arXiv subjects

Jingkuan Song

Publications and source records attributed to Jingkuan Song.

At least 55 records · Page 3Linked to original sources

arXiv2025

InSpire: Vision-Language-Action Models with Intrinsic Spatial Reasoning

Leveraging pretrained Vision-Language Models (VLMs) to map language instruction and visual observations to raw low-level actions, Vision-Language-Action models (VLAs) hold great promise for achieving general-purpose robotic systems. Despite their advancements, existing VLAs tend to spuriously correlate task-irrelevant visual features with actions, limiting their generalization capacity beyond the training data. To tackle this challenge, we propose Intrinsic Spatial Reasoning (InSpire), a simple yet effective approach that mitigates the adverse effects of spurious correlations by boosting the spatial reasoning ability of VLAs. Specifically, InSpire redirects the VLA's attention to task-relevant factors by prepending the question "In which direction is the [object] relative to the robot?" to the language instruction and aligning the answer "right/left/up/down/front/back/grasped" and predicted actions with ground-truth. Notably, InSpire can be used as a plugin to enhance existing autoregressive VLAs, requiring no extra training data or interaction with other large models. Extensive experimental results in both simulation and real-world environments demonstrate the effectiveness and flexibility of our approach.

cs.RO↗

arXiv2025

An Empirical Analysis of VLM-based OOD Detection: Mechanisms, Advantages, and Sensitivity

Vision-Language Models (VLMs), such as CLIP, have demonstrated remarkable zero-shot out-of-distribution (OOD) detection capabilities, vital for reliable AI systems. Despite this promising capability, a comprehensive understanding of (1) why they work so effectively, (2) what advantages do they have over single-modal methods, and (3) how is their behavioral robustness -- remains notably incomplete within the research community. This paper presents a systematic empirical analysis of VLM-based OOD detection using in-distribution (ID) and OOD prompts. (1) Mechanisms: We systematically characterize and formalize key operational properties within the VLM embedding space that facilitate zero-shot OOD detection. (2) Advantages: We empirically quantify the superiority of these models over established single-modal approaches, attributing this distinct advantage to the VLM's capacity to leverage rich semantic novelty. (3) Sensitivity: We uncovers a significant and previously under-explored asymmetry in their robustness profile: while exhibiting resilience to common image noise, these VLM-based methods are highly sensitive to prompt phrasing. Our findings contribute a more structured understanding of the strengths and critical vulnerabilities inherent in VLM-based OOD detection, offering crucial, empirically-grounded guidance for developing more robust and reliable future designs.

cs.CV↗

arXiv2025

From Channel Bias to Feature Redundancy: Uncovering the "Less is More" Principle in Few-Shot Learning

Deep neural networks often fail to adapt representations to novel tasks under distribution shifts, especially when only a few examples are available. This paper identifies a core obstacle behind this failure: channel bias, where networks develop a rigid emphasis on feature dimensions that were discriminative for the source task, but this emphasis is misaligned and fails to adapt to the distinct needs of a novel task. This bias leads to a striking and detrimental consequence: feature redundancy. We demonstrate that for few-shot tasks, classification accuracy is significantly improved by using as few as 1-5% of the most discriminative feature dimensions, revealing that the vast majority are actively harmful. Our theoretical analysis confirms that this redundancy originates from confounding feature dimensions-those with high intra-class variance but low inter-class separability-which are especially problematic in low-data regimes. This "less is more" phenomenon is a defining characteristic of the few-shot setting, diminishing as more samples become available. To address this, we propose a simple yet effective soft-masking method, Augmented Feature Importance Adjustment (AFIA), which estimates feature importance from augmented data to mitigate the issue. By establishing the cohesive link from channel bias to its consequence of extreme feature redundancy, this work provides a foundational principle for few-shot representation transfer and a practical method for developing more robust few-shot learning algorithms.

cs.CV↗

arXiv2025

Unlocking Smarter Device Control: Foresighted Planning with a World Model-Driven Code Execution Approach

