arXiv subjects

Xiaofeng Zhang

Publications and source records attributed to Xiaofeng Zhang.

At least 37 records · Page 2Linked to original sources

arXiv2026

SoPE: Spherical Coordinate-Based Positional Embedding for Enhancing Spatial Perception of 3D LVLMs

3D Large Vision-Language Models (3D LVLMs) built upon Large Language Models (LLMs) have achieved remarkable progress across various multimodal tasks. However, their inherited position-dependent modeling mechanism, Rotary Position Embedding (RoPE), remains suboptimal for 3D multimodal understanding. The vanilla RoPE formulation fails to preserve essential three-dimensional spatial structures when encoding 3D tokens, and its relative distance computation overlooks angular dependencies, hindering the model's ability to capture directional variations in visual representations. To overcome these limitations, we introduce Spherical Coordinate-based Positional Embedding (SoPE). Our method maps point-cloud token indices into a 3D spherical coordinate space, enabling unified modeling of spatial locations and directional angles. This formulation preserves the inherent geometric structure of point-cloud data, enhances spatial awareness, and yields more consistent and expressive geometric representations for multimodal learning. In addition, we introduce a multi-scale frequency mixing strategy to fuse feature information across different frequency domains. Experimental results on multiple 3D scene benchmarks validate the effectiveness of our approach, while real-world deployment experiments further demonstrate its strong generalization capability.

cs.CV↗

arXiv2026

C^2ROPE: Causal Continuous Rotary Positional Encoding for 3D Large Multimodal-Models Reasoning

Recent advances in 3D Large Multimodal Models (LMMs) built on Large Language Models (LLMs) have established the alignment of 3D visual features with LLM representations as the dominant paradigm. However, the inherited Rotary Position Embedding (RoPE) introduces limitations for multimodal processing. Specifically, applying 1D temporal positional indices disrupts the continuity of visual features along the column dimension, resulting in spatial locality loss. Moreover, RoPE follows the prior that temporally closer image tokens are more causally related, leading to long-term decay in attention allocation and causing the model to progressively neglect earlier visual tokens as the sequence length increases. To address these issues, we propose C^2RoPE, an improved RoPE that explicitly models local spatial Continuity and spatial Causal relationships for visual processing. C^2RoPE introduces a spatio-temporal continuous positional embedding mechanism for visual tokens. It first integrates 1D temporal positions with Cartesian-based spatial coordinates to construct a triplet hybrid positional index, and then employs a frequency allocation strategy to encode spatio-temporal positional information across the three index components. Additionally, we introduce Chebyshev Causal Masking, which determines causal dependencies by computing the Chebyshev distance of image tokens in 2D space. Evaluation results across various benchmarks, including 3D scene reasoning and 3D visual question answering, demonstrate C^2RoPE's effectiveness. The code is be available at https://github.com/ErikZ719/C2RoPE.

cs.CV↗

arXiv2026

Hallucination Begins Where Saliency Drops

Recent studies have examined attention dynamics in large vision-language models (LVLMs) to detect hallucinations. However, existing approaches remain limited in reliably distinguishing hallucinated from factually grounded outputs, as they rely solely on forward-pass attention patterns and neglect gradient-based signals that reveal how token influence propagates through the network. To bridge this gap, we introduce LVLMs-Saliency, a gradient-aware diagnostic framework that quantifies the visual grounding strength of each output token by fusing attention weights with their input gradients. Our analysis uncovers a decisive pattern: hallucinations frequently arise when preceding output tokens exhibit low saliency toward the prediction of the next token, signaling a breakdown in contextual memory retention. Leveraging this insight, we propose a dual-mechanism inference-time framework to mitigate hallucinations: (1) Saliency-Guided Rejection Sampling (SGRS), which dynamically filters candidate tokens during autoregressive decoding by rejecting those whose saliency falls below a context-adaptive threshold, thereby preventing coherence-breaking tokens from entering the output sequence; and (2) Local Coherence Reinforcement (LocoRE), a lightweight, plug-and-play module that strengthens attention from the current token to its most recent predecessors, actively counteracting the contextual forgetting behavior identified by LVLMs-Saliency. Extensive experiments across multiple LVLMs demonstrate that our method significantly reduces hallucination rates while preserving fluency and task performance, offering a robust and interpretable solution for enhancing model reliability. Code is available at: https://github.com/zhangbaijin/LVLMs-Saliency

