Search arXiv⌕ Search

arXiv subjects

Keze Wang

Publications and source records attributed to Keze Wang.

At least 19 recordsLinked to original sources
arXiv2026

Token-Budget Distillation: Transferring Full-Token Semantics to Compressed Video Vision-Language Models

Adapting video vision-language models (VLMs) is computationally expensive because video inputs produce a large number of visual tokens, making both fine-tuning and inference costly. Although visual token compression can reduce this overhead, direct adaptation on compressed inputs often causes semantic drift and noticeable performance degradation. We present Token-Budget Distillation (TBD), a parameter-efficient fine-tuning framework for adapting video VLMs under a fixed token budget. TBD freezes the pretrained backbone, updates only LoRA adapters, and integrates FlashVID-based visual token compression into the video pathway. To preserve full-token semantics under compression, TBD employs a dual-path teacher-student design, where a full-token teacher provides stable supervision and a compressed student is optimized with task loss, answer-region KL distillation, GT-anchored margin distillation, and reliability-aware KD control. This design enables the student to recover the semantic behavior of the full-token model while remaining efficient under aggressive token reduction. We evaluate TBD on three video VLM backbones, including LLaVA-Video, LLaVA-OneVision, and Qwen3-VL-8B-Instruct, across four video understanding benchmarks. TBD consistently outperforms compression-only baselines under both moderate and aggressive compression. On LLaVA-Video at retention ratio R = 10 percent, TBD preserves 97.0 percent of the Vanilla model's average accuracy; on LLaVA-OneVision at R = 10 percent, it achieves an average score of 58.4 and matches 100.0 percent relative accuracy.

cs.CV↗
arXiv2026

CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration

Training-free feature reuse accelerates diffusion transformer inference by replacing full denoising computation at selected steps with cached approximations. Existing methods mainly improve the reconstruction of skipped features, but we find that once a stable first-order reconstructor is used, increasing reconstruction complexity brings limited gains, whereas the placement of exact-computation steps has a much larger effect on fidelity. More importantly, absolute shallow-probe variation is jointly affected by input-trajectory displacement and network response strength, and therefore does not directly reflect cache risk. We address this confounding by normalizing probe-output variation with the corresponding input displacement, yielding a directional gain per unit trajectory movement. Based on this signal, we propose CacheScope, which identifies high-risk reuse steps online and selectively falls back to full denoising computation while leaving the underlying cache reconstructor unchanged. CacheScope requires no additional training, gradients, auxiliary networks, or prompt-specific offline schedule optimization. Across FLUX.1-dev, Wan2.1-T2V-1.3B, and Open-Sora 1.2, CacheScope achieves up to 3.57x acceleration and reduces median LPIPS by up to 70.1% over the strongest probe-based baseline.

cs.LG↗
arXiv2026

Recurrent Latent Visual Search for GUI Grounding

GUI grounding is a critical capability for GUI agents powered by vision-language models, helping them execute user instructions by locating the corresponding elements in screenshots. Single-step grounding struggles with small elements and dense layouts, motivating multi-step visual search. However, existing approaches commonly rely on textual reasoning misaligned with visual space or costly multi-round interactions with external visual tools. To make multi-step visual search an explicit spatial process within the model, we propose ReLaViS, which performs Recurrent Latent Visual Search in a single interaction round. At each step, a spatial search head uses the hidden state to query the screenshot's visual tokens, producing a spatial search distribution that explicitly represents the search focus. This distribution then aggregates the visual tokens into latent visual evidence, which is recurrently fed back as the next input embedding to condition subsequent search. We further introduce a GUI-aware coarse-to-fine inductive bias through trajectories constructed from flat element annotations, supervising search from the global interface through intermediate element groups to the target. Built on Qwen2.5-VL-7B, ReLaViS improves ScreenSpot-Pro accuracy by 3.1 percentage points to 56.3% with only a 3.5% increase in inference FLOPs and outperforms the matched single-step baseline on all five benchmarks.

cs.CV↗
arXiv2026

How Do Document Parsers Break? Auditing Structural Vulnerability in Document Intelligence