The automatic control of mobile devices is essential for efficiently performing complex tasks that involve multiple sequential steps. However, these tasks pose significant challenges due to the limited environmental information available at each step, primarily through visual observations. As a result, current approaches, which typically rely on reactive policies, focus solely on immediate observations and often lead to suboptimal decision-making. To address this problem, we propose \textbf{Foresighted Planning with World Model-Driven Code Execution (FPWC)},a framework that prioritizes natural language understanding and structured reasoning to enhance the agent's global understanding of the environment by developing a task-oriented, refinable \emph{world model} at the outset of the task. Foresighted actions are subsequently generated through iterative planning within this world model, executed in the form of executable code. Extensive experiments conducted in simulated environments and on real mobile devices demonstrate that our method outperforms previous approaches, particularly achieving a 44.4\% relative improvement in task success rate compared to the state-of-the-art in the simulated environment. Code and demo are provided in the supplementary material.

cs.CV↗

arXiv2025

Dynamic Pattern Alignment Learning for Pretraining Lightweight Human-Centric Vision Models

Human-centric vision models (HVMs) have achieved remarkable generalization due to large-scale pretraining on massive person images. However, their dependence on large neural architectures and the restricted accessibility of pretraining data significantly limits their practicality in real-world applications. To address this limitation, we propose Dynamic Pattern Alignment Learning (DPAL), a novel distillation-based pretraining framework that efficiently trains lightweight HVMs to acquire strong generalization from large HVMs. In particular, human-centric visual perception are highly dependent on three typical visual patterns, including global identity pattern, local shape pattern and multi-person interaction pattern. To achieve generalizable lightweight HVMs, we firstly design a dynamic pattern decoder (D-PaDe), acting as a dynamic Mixture of Expert (MoE) model. It incorporates three specialized experts dedicated to adaptively extract typical visual patterns, conditioned on both input image and pattern queries. And then, we present three levels of alignment objectives, which aims to minimize generalization gap between lightweight HVMs and large HVMs at global image level, local pixel level, and instance relation level. With these two deliberate designs, the DPAL effectively guides lightweight model to learn all typical human visual patterns from large HVMs, which can generalize to various human-centric vision tasks. Extensive experiments conducted on 15 challenging datasets demonstrate the effectiveness of the DPAL. Remarkably, when employing PATH-B as the teacher, DPAL-ViT/Ti (5M parameters) achieves surprising generalizability similar to existing large HVMs such as PATH-B (84M) and Sapiens-L (307M), and outperforms previous distillation-based pretraining methods including Proteus-ViT/Ti (5M) and TinyMiM-ViT/Ti (5M) by a large margin.

cs.CV↗

arXiv2025

Shortcut Learning in Generalist Robot Policies: The Role of Dataset Diversity and Fragmentation

Generalist robot policies trained on large-scale datasets such as Open X-Embodiment (OXE) demonstrate strong performance across a wide range of tasks. However, they often struggle to generalize beyond the distribution of their training data. In this paper, we investigate the underlying cause of this limited generalization capability. We identify shortcut learning -- the reliance on task-irrelevant features -- as a key impediment to generalization. Through comprehensive theoretical and empirical analysis, we uncover two primary contributors to shortcut learning: (1) limited diversity within individual sub-datasets, and (2) significant distributional disparities across sub-datasets, leading to dataset fragmentation. These issues arise from the inherent structure of large-scale datasets like OXE, which are typically composed of multiple sub-datasets collected independently across varied environments and embodiments. Our findings provide critical insights into dataset collection strategies that can reduce shortcut learning and enhance the generalization ability of generalist robot policies. Moreover, in scenarios where acquiring new large-scale data is impractical, we demonstrate that carefully selected robotic data augmentation strategies can effectively reduce shortcut learning in existing offline datasets, thereby improving generalization capabilities of generalist robot policies, e.g., $π_0$, in both simulation and real-world environments. More information at https://lucky-light-sun.github.io/proj/shortcut-learning-in-grps/.

cs.RO↗

arXiv2025

VSAG: An Optimized Search Framework for Graph-based Approximate Nearest Neighbor Search