cs.CV↗

arXiv2026

Context Tokens are Anchors: Understanding the Repetition Curse in dMLLMs from an Information Flow Perspective

Recent diffusion-based Multimodal Large Language Models (dMLLMs) suffer from high inference latency and therefore rely on caching techniques to accelerate decoding. However, the application of cache mechanisms often introduces undesirable repetitive text generation, a phenomenon we term the \textbf{Repeat Curse}. To better investigate underlying mechanism behind this issue, we analyze repetition generation through the lens of information flow. Our work reveals three key findings: (1) context tokens aggregate semantic information as anchors and guide the final predictions; (2) as information propagates across layers, the entropy of context tokens converges in deeper layers, reflecting the model's growing prediction certainty; (3) Repetition is typically linked to disruptions in the information flow of context tokens and to the inability of their entropy to converge in deeper layers. Based on these insights, we present \textbf{CoTA}, a plug-and-play method for mitigating repetition. CoTA enhances the attention of context tokens to preserve intrinsic information flow patterns, while introducing a penalty term to the confidence score during decoding to avoid outputs driven by uncertain context tokens. With extensive experiments, CoTA demonstrates significant effectiveness in alleviating repetition and achieves consistent performance improvements on general tasks. Code is available at https://github.com/ErikZ719/CoTA

cs.CV↗

arXiv2026

CulturalFrames: Assessing Cultural Expectation Alignment in Text-to-Image Models and Evaluation Metrics

The increasing ubiquity of text-to-image (T2I) models as tools for visual content generation raises concerns about their ability to accurately represent diverse cultural contexts -- where missed cues can stereotype communities and undermine usability. In this work, we present the first study to systematically quantify the alignment of T2I models and evaluation metrics with respect to both explicit (stated) as well as implicit (unstated, implied by the prompt's cultural context) cultural expectations. To this end, we introduce CulturalFrames, a novel benchmark designed for rigorous human evaluation of cultural representation in visual generations. Spanning 10 countries and 5 socio-cultural domains, CulturalFrames comprises 983 prompts, 3637 corresponding images generated by 4 state-of-the-art T2I models, and over 10k detailed human annotations. We find that across models and countries, cultural expectations are missed an average of 44% of the time. Among these failures, explicit expectations are missed at a surprisingly high average rate of 68%, while implicit expectation failures are also significant, averaging 49%. Furthermore, we show that existing T2I evaluation metrics correlate poorly with human judgments of cultural alignment, irrespective of their internal reasoning. Collectively, our findings expose critical gaps, provide a concrete testbed, and outline actionable directions for developing culturally informed T2I models and metrics that improve global usability.

cs.CV↗

arXiv2026

Uncovering Intrinsic Capabilities: A Paradigm for Data Curation in Vision-Language Models

Large vision-language models (VLMs) achieve strong benchmark performance, but controlling their behavior through instruction tuning remains difficult. Reducing the budget of instruction tuning dataset often causes regressions, as heuristic strategies treat models as black boxes and overlook the latent capabilities that govern learning. We introduce Capability-Attributed Data Curation (CADC), a framework that shifts curation from task-specific heuristics to intrinsic capability analysis. CADC discovers intrinsic capabilities in an unsupervised manner from gradient-based learning trajectories, attributes training data to these capabilities via influence estimation, and curates capability-aware curricula through balanced selection and staged sequencing. This transforms black-box instruction tuning into a controllable, capability-driven process. With as little as 5% of the original data, CADC surpasses full-data training on multimodal benchmarks. These results validate intrinsic capabilities as the fundamental building blocks of model learning and establish CADC as a principle paradigm for instruction data curation.