Document Layout Analysis (DLA) pipelines provide structured page representations for retrieval-augmented generation, long-document question answering, and related applications. Yet their robustness evaluation remains largely area-centric. We identify this Footprint Bias and propose ProSA, a lightweight output-level auditing framework that decouples controlled probing, policy-driven targeting, and structure-aware diagnosis. ProSA combines Block-level Structural Loss Rate (B-SLR), granularity-aware exposure descriptors, and pathway attribution to analyze where structural identity is lost, at what exposure granularity failures emerge, and how failures propagate. Across MinerU and PP-StructureV3 on 1,000 pages, affected area weakly tracks perturbation-induced OCR instability ($R^2=0.384/0.110$), whereas B-SLR aligns much more closely with it ($R^2=0.727/0.916$). Exposure descriptors further separate occlusion- and topology-dominant pathways, while matched-footprint structural probes cause much larger downstream QA/retrieval drops than area-matched erasure. These results shift DLA robustness evaluation from footprint-based measurement toward structure-aware vulnerability auditing.

cs.CL↗
arXiv2026

Do Dynamic Routers Need Memory? HeRo: History-Aware Routing for Efficient LLM Inference

Dynamic layer routing reduces the inference cost of Large Language Models (LLMs) by learning to skip layers for individual tokens. Existing methods, however, treat each routing decision as a local operation conditioned solely on the current hidden state which is a formulation that overlooks the sequential, path-dependent nature of routing across depth: earlier decisions shape the representations seen by downstream routers, and the layer-usage objective couples all decisions jointly. We propose History-Aware Routing (HeRo), a dynamic routing framework that resolves this mismatch by introducing a router memory mechanism to maintain an explicit routing state across model depth. The memory is constructed via linear attention, incrementally aggregating preceding routing scores and their induced residual updates into a compact history representation. At each routed layer, the router conditions jointly on this accumulated state and the current hidden representation to select the executed branch. Instantiated for token-wise FFN routing, HeRo trains only lightweight routers and adapters on a frozen backbone, requiring no modification to pretrained parameters. Across Llama 3.1-8B, Llama 2-7B, and Llama 2-13B, HeRo consistently achieves the highest aggregate performance retention among ten baselines. On Llama 3.1-8B, it bypasses 26.87% of model parameters while achieving 100.24% of dense model performance across seven benchmarks, and retains 97.01% while bypassing 38.82% of model parameters under a tighter computation budget. Ablation studies confirm that removing routing history consistently degrades performance, most notably on multistep reasoning and code generation, validating that explicit routing memory enables more accurate and adaptive dynamic routing than solely conditioning on hidden state.

cs.AI↗
arXiv2026

GAM-Agent: Game-Theoretic and Uncertainty-Aware Collaboration for Complex Visual Reasoning

We propose GAM-Agent, a game-theoretic multi-agent framework for enhancing vision-language reasoning. Unlike prior single-agent or monolithic models, GAM-Agent formulates the reasoning process as a non-zero-sum game between base agents--each specializing in visual perception subtasks--and a critical agent that verifies logic consistency and factual correctness. Agents communicate via structured claims, evidence, and uncertainty estimates. The framework introduces an uncertainty-aware controller to dynamically adjust agent collaboration, triggering multi-round debates when disagreement or ambiguity is detected. This process yields more robust and interpretable predictions. Experiments on four challenging benchmarks--MMMU, MMBench, MVBench, and V*Bench--demonstrate that GAM-Agent significantly improves performance across various VLM backbones. Notably, GAM-Agent boosts the accuracy of small-to-mid scale models (e.g., Qwen2.5-VL-7B, InternVL3-14B) by 5--6\%, and still enhances strong models like GPT-4o by up to 2--3\%. Our approach is modular, scalable, and generalizable, offering a path toward reliable and explainable multi-agent multimodal reasoning.