Approximate nearest neighbor search (ANNS) is a fundamental problem in vector databases and AI infrastructures. Recent graph-based ANNS algorithms have achieved high search accuracy with practical efficiency. Despite the advancements, these algorithms still face performance bottlenecks in production, due to the random memory access patterns of graph-based search and the high computational overheads of vector distance. In addition, the performance of a graph-based ANNS algorithm is highly sensitive to parameters, while selecting the optimal parameters is cost-prohibitive, e.g., manual tuning requires repeatedly re-building the index. This paper introduces VSAG, an open-source framework that aims to enhance the in production performance of graph-based ANNS algorithms. VSAG has been deployed at scale in the services of Ant Group, and it incorporates three key optimizations: (i) efficient memory access: it reduces L3 cache misses with pre-fetching and cache-friendly vector organization; (ii) automated parameter tuning: it automatically selects performance-optimal parameters without requiring index rebuilding; (iii) efficient distance computation: it leverages modern hardware, scalar quantization, and smartly switches to low-precision representation to dramatically reduce the distance computation costs. We evaluate VSAG on real-world datasets. The experimental results show that VSAG achieves the state-of-the-art performance and provides up to 4x speedup over HNSWlib (an industry-standard library) while ensuring the same accuracy.

cs.DB↗

arXiv2025

Towards Generalized and Training-Free Text-Guided Semantic Manipulation

Text-guided semantic manipulation refers to semantically editing an image generated from a source prompt to match a target prompt, enabling the desired semantic changes (e.g., addition, removal, and style transfer) while preserving irrelevant contents. With the powerful generative capabilities of the diffusion model, the task has shown the potential to generate high-fidelity visual content. Nevertheless, existing methods either typically require time-consuming fine-tuning (inefficient), fail to accomplish multiple semantic manipulations (poorly extensible), and/or lack support for different modality tasks (limited generalizability). Upon further investigation, we find that the geometric properties of noises in the diffusion model are strongly correlated with the semantic changes. Motivated by this, we propose a novel $\textit{GTF}$ for text-guided semantic manipulation, which has the following attractive capabilities: 1) $\textbf{Generalized}$: our $\textit{GTF}$ supports multiple semantic manipulations (e.g., addition, removal, and style transfer) and can be seamlessly integrated into all diffusion-based methods (i.e., Plug-and-play) across different modalities (i.e., modality-agnostic); and 2) $\textbf{Training-free}$: $\textit{GTF}$ produces high-fidelity results via simply controlling the geometric relationship between noises without tuning or optimization. Our extensive experiments demonstrate the efficacy of our approach, highlighting its potential to advance the state-of-the-art in semantics manipulation.

cs.CV↗

arXiv2025

A Closer Look at Conditional Prompt Tuning for Vision-Language Models

Despite the great promise of Prompt Tuning (PT) in adapting large Vision-Language Pretrained Models (VLPMs) to downstream tasks, they often struggle to overcome the Base-New Tradeoff (BNT) dilemma: as VLPMs are better tuned to a base task, their ability to generalize to new tasks diminishes. Recent work on conditional PT addresses this problem by replacing static prompts with dynamic Visual Image Information (VII)-conditioned prompts, improving the model's generalization to new tasks to some extent. In this work, we first identify a critical issue with existing conditional PT methods: using VII as the "condition" of prompts yields suboptimal performance, and even random noise-conditioned prompts can outperform the VII-conditioned counterparts. On further analysis, we find that learning dynamic prompts conditioned on Textual Class Information (TCI) is the key to solving the BNT problem. Motivated by this, we then propose Class-adaptive Prompt Tuning (CaPT), which enables fast adaptation of tuned models to new classes by learning TCI-conditioned prompts from base classes. Remarkably, CaPT can be used as a plugin to mitigate the BNT problem for existing unconditional PT schemes. Extensive experiments on 11 datasets show that CaPT consistently improves the performance of five strong unconditional PT baselines with negligible additional computational cost. Additionally, by integrating CaPT with our recently proposed DePT framework, we devise a new conditional PT approach, termed DeCaPT, which outperforms the H ACC of the state-of-the-art conditional PT scheme by 3.49%, averaged over the 11 datasets. Code: https://github.com/Koorye/CaPT.