cs.CV↗

arXiv2026

Fine-Grained Preference Optimization Improves Spatial Reasoning in VLMs

Current Vision-Language Models (VLMs) struggle with fine-grained spatial reasoning, particularly when multi-step logic and precise spatial alignment are required. In this work, we introduce SpatialReasoner-R1, a vision-language reasoning model designed to address these limitations. To construct high-quality supervision for spatial reasoning, we design a Multi-Model Monte Carlo Tree Search (M3CTS) method that generates diverse, logically consistent Long Chain-of-Thought (LongCOT) reasoning trajectories. In addition, we propose a fine-grained Direct Preference Optimization (fDPO) method that introduces segment-specific preference granularity for descriptive grounding and logical reasoning, guided by a spatial reward mechanism that evaluates candidate responses based on visual consistency, spatial grounding, and logical coherence. Experimental results demonstrate that fDPO achieves relative performance gains of 4.1% and 9.0% over standard DPO on spatial qualitative and quantitative tasks, respectively. SpatialReasoner-R1, trained with fDPO, sets a new SoTA on SpatialRGPT-Bench, outperforming the strongest baseline by 9.4% in average accuracy, while maintaining competitive performance on general vision-language tasks.

cs.CV↗

arXiv2025

Diversity Recommendation via Causal Deconfounding of Co-purchase Relations and Counterfactual Exposure

Beyond user-item modeling, item-to-item relationships are increasingly used to enhance recommendation. However, common methods largely rely on co-occurrence, making them prone to item popularity bias and user attributes, which degrades embedding quality and performance. Meanwhile, although diversity is acknowledged as a key aspect of recommendation quality, existing research offers limited attention to it, with a notable lack of causal perspectives and theoretical grounding. To address these challenges, we propose Cadence: Diversity Recommendation via Causal Deconfounding of Co-purchase Relations and Counterfactual Exposure - a plug-and-play framework built upon LightGCN as the backbone, primarily designed to enhance recommendation diversity while preserving accuracy. First, we compute the Unbiased Asymmetric Co-purchase Relationship (UACR) between items - excluding item popularity and user attributes - to construct a deconfounded directed item graph, with an aggregation mechanism to refine embeddings. Second, we leverage UACR to identify diverse categories of items that exhibit strong causal relevance to a user's interacted items but have not yet been engaged with. We then simulate their behavior under high-exposure scenarios, thereby significantly enhancing recommendation diversity while preserving relevance. Extensive experiments on real-world datasets demonstrate that our method consistently outperforms state-of-the-art diversity models in both diversity and accuracy, and further validates its effectiveness, transferability, and efficiency over baselines.

cs.IR↗

arXiv2025

D$^{3}$ToM: Decider-Guided Dynamic Token Merging for Accelerating Diffusion MLLMs

Diffusion-based multimodal large language models (Diffusion MLLMs) have recently demonstrated impressive non-autoregressive generative capabilities across vision-and-language tasks. However, Diffusion MLLMs exhibit substantially slower inference than autoregressive models: Each denoising step employs full bidirectional self-attention over the entire sequence, resulting in cubic decoding complexity that becomes computationally impractical with thousands of visual tokens. To address this challenge, we propose D$^{3}$ToM, a Decider-guided dynamic token merging method that dynamically merges redundant visual tokens at different denoising steps to accelerate inference in Diffusion MLLMs. At each denoising step, D$^{3}$ToM uses decider tokens-the tokens generated in the previous denoising step-to build an importance map over all visual tokens. Then it maintains a proportion of the most salient tokens and merges the remainder through similarity-based aggregation. This plug-and-play module integrates into a single transformer layer, physically shortening the visual token sequence for all subsequent layers without altering model parameters. Moreover, D$^{3}$ToM employs a merge ratio that dynamically varies with each denoising step, aligns with the native decoding process of Diffusion MLLMs, achieving superior performance under equivalent computational budgets. Extensive experiments show that D$^{3}$ToM accelerates inference while preserving competitive performance. The code is released at https://github.com/bcmi/D3ToM-Diffusion-MLLM.