cs.AI↗
arXiv2026

Hybrid Token Compression for Vision-Language Models

Vision-language models (VLMs) rely on hundreds of visual tokens, leading to high computational and memory costs. Existing compression methods face a trade-off: continuous compression can weaken high-level semantics, while discrete quantization may lose fine-grained appearance details. We introduce HTC-VLM, a hybrid visual token compression framework that disentangles semantics and appearance through two complementary pathways. A continuous pathway preserves detailed ViT patch features, while a discrete pathway provides semantic anchors using MGVQ quantization represented by four tokens. The two pathways are fused into a 580-token hybrid sequence and compressed into a single token using a disentanglement attention mask and a bottleneck. Under the same one-token output budget, HTC-VLM achieves 87.2% average performance retention across seven benchmarks (GQA, VQAv2, MMBench, MME, POPE, SEED-Bench, and ScienceQA-Image), outperforming the leading continuous baseline at 81.0%. Attention analysis shows that the compressed token prioritizes discrete anchors, supporting their role as semantic guidance. We further study token-budget scaling, cross-architecture generalization, inference efficiency, robustness to codebook and masking variations, and the information-theoretic properties of the hybrid bottleneck. These results show that combining continuous appearance features with discrete semantic anchors enables effective extreme visual token compression for efficient VLMs.

cs.CV↗
arXiv2026

UserToolBench: A User-Profile-Hidden Benchmark for Personalized Decision Making in Tool-Use LLMs

Tool-use LLMs are increasingly asked to act on users' behalf, but existing benchmarks usually focus on profile recall, style imitation, generic tool use, or response-level personalization. We introduce UserToolBench , a benchmark for personalized decision making in tool-use LLMs. UserToolBench tests whether a model can infer latent user preferences from interaction history, recognize when clarification is needed, and produce user-aligned tool-call trajectories under incomplete information. The benchmark is built from privacy-sanitized real interaction traces and combines structured persona profiles, public API-style tool ecosystems, and long-horizon multi-turn trajectories. It includes 10 user profiles, 36 tool sets, 1,065 turns, 170 unique tools, and evaluation-focused task types covering lack-of-information, single-tool, and multi-tool settings. Experiments with strong tool-use LLMs show that current models still have difficulty with personalized delegation. Multi-tool coordination, missing-constraint inference, and long-horizon behavioral consistency remain major bottlenecks. These results suggest that personalization evaluation should move beyond asking whether outputs sound user-specific and instead ask whether LLMs make correct decisions for the users they represent.

cs.LG↗
arXiv2026

Agentic Designer: Progressive Multi-Agent Collaboration for Structure-Aware Interior Layout Generation

Generating realistic interior furniture layouts that strictly adhere to architectural constraints (e.g., walls, doors, and windows) remains a fundamental challenge in automated spatial design. Existing approaches, primarily based on one-shot generation using diffusion models or Large Language Models (LLMs), lack explicit mechanisms for intermediate geometric constraint verification, often resulting in structural collisions and functionally infeasible arrangements under complex room constraints. To address these challenges, we propose Agentic Designer, a progressive, multi-agent framework that formulates structure-aware interior layout generation as an iterative and constraint-verified decision process. By decomposing layout synthesis into modular stages of proposal, verification, and adjustment, the framework coordinates three specialized agents, a Generator, an Evaluator, and a Refiner, through a Progressive Consensus Mechanism. This mechanism enforces stepwise geometric validation and correction before each placement is committed, thereby preventing error accumulation. To facilitate this structure-aware paradigm and standardize evaluation, we establish InStruct, a comprehensive benchmark that integrates a dataset comprising over 18,000 high-quality, parametrically annotated samples with a novel suite of structure-centric metrics. Extensive quantitative evaluations, qualitative analyses, and user studies show that Agentic Designer significantly outperforms state-of-the-art methods, demonstrating substantial improvements in strict structural adherence and functional design coherence.

cs.CV↗
arXiv2026

Category-Adaptive Cross-Modal Semantic Refinement and Transfer for Open-Vocabulary Multi-Label Recognition