cs.CV↗

arXiv2025

Scale-Aware Pre-Training for Human-Centric Visual Perception: Enabling Lightweight and Generalizable Models

Human-centric visual perception (HVP) has recently achieved remarkable progress due to advancements in large-scale self-supervised pretraining (SSP). However, existing HVP models face limitations in adapting to real-world applications, which require general visual patterns for downstream tasks while maintaining computationally sustainable costs to ensure compatibility with edge devices. These limitations primarily arise from two issues: 1) the pretraining objectives focus solely on specific visual patterns, limiting the generalizability of the learned patterns for diverse downstream tasks; and 2) HVP models often exhibit excessively large model sizes, making them incompatible with real-world applications.To address these limitations, we introduce Scale-Aware Image Pretraining (SAIP), a novel SSP framework pretraining lightweight vision models to acquire general patterns for HVP. Specifically, SAIP incorporates three learning objectives based on the principle of cross-scale consistency: 1) Cross-scale Matching (CSM) which contrastively learns image-level invariant patterns from multi-scale single-person images; 2) Cross-scale Reconstruction (CSR) which learns pixel-level consistent visual structures from multi-scale masked single-person images; and 3) Cross-scale Search (CSS) which learns to capture diverse patterns from multi-scale multi-person images. Three objectives complement one another, enabling lightweight models to learn multi-scale generalizable patterns essential for HVP downstream tasks.Extensive experiments conducted across 12 HVP datasets demonstrate that SAIP exhibits remarkable generalization capabilities across 9 human-centric vision tasks. Moreover, it achieves significant performance improvements over existing methods, with gains of 3%-13% in single-person discrimination tasks, 1%-11% in dense prediction tasks, and 1%-6% in multi-person visual understanding tasks.

cs.CV↗

arXiv2025

Reliable Few-shot Learning under Dual Noises

Recent advances in model pre-training give rise to task adaptation-based few-shot learning (FSL), where the goal is to adapt a pre-trained task-agnostic model for capturing task-specific knowledge with a few-labeled support samples of the target task.Nevertheless, existing approaches may still fail in the open world due to the inevitable in-distribution (ID) and out-of-distribution (OOD) noise from both support and query samples of the target task. With limited support samples available, i) the adverse effect of the dual noises can be severely amplified during task adaptation, and ii) the adapted model can produce unreliable predictions on query samples in the presence of the dual noises. In this work, we propose DEnoised Task Adaptation (DETA++) for reliable FSL. DETA++ uses a Contrastive Relevance Aggregation (CoRA) module to calculate image and region weights for support samples, based on which a clean prototype loss and a noise entropy maximization loss are proposed to achieve noise-robust task adaptation. Additionally,DETA++ employs a memory bank to store and refine clean regions for each inner-task class, based on which a Local Nearest Centroid Classifier (LocalNCC) is devised to yield noise-robust predictions on query samples. Moreover, DETA++ utilizes an Intra-class Region Swapping (IntraSwap) strategy to rectify ID class prototypes during task adaptation, enhancing the model's robustness to the dual noises. Extensive experiments demonstrate the effectiveness and flexibility of DETA++.

cs.CV↗

arXiv2025

OmniCharacter: Towards Immersive Role-Playing Agents with Seamless Speech-Language Personality Interaction

Role-Playing Agents (RPAs), benefiting from large language models, is an emerging interactive AI system that simulates roles or characters with diverse personalities. However, existing methods primarily focus on mimicking dialogues among roles in textual form, neglecting the role's voice traits (e.g., voice style and emotions) as playing a crucial effect in interaction, which tends to be more immersive experiences in realistic scenarios. Towards this goal, we propose OmniCharacter, a first seamless speech-language personality interaction model to achieve immersive RPAs with low latency. Specifically, OmniCharacter enables agents to consistently exhibit role-specific personality traits and vocal traits throughout the interaction, enabling a mixture of speech and language responses. To align the model with speech-language scenarios, we construct a dataset named OmniCharacter-10K, which involves more distinctive characters (20), richly contextualized multi-round dialogue (10K), and dynamic speech response (135K). Experimental results showcase that our method yields better responses in terms of both content and style compared to existing RPAs and mainstream speech-language models, with a response latency as low as 289ms. Code and dataset are available at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/OmniCharacter.

