Condition-based video generation aims to create video content based on given information that describes specific subjects. However, most existing works can only utilize a single condition to guide the denoising process, thereby limiting their applicability to specific scenarios. Although some works attempt to accommodate multiple conditions within one framework, they often require multiple encoders, leading to inefficiencies in integrating multi-condition features. In this work, we present a framework that, with the support of the proposed Unified Adapter (UniAdapter), enables simultaneous multi-condition control of video generation within a single model. To effectively merge these conditions, we propose a novel Probabilistic Multi-condition Concatenator (PMC) module, which employs a unified encoder to accommodate multiple conditions and concatenate condition features at the pixel level to achieve fine-grained control. Following the PMC module, we employ 2D down-sampling blocks to refine features for injection into the Video Diffusion Model (VDM). Moreover, our UniAdapter is designed to be model-agnostic and compatible with any U-Net-based VDM, offering a versatile solution for improving video generation quality. Experimental results on public benchmarks UCF-101 and MSR-VTT show that our method achieves superior results in both quantitative and qualitative evaluations.

Remarkable progress in 2D Vision-Language Models (VLMs) has spurred interest in extending them to 3D settings for tasks like 3D Question Answering, Dense Captioning, and Visual Grounding. Unlike 2D VLMs that typically process images through an image encoder, 3D scenes, with their intricate spatial structures, allow for diverse model architectures. Based on their encoder design, this paper categorizes recent 3D VLMs into 3D object-centric, 2D image-based, and 3D scene-centric approaches. Despite the architectural similarity of 3D scene-centric VLMs to their 2D counterparts, they have exhibited comparatively lower performance compared with the latest 3D object-centric and 2D image-based approaches. To understand this gap, we conduct an in-depth analysis, revealing that 3D scene-centric VLMs show limited reliance on the 3D scene encoder, and the pre-train stage appears less effective than in 2D VLMs. Furthermore, we observe that data scaling benefits are less pronounced on larger datasets. Our investigation suggests that while these models possess cross-modal alignment capabilities, they tend to over-rely on linguistic cues and overfit to frequent answer distributions, thereby diminishing the effective utilization of the 3D encoder. To address these limitations and encourage genuine 3D scene understanding, we introduce a novel 3D Relevance Discrimination QA dataset designed to disrupt shortcut learning and improve 3D understanding. Our findings highlight the need for advanced evaluation and improved strategies for better 3D understanding in 3D VLMs.