Benefiting from the generalization capability of CLIP, recent vision language pre-training (VLP) models have demonstrated the ability to capture a wide range of visual concepts in daily images. However, due to the presence of unseen categories in open-vocabulary settings, existing algorithms struggle to capture semantic correlations between categories, leading to suboptimal performance on open-vocabulary multi-label recognition (OV-MLR). Furthermore, the substantial variation in the number of discriminative areas across diverse object categories is misaligned with the fixed-number patch matching used in current methods, introducing noisy visual cues that hinder the capture of target semantics. To address these challenges, we propose a novel category-adaptive cross-modal semantic refinement and transfer (C$^2$SRT) framework to model semantic correlations both within each category and across different categories, in a category-adaptive manner. The proposed framework consists of two complementary modules, i.e., intra-category semantic refinement (ISR) module and inter-category semantic transfer (IST) module. Specifically, the ISR module leverages the cross-modal knowledge of the VLP model to adaptively select a set of local discriminative regions that represent the semantics of the target category. The IST module adaptively discovers a set of correlated categories for a target category by constructing a category-adaptive correlation graph and transfers semantic knowledge from the correlated seen categories to unseen ones. Experiments on OV-MLR benchmarks demonstrate that the proposed C$^2$SRT framework improves over current methods.

cs.CV↗
arXiv2026

ROSE: Benchmarking the Perception-to-Action Gap in Multimodal Models

Multimodal large language models (MLLMs) are increasingly expected to act on visual information, yet the same scene may require different actions under different task contexts. How reliably can a model turn the same visual evidence into the action required by the current context? To answer this question, we introduce \textsc{ROSE} (\textbf{R}eference-conditioned \textbf{O}ddity and \textbf{S}ymbolic \textbf{E}xecution), a controlled benchmark that holds the visual scene fixed while varying region constraints and required symbolic outputs. Through coupled counting and coordinate-action tasks, \textsc{ROSE} tests whether models can infer an implicit majority reference and act on the resulting fine-grained visual evidence under changing contexts. Across nine recent MLLMs, performance drops by as much as 44.5 percentage points from counting-oriented tasks to region-conditioned action, despite 98.8\% human performance. The gap persists on paired scenes and regions for which the same model returns the correct count, while global-click and matched local controls show that coordinate grounding explains only part of the loss, revealing a distinct, model-dependent bottleneck in turning shared visual evidence into context-specific actions.

cs.CV↗
arXiv2026

Kolmogorov-Arnold Fourier Networks

Although Kolmogorov-Arnold-based interpretable networks (KANs) possess strong theoretical expressiveness, they suffer from severe parameter explosion and limited ability to capture high-frequency features in high-dimensional tasks. To address these issues, we propose the Kolmogorov-Arnold Fourier Network (KAF), which fundamentally redefines the KAN paradigm through spectral reparameterization. Our key contributions include: (1) proposing a fundamental basis transformation from the local, grid-based B-spline representation to a global, adaptive spectral representation. This shift changes the network's inductive bias, reducing parameter complexity from $O(G)$ to $O(1)$ while preserving expressiveness; (2) introducing trainable Random Fourier Features (RFF) initialized via a spectral alignment strategy, which allows the model to break the smoothness limitation of fixed kernels and accurately capture high-frequency components; and (3) implementing an adaptive hybrid GELU-Fourier activation mechanism that progressively enhances frequency representation during training. Comprehensive experiments demonstrate the superiority of KAF across computer vision (CV), natural language processing (NLP), audio, and partial differential equation (PDE) solving tasks, achieving state-of-the-art performance with improved efficiency. The code is available at https://github.com/kolmogorovArnoldFourierNetwork/KAF.