cs.CL↗

arXiv2025

SeMv-3D: Towards Concurrency of Semantic and Multi-view Consistency in General Text-to-3D Generation

General Text-to-3D (GT23D) generation is crucial for creating diverse 3D content across objects and scenes, yet it faces two key challenges: 1) ensuring semantic consistency between input text and generated 3D models, and 2) maintaining multi-view consistency across different perspectives within 3D. Existing approaches typically address only one of these challenges, often leading to suboptimal results in semantic fidelity and structural coherence. To overcome these limitations, we propose SeMv-3D, a novel framework that jointly enhances semantic alignment and multi-view consistency in GT23D generation. At its core, we introduce Triplane Prior Learning (TPL), which effectively learns triplane priors by capturing spatial correspondences across three orthogonal planes using a dedicated Orthogonal Attention mechanism, thereby ensuring geometric consistency across viewpoints. Additionally, we present Prior-based Semantic Aligning in Triplanes (SAT), which enables consistent any-view synthesis by leveraging attention-based feature alignment to reinforce the correspondence between textual semantics and triplane representations. Extensive experiments demonstrate that our method sets a new state-of-the-art in multi-view consistency, while maintaining competitive performance in semantic consistency compared to methods focused solely on semantic alignment. These results emphasize the remarkable ability of our approach to effectively balance and excel in both dimensions, establishing a new benchmark in the field.

cs.CV↗

arXiv2025

Alleviating Hallucinations in Large Vision-Language Models through Hallucination-Induced Optimization

Although Large Visual Language Models (LVLMs) have demonstrated exceptional abilities in understanding multimodal data, they invariably suffer from hallucinations, leading to a disconnect between the generated text and the corresponding images. Almost all current visual contrastive decoding methods attempt to mitigate these hallucinations by introducing visual uncertainty information that appropriately widens the contrastive logits gap between hallucinatory and targeted ones. However, due to uncontrollable nature of the global visual uncertainty, they struggle to precisely induce the hallucinatory tokens, which severely limits their effectiveness in mitigating hallucinations and may even lead to the generation of undesired hallucinations. To tackle this issue, we conducted the theoretical analysis to promote the effectiveness of contrast decoding. Building on this insight, we introduce a novel optimization strategy named Hallucination-Induced Optimization (HIO). This strategy seeks to amplify the contrast between hallucinatory and targeted tokens relying on a fine-tuned theoretical preference model (i.e., Contrary Bradley-Terry Model), thereby facilitating efficient contrast decoding to alleviate hallucinations in LVLMs. Extensive experimental research demonstrates that our HIO strategy can effectively reduce hallucinations in LVLMs, outperforming state-of-the-art methods across various benchmarks.

cs.CV↗

arXiv2025

Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves

Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks by learning a tiny set of context vectors. Nevertheless, in this work, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves the memory and time efficiency significantly. Upon further investigation, we find that reducing both the length and width of the feature-gradient propagation flows of the full fine-tuning (FT) baseline is key to achieving effective and efficient knowledge transfer. Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks. Unlike existing PT or adapter-based methods, Skip Tuning applies Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip) upon the FT baseline without introducing extra context vectors or adapter modules. Extensive experiments across a wide spectrum of benchmarks demonstrate the superior effectiveness and efficiency of our Skip Tuning over both PT and adapter-based methods. Code: https://github.com/Koorye/SkipTuning.

cs.CV↗

arXiv2025

Attention Hijackers: Detect and Disentangle Attention Hijacking in LVLMs for Hallucination Mitigation