cs.CV↗

arXiv2025

Remember Me: Bridging the Long-Range Gap in LVLMs with Three-Step Inference-Only Decay Resilience Strategies

Large Vision-Language Models (LVLMs) have achieved impressive performance across a wide range of multimodal tasks. However, they still face critical challenges in modeling long-range dependencies under the usage of Rotary Positional Encoding (ROPE). Although it can facilitate precise modeling of token positions, it induces progressive attention decay as token distance increases, especially with progressive attention decay over distant token pairs, which severely impairs the model's ability to remember global context. To alleviate this issue, we propose inference-only Three-step Decay Resilience Strategies (T-DRS), comprising (1) Semantic-Driven DRS (SD-DRS), amplifying semantically meaningful but distant signals via content-aware residuals, (2) Distance-aware Control DRS (DC-DRS), which can purify attention by smoothly modulating weights based on positional distances, suppressing noise while preserving locality, and (3) re-Reinforce Distant DRS (reRD-DRS), consolidating the remaining informative remote dependencies to maintain global coherence. Together, the T-DRS recover suppressed long-range token pairs without harming local inductive biases. Extensive experiments on Vision Question Answering (VQA) benchmarks demonstrate that T-DRS can consistently improve performance in a training-free manner. The code can be accessed in https://github.com/labixiaoq-qq/Remember-me

cs.CV↗

arXiv2025

Improving the Physics of Video Generation with VJEPA-2 Reward Signal

This is a short technical report describing the winning entry of the PhysicsIQ Challenge, presented at the Perception Test Workshop at ICCV 2025. State-of-the-art video generative models exhibit severely limited physical understanding, and often produce implausible videos. The Physics IQ benchmark has shown that visual realism does not imply physics understanding. Yet, intuitive physics understanding has shown to emerge from SSL pretraining on natural videos. In this report, we investigate whether we can leverage SSL-based video world models to improve the physics plausibility of video generative models. In particular, we build ontop of the state-of-the-art video generative model MAGI-1 and couple it with the recently introduced Video Joint Embedding Predictive Architecture 2 (VJEPA-2) to guide the generation process. We show that by leveraging VJEPA-2 as reward signal, we can improve the physics plausibility of state-of-the-art video generative models by ~6%.

cs.CV↗

arXiv2025

On Generalization and Distributional Update for Mimicking Observations with Adequate Exploration

Learning from observations (LfO) replicates expert behavior without needing access to the expert's actions, making it more practical than learning from demonstrations (LfD) in many real-world scenarios. However, directly applying the on-policy training scheme in LfO worsens the sample inefficiency problem, while employing the traditional off-policy training scheme in LfO magnifies the instability issue. This paper seeks to develop an efficient and stable solution for the LfO problem. Specifically, we begin by exploring the generalization capabilities of both the reward function and policy in LfO, which provides a theoretical foundation for computation. Building on this, we modify the policy optimization method in generative adversarial imitation from observation (GAIfO) with distributional soft actor-critic (DSAC), and propose the Mimicking Observations through Distributional Update Learning with adequate Exploration (MODULE) algorithm to solve the LfO problem. MODULE incorporates the advantages of (1) high sample efficiency and training robustness enhancement in soft actor-critic (SAC), and (2) training stability in distributional reinforcement learning (RL). Extensive experiments in MuJoCo environments showcase the superior performance of MODULE over current LfO methods.