cs.LG↗
arXiv2026

PhyDetEx: Detecting and Explaining the Physical Plausibility of T2V Models

Driven by the growing capacity and training scale, Text-to-Video (T2V) generation models have recently achieved substantial progress in video quality, length, and instruction-following capability. However, whether these models can understand physics and generate physically plausible videos remains a question. While Vision-Language Models (VLMs) have been widely used as general-purpose evaluators in various applications, they struggle to identify the physically impossible content from generated videos. To investigate this issue, we construct a \textbf{PID} (\textbf{P}hysical \textbf{I}mplausibility \textbf{D}etection) dataset, which consists of a \textit{test split} of 500 manually annotated videos and a \textit{train split} of 2,588 paired videos, where each implausible video is generated by carefully rewriting the caption of its corresponding real-world video to induce T2V models producing physically implausible content. With the constructed dataset, we introduce a lightweight fine-tuning approach, enabling VLMs to not only detect physically implausible events but also generate textual explanations on the violated physical principles. Taking the fine-tuned VLM as a physical plausibility detector and explainer, namely \textbf{PhyDetEx}, we benchmark a series of state-of-the-art T2V models to assess their adherence to physical laws. Our findings show that although recent T2V models have made notable progress toward generating physically plausible content, understanding and adhering to physical laws remains a challenging issue, especially for open-source models. Our dataset, training code, and checkpoints are available at \href{https://github.com/Zeqing-Wang/PhyDetEx}{https://github.com/Zeqing-Wang/PhyDetEx}.

cs.CV↗
arXiv2026

LASAR: Towards Spatio-temporal Reasoning with Latent Cognitive Map

A fundamental challenge in embodied AI is verifying if agents build internal models of spatial structure or merely learn to mimic task-specific expert trajectories. This is critical as foundational approaches rooted in action-centric tasks (e.g., VLN) and reasoning-centric tasks (e.g., EQA) often share a common limitation: they lack a learning signal that forces them to encode fine-grained spatial relationships (like topology or distance) over long-range, fragmented experiences. To address this, we first propose LASAR, an architecture featuring a dual-memory system designed to maintain both episodic experiences and a semantic cognitive map. We then introduce Spatio-temporal Contextual Representation Learning (ST-CRL), a contrastive objective designed to train this architecture. ST-CRL leverages spatio-temporal cues from cognitive queries generated through annotated spatio-temporal context in simulation to build sample pairs, thereby forming the internal cognitive map from the agent's experiences. Experiments demonstrate that our method achieves 2\%-3.5\% gains in both zero-shot generalization on standard VLN-CE and VSI-Bench benchmarks. We also demonstrate that our proposed cognitive map has high self-consistency.

cs.CV↗
arXiv2026

VideoVerse: Does Your T2V Generator Have World Model Capability to Synthesize Videos?

The recent rapid advancement of Text-to-Video (T2V) generation technologies are engaging the trained models with more world model ability, making the existing benchmarks increasingly insufficient to evaluate state-of-the-art T2V models. First, current evaluation dimensions, such as per-frame aesthetic quality and temporal consistency, are no longer able to differentiate state-of-the-art T2V models. Second, event-level temporal causality-an essential property that differentiates videos from other modalities-remains largely unexplored. Third, existing benchmarks lack a systematic assessment of world knowledge, which are essential capabilities for building world models. To address these issues, we introduce VideoVerse, a comprehensive benchmark focusing on evaluating whether the current T2V model could understand complex temporal causality and world knowledge to synthesize videos. We collect representative videos across diverse domains and extract their event-level descriptions with inherent temporal causality, which are then rewritten into text-to-video prompts by independent annotators. For each prompt, we design ten evaluation dimensions covering dynamic and static properties, resulting in 300 prompts, 815 events, and 793 evaluation questions. Consequently, a human preference-aligned QA-based evaluation pipeline is developed by using modern vision-language models to systematically benchmark leading open- and closed-source T2V systems, revealing the current gap between T2V models and desired world modeling abilities.

cs.CV↗
arXiv2026

Cost-Effective Communication: An Auction-based Method for Language Agent Interaction