Despite their success, Large Vision-Language Models (LVLMs) remain vulnerable to hallucinations. While existing studies attribute the cause of hallucinations to insufficient visual attention to image tokens, our findings indicate that hallucinations also arise from interference from instruction tokens during decoding. Intuitively, certain instruction tokens continuously distort LVLMs' visual perception during decoding, hijacking their visual attention toward less discriminative visual regions. This distortion prevents them integrating broader contextual information from images, ultimately leading to hallucinations. We term this phenomenon 'Attention Hijacking', where disruptive instruction tokens act as 'Attention Hijackers'. To address this, we propose a novel, training-free strategy namely Attention HIjackers Detection and Disentanglement (AID), designed to isolate the influence of Hijackers, enabling LVLMs to rely on their context-aware intrinsic attention map. Specifically, AID consists of three components: First, Attention Hijackers Detection identifies Attention Hijackers by calculating instruction-driven visual salience. Next, Attention Disentanglement mechanism is proposed to mask the visual attention of these identified Hijackers, and thereby mitigate their disruptive influence on subsequent tokens. Finally, Re-Disentanglement recalculates the balance between instruction-driven and image-driven visual salience to avoid over-masking effects. Extensive experiments demonstrate that AID significantly reduces hallucination across various LVLMs on several benchmarks.

cs.CV↗

arXiv2025

Practical No-box Adversarial Attacks with Training-free Hybrid Image Transformation

In recent years, the adversarial vulnerability of deep neural networks (DNNs) has raised increasing attention. Among all the threat models, no-box attacks are the most practical but extremely challenging since they neither rely on any knowledge of the target model or similar substitute model, nor access the dataset for training a new substitute model. Although a recent method has attempted such an attack in a loose sense, its performance is not good enough and computational overhead of training is expensive. In this paper, we move a step forward and show the existence of a \textbf{training-free} adversarial perturbation under the no-box threat model, which can be successfully used to attack different DNNs in real-time. Motivated by our observation that high-frequency component (HFC) domains in low-level features and plays a crucial role in classification, we attack an image mainly by manipulating its frequency components. Specifically, the perturbation is manipulated by suppression of the original HFC and adding of noisy HFC. We empirically and experimentally analyze the requirements of effective noisy HFC and show that it should be regionally homogeneous, repeating and dense. Extensive experiments on the ImageNet dataset demonstrate the effectiveness of our proposed no-box method. It attacks ten well-known models with a success rate of \textbf{98.13\%} on average, which outperforms state-of-the-art no-box attacks by \textbf{29.39\%}. Furthermore, our method is even competitive to mainstream transfer-based black-box attacks.

cs.CV↗

arXiv2024

MMEvol: Empowering Multimodal Large Language Models with Evol-Instruct

The development of Multimodal Large Language Models (MLLMs) has seen significant advancements with increasing demands in various fields (e.g., multimodal agents, embodied intelligence). While model-driven approaches attempt to enhance MLLMs capabilities through diverse architectures, the gains have become increasingly marginal. Conversely, data-driven methods, which scale up image-text instruction data, are more effective but face limited data diversity and complexity challenges. The absence of high-quality data constitutes a significant development barrier for MLLMs. To address the data quality bottleneck, we propose MMEvol, a novel multimodal instruction data evolution framework. This framework iteratively improve data quality through a refined combination of fine-grained perception, cognitive reasoning, and interaction evolution, generating a more complex and diverse image-text instruction dataset that empowers MLLMs with enhanced capabilities. Beginning with an initial set of instructions, SEED-163K, we utilize MMEvol to systematically broaden the diversity of instruction types, extend visual reasoning steps to improve cognitive reasoning abilities, and thoroughly explore fine-grained information within images to enhance visual understanding and robustness. To comprehensively evaluate the effectiveness of our approach, we conduct extensive qualitative analysis and quantitative experiments across 13 vision-language tasks. Compared to baseline models trained with the initial seed data, the results demonstrate that our method achieves an average accuracy improvement of 3.1 percentage points. Furthermore, our approach reaches state-of-the-art (SOTA) performance in nine tasks using significantly less data compared to state-of-the-art models.

cs.CL↗