stat.ML↗

arXiv2025

Increasing the Utility of Synthetic Images through Chamfer Guidance

Conditional image generative models hold considerable promise to produce infinite amounts of synthetic training data. Yet, recent progress in generation quality has come at the expense of generation diversity, limiting the utility of these models as a source of synthetic training data. Although guidance-based approaches have been introduced to improve the utility of generated data by focusing on quality or diversity, the (implicit or explicit) utility functions oftentimes disregard the potential distribution shift between synthetic and real data. In this work, we introduce Chamfer Guidance: a training-free guidance approach which leverages a handful of real exemplar images to characterize the quality and diversity of synthetic data. We show that by leveraging the proposed Chamfer Guidance, we can boost the diversity of the generations w.r.t. a dataset of real images while maintaining or improving the generation quality on ImageNet-1k and standard geo-diversity benchmarks. Our approach achieves state-of-the-art few-shot performance with as little as 2 exemplar real images, obtaining 96.4% in terms of precision, and 86.4% in terms of distributional coverage, which increase to 97.5% and 92.7%, respectively, when using 32 real images. We showcase the benefits of the Chamfer Guidance generation by training downstream image classifiers on synthetic data, achieving accuracy boost of up to 15% for in-distribution over the baselines, and up to 16% in out-of-distribution. Furthermore, our approach does not require using the unconditional model, and thus obtains a 31% reduction in FLOPs w.r.t. classifier-free-guidance-based approaches at sampling time.

cs.CV↗

arXiv2025

SalaMAnder: Shapley-based Mathematical Expression Attribution and Metric for Chain-of-Thought Reasoning

Chain-of-Thought (CoT) prompting enhances the math reasoning capability of large language models (LLMs) to a large margin. However, the mechanism underlying such improvements remains unexplored. In this paper, we present \textbf{SalaMAnder} (\textbf{S}h\textbf{a}p\textbf{l}ey-b\textbf{a}sed \textbf{M}athematical Expression \textbf{A}ttribution a\textbf{nd} M\textbf{e}t\textbf{r}ic), a theoretically grounded methodology as well as a mathematically rigorous evaluation metric for quantifying component-level contributions in few-shot CoT reasoning. Concretely, we leverage the Shapley value for mathematical expression attribution and develop an efficient stratified sampling algorithm that significantly reduces the computational complexity. Besides, we develop the \textbf{CoSP} (\textbf{C}ardinality \textbf{o}f \textbf{S}hapley \textbf{P}ositives) metric through covariance analysis. Comprehensive validation across popular LLM models and diverse mathematical benchmarks demonstrates that the CoSP metric within our SalaMAnder framework exhibits a robust monotonic correlation with model performance, not only providing theoretical explanations for the empirical success of existing few-shot CoT but also establishing mathematically rigorous principles for prompt construction optimization. Furthermore, we verify the reliability of the explanation, based on which we unify the insights of previous work.

cs.AI↗

arXiv2025

CMAMRNet: A Contextual Mask-Aware Network Enhancing Mural Restoration Through Comprehensive Mask Guidance

Murals, as invaluable cultural artifacts, face continuous deterioration from environmental factors and human activities. Digital restoration of murals faces unique challenges due to their complex degradation patterns and the critical need to preserve artistic authenticity. Existing learning-based methods struggle with maintaining consistent mask guidance throughout their networks, leading to insufficient focus on damaged regions and compromised restoration quality. We propose CMAMRNet, a Contextual Mask-Aware Mural Restoration Network that addresses these limitations through comprehensive mask guidance and multi-scale feature extraction. Our framework introduces two key components: (1) the Mask-Aware Up/Down-Sampler (MAUDS), which ensures consistent mask sensitivity across resolution scales through dedicated channel-wise feature selection and mask-guided feature fusion; and (2) the Co-Feature Aggregator (CFA), operating at both the highest and lowest resolutions to extract complementary features for capturing fine textures and global structures in degraded regions. Experimental results on benchmark datasets demonstrate that CMAMRNet outperforms state-of-the-art methods, effectively preserving both structural integrity and artistic details in restored murals. The code is available at~\href{https://github.com/CXH-Research/CMAMRNet}{https://github.com/CXH-Research/CMAMRNet}.

