Large Vision Language Models

Multimodal LLMs are evolving from vision-language to tri-modality that see, hear, and read, yet pipelines and benchmarks remain English-centric and compute-heavy. The tutorial offers an overview of this emerging research area for multilingual multimodality across text, speech, and vision under limited data/compute budgets, synthesizing foundations, recent multilingual models (PALO, Maya), speech-text LLMs. We cover low-cost data creation/curation; adapter stacks for tri-modal alignment; culture-aware evaluation beyond English and hands on resources for fine-tuning a compact multilingual VLM and wiring a speech->text->LLM pipeline. The content will be delivered as an interactive half-day tutorial, designed for researchers and practitioners working on multilingual, multimodal AI in low-resource language settings.

Vision-Language-Action (VLA) models leverage powerful perceptual priors from web-scale Vision-Language Model (VLM) pre-training, yet they remain surprisingly brittle in practice, frequently failing at simple robotic tasks. To mitigate this, we propose Contrastive Conceptor Activation Steering (COAST). COAST builds on the notion of a "conceptor", a linear operator that soft-projects data into the principal components of a target distribution. COAST uses conceptors to identify success-critical subspaces for a target robotic task from a few examples of success and failure rollouts. At inference time, it steers VLA latents into these identified success subspaces to improve task outcomes. Across three architecturally distinct neural policies (flow-matching VLA, autoregressive VLA, and Diffusion Policy), COAST improves absolute mean simulation and real-robot task success rate by over 20 and 40% respectively. The activation subspace geometry reveals that failure modes share substantial structure across tasks while success representations remain largely task-specific. When tasks share similar failure modes, this structure enables previously fitted conceptors to improve performance on new tasks without refitting. Ultimately, our results suggest that current VLAs retain substantial task-relevant knowledge in their latent representations, and that the action expert's decoding bottleneck could be mitigated by steering its residual stream toward task-relevant subspaces. COAST provides a lightweight, training-free path to unlocking these latent capabilities by steering the model towards its own "success" distributions.

Brain tumor diagnosis is largely dependent on Magnetic Resonance Imaging (MRI) evaluation, which requires radiologists to synthesize thousands of images across multiple 3D sequences and longitudinal studies. This process requires advanced neuro-radiology training, poses substantial cognitive load, and is highly time-consuming. Despite increasing demands in radiology, this expertise is difficult to scale, straining the current health systems. Vision-Language Models (VLMs) provide an opportunity to reduce this burden through a semi-automated, interactive interpretation of complex brain MRIs. However, they are currently underutilized in neuro-oncology due to a lack of specialized benchmarks for evaluating them. We introduce a clinically relevant visual question answering (VQA) benchmark -- the UCSF-PDGM-VQA dataset -- consisting of 2,387 QA pairs from 473 glioma-related MRI studies in the public UCSF-PDGM dataset. We further establish a performance baseline for six state-of-the-art vision-language models (VLMs) and one large language model on this dataset. We find that current models are incapable of effectively processing multi-sequence, 3-dimensional MRI scans, thus resulting in a suppression of visual features and over-reliance on language priors, causing modality collapse. These findings underscore a critical deficiency in current model reliability and safety within clinical settings, necessitating the development of robust, domain-specific VLMs.

Distilling vision-language models into faster hybrid architectures, such as 3:1 Mamba-2/attention mixes, is now standard practice for making inference efficient. Aggregate benchmarks suggest that this works but they hide selective failures. When we distill Qwen3-VL-8B-Instruct into a 3:1 Mamba-2/attention hybrid, student model stays within 2 points of the teacher across visual reasoning benchmarks like MMStar, MMBench, and MMMU-Pro, while dropping 13 points on optical-character-recognition and document tasks. The student can still understand the scene but loses the fine-grained text needed to answer. We localize much of the failure to a specific kind of position. In a high-resolution image, most patches are sky, wall, or smooth texture, while a small fraction carries text, edges, object boundaries, or other local details. In a token-level diagnostic, the top 10% highest-density patches have 3.6$\times$ larger residual drift than the bottom 10% lowest-density patches and 3.5$\times$ larger teacher-masking answer contribution. Uniform weighting devotes many loss terms to low-information background patches, whereas sparse answer-bearing patches receive no special protection. The required intervention is minimal: we replace uniform residual alignment with density-weighted residual alignment, using patch self-dissimilarity as a training-free proxy for position importance. We call this HEED. Compared with normal end-to-end distillation, HEED increases performance by 8.7 points on OCRBench v2 and 5.13 points on a 10-benchmark average. The gain is realized on different teacher models and hybrid architectures. After standard post-training, the student reaches teacher-level performance on the 10-benchmark average with a 4.12$\times$ throughput and a 68% memory saving at 128k context, with no additional parameters and no inference-time cost.