Multi-agent systems (MAS) built on large language models (LLMs) often suffer from inefficient "free-for-all" communication, leading to exponential token costs and low signal-to-noise ratios that hinder their practical deployment. We challenge the notion that more communication is always beneficial, hypothesizing instead that the core issue is the absence of resource rationality. We argue that "free" communication, by ignoring the principle of scarcity, inherently breeds inefficiency and unnecessary expenses. To address this, we introduce the Dynamic Auction-based Language Agent (DALA), a novel framework that treats communication bandwidth as a scarce and tradable resource. Specifically, our DALA regards inter-agent communication as a centralized auction, where agents learn to bid for the opportunity to speak based on the predicted value density of their messages. Thus, our DALA intrinsically encourages agents to produce concise, informative messages while filtering out low-value communication. Extensive and comprehensive experiments demonstrate that our economically-driven DALA achieves new state-of-the-art performance across seven challenging reasoning benchmarks, including 84.32% on MMLU and a 91.21% pass@1 rate on HumanEval. Note that this is accomplished with remarkable efficiency, i.e., our DALA uses only 6.25 million tokens, a fraction of the resources consumed by current state-of-the-art methods on GSM8K. Further analysis reveals that our DALA cultivates the emergent skill of strategic silence, effectively adapting its communication strategies from verbosity to silence in a dynamical manner via resource constraints. Our code and updates are available at https://github.com/waltstephen/Cost-Effective-Communication.

cs.AI↗
arXiv2026

3DAlign-DAER: Dynamic Attention Policy and Efficient Retrieval Strategy for Fine-grained 3D-Text Alignment at Scale

Despite recent advancements in 3D-text cross-modal alignment, existing state-of-the-art methods still struggle to align fine-grained textual semantics with detailed geometric structures, and their alignment performance degrades significantly when scaling to large-scale 3D databases. To overcome this limitation, we introduce 3DAlign-DAER, a unified framework designed to align text and 3D geometry via the proposed dynamic attention policy and the efficient retrieval strategy, capturing subtle correspondences for diverse cross-modal retrieval and classification tasks. Specifically, during the training, our proposed dynamic attention policy (DAP) employs the Hierarchical Attention Fusion (HAF) module to represent the alignment as learnable fine-grained token-to-point attentions. To optimize these attentions across different tasks and geometric hierarchies, our DAP further exploits the Monte Carlo tree search to dynamically calibrate HAF attention weights via a hybrid reward signal and further enhances the alignment between textual descriptions and local 3D geometry. During the inference, our 3DAlign-DAER introduces an Efficient Retrieval Strategy (ERS) to leverage efficient hierarchical searching in the large-scale embedding spaces, outperforming traditional methods (e.g., KNN) in accuracy and efficiency. Furthermore, to facilitate text-3D alignment research and train our 3DAlign-DAER, we construct Align3D-2M, a large-scale dataset featuring 2M text-3D pairs, to provide sufficient fine-grained cross-modal annotations. Extensive and comprehensive experiments demonstrate the superior performance of our 3DAlign-DAER on diverse benchmarks. We will release our codes, models, and datasets. Our code and updates are available at https://github.com/waltstephen/Cost-Effective-Communication.

cs.CV↗
arXiv2026

Stable Language Guidance for Vision-Language-Action Models

Vision-Language-Action (VLA) models have demonstrated impressive capabilities in generalized robotic control; however, they remain notoriously brittle to linguistic perturbations. We identify a critical ``modality collapse'' phenomenon where strong visual priors overwhelm sparse linguistic signals, causing agents to overfit to specific instruction phrasings while ignoring the underlying semantic intent. To address this, we propose Residual Semantic Steering (RSS), a probabilistic framework that disentangles physical affordance from semantic execution. RSS introduces two theoretical innovations: (1) Monte Carlo Syntactic Integration, which approximates the true semantic posterior via dense, LLM-driven distributional expansion, and (2) Residual Affordance Steering, a dual-stream decoding mechanism that explicitly isolates the causal influence of language by subtracting the visual affordance prior. Theoretical analysis suggests that RSS effectively maximizes the mutual information between action and intent while suppressing visual distractors. Empirical results across diverse manipulation benchmarks demonstrate that RSS achieves state-of-the-art robustness, maintaining performance even under adversarial linguistic perturbations. We release our code at https://github.com/Doo-mon/RSS.

cs.RO↗