cs.CV↗

arXiv2025

Dual-Phase Playtime-guided Recommendation: Interest Intensity Exploration and Multimodal Random Walks

The explosive growth of the video game industry has created an urgent need for recommendation systems that can scale with expanding catalogs and maintain user engagement. While prior work has explored accuracy and diversity in recommendations, existing models underutilize playtime, a rich behavioral signal unique to gaming platforms, and overlook the potential of multimodal information to enhance diversity. In this paper, we propose DP2Rec, a novel Dual-Phase Playtime-guided Recommendation model designed to jointly optimize accuracy and diversity. First, we introduce a playtime-guided interest intensity exploration module that separates strong and weak preferences via dual-beta modeling, enabling fine-grained user profiling and more accurate recommendations. Second, we present a playtime-guided multimodal random walks module that simulates player exploration using transitions guided by both playtime-derived interest similarity and multimodal semantic similarity. This mechanism preserves core preferences while promoting cross-category discovery through latent semantic associations and adaptive category balancing. Extensive experiments on a real-world game dataset show that DP2Rec outperforms existing methods in both recommendation accuracy and diversity.

cs.IR↗

arXiv2025

MCA-LLaVA: Manhattan Causal Attention for Reducing Hallucination in Large Vision-Language Models

Hallucinations pose a significant challenge in Large Vision Language Models (LVLMs), with misalignment between multimodal features identified as a key contributing factor. This paper reveals the negative impact of the long-term decay in Rotary Position Encoding (RoPE), used for positional modeling in LVLMs, on multimodal alignment. Concretely, under long-term decay, instruction tokens exhibit uneven perception of image tokens located at different positions within the two-dimensional space: prioritizing image tokens from the bottom-right region since in the one-dimensional sequence, these tokens are positionally closer to the instruction tokens. This biased perception leads to insufficient image-instruction interaction and suboptimal multimodal alignment. We refer to this phenomenon as image alignment bias. To enhance instruction's perception of image tokens at different spatial locations, we propose MCA-LLaVA, based on Manhattan distance, which extends the long-term decay to a two-dimensional, multi-directional spatial decay. MCA-LLaVA integrates the one-dimensional sequence order and two-dimensional spatial position of image tokens for positional modeling, mitigating hallucinations by alleviating image alignment bias. Experimental results of MCA-LLaVA across various hallucination and general benchmarks demonstrate its effectiveness and generality. The code can be accessed in https://github.com/ErikZ719/MCA-LLaVA.

cs.CV↗

arXiv2025

Bias Analysis in Unconditional Image Generative Models

The widespread adoption of generative AI models has raised growing concerns about representational harm and potential discriminatory outcomes. Yet, despite growing literature on this topic, the mechanisms by which bias emerges - especially in unconditional generation - remain disentangled. We define the bias of an attribute as the difference between the probability of its presence in the observed distribution and its expected proportion in an ideal reference distribution. In our analysis, we train a set of unconditional image generative models and adopt a commonly used bias evaluation framework to study bias shift between training and generated distributions. Our experiments reveal that the detected attribute shifts are small. We find that the attribute shifts are sensitive to the attribute classifier used to label generated images in the evaluation framework, particularly when its decision boundaries fall in high-density regions. Our empirical analysis indicates that this classifier sensitivity is often observed in attributes values that lie on a spectrum, as opposed to exhibiting a binary nature. This highlights the need for more representative labeling practices, understanding the shortcomings through greater scrutiny of evaluation frameworks, and recognizing the socially complex nature of attributes when evaluating bias.

cs.CV↗