While large vision-language models (VLMs) are increasingly adopted as the perceptual backbone for embodied agents, existing benchmarks often rely on question-answering or multiple-choice formats. These protocols allow models to exploit linguistic priors rather than demonstrating genuine visual grounding. To address this, we present EPIC-Bench, Embodied PerceptIon BenChmark, a fine-grained grounding benchmark designed to systematically evaluate the visual perceptual capabilities of VLMs in real-world embodied environments. Comprising 6.6k meticulously annotated tuples (Image, Text, Mask), EPIC-Bench spans 23 fine-grained tasks across three core stages of the embodied interaction pipeline: Target Localization, Navigation, and Manipulation. Extensive evaluations of over 89 leading VLMs reveal that while advanced reasoning models show promise, current VLMs universally struggle with complex visual-text alignment for physical interactions. Specifically, models exhibit critical bottlenecks in multi-target counting, part-whole relationship understanding, and affordance region detection. EPIC-Bench provides a robust foundation and actionable insights for advancing the next generation of vision-driven embodied models.

Domain generalization aims to enhance model robustness against unseen domains with embedding distribution shifts. While large-scale vision-language models like CLIP exhibit strong generalization, their direct image-text embedding alignment suffers from inherent information asymmetry: images encode both class semantics and domain-specific styles, whereas text prompts primarily convey basic class cues. This asymmetry hinders generalization to novel domains in realistic scenarios. To address this, we propose Style-Homogenized Embedding alignment for Domain-generalization (SHED), a novel CLIP-based method that aligns style-homogenized embeddings instead of raw representations from encoders in CLIP. During training, SHED removes domain-specific style centroids from both image embeddings computed per source domains and text embeddings which are averaged across diverse prompt templates and stripped of a global centroid. For inference, considering the lack of target domain information, SHED projects diverse textual domain centroids into the visual space and aggregates predictions via membership weighting. Extensive experiments on five benchmarks show SHED achieves state-of-the-art performance, outperforming prior methods significantly (e.g., +4.0% on DomainNet vs. standard fine-tuning).

Existing vision-language forgery detection and grounding methods operate under a closed-world paradigm, assuming verification can be completed by the model alone. However, self-contained MLLMs are constrained by finite parametric knowledge, static training corpora, and limited perceptual resolution, creating a practical ceiling in dynamic open-world forensics -- particularly for real-time event verification requiring external clues and forgery segmentation demanding fine-grained scrutiny of local manipulations. To address these limitations, we shift from scaling up the self-contained model toward reaching beyond it. We propose \textbf{OmniVL-Guard Pro}, a tool-augmented agent that extends unified forensics from closed-world prediction to open-world clues-driven reasoning. OmniVL-Guard Pro integrates a tool environment spanning real-time event search, local cropping and zooming, edge-anomaly screening, face detection, video frame extraction, and SAM3-based segmentation. To generate high-quality tool-reasoning trajectories, we introduce \textbf{Tree-Structured Self-Evolving Tool Trajectory Generation}, which produces diverse trajectories through seed guidance, guider-free self-evolution, and weakly-hinted hard sample synthesis, yielding the Full-Spectrum Tool Reasoning (FSTR) dataset for training. We further propose \textbf{Checker-Guided Agentic Reinforcement Learning} (CGARL), which provides process-level supervision to penalize cases where the answer is correct but the reasoning is distorted. Extensive experiments demonstrate that OmniVL-Guard Pro achieves state-of-the-art performance across various tasks, and exhibits strong zero-shot generalization. The FSTR dataset and code for OmniVL-Guard Pro will be publicly released at https://github.com/shen8424/OmniVL-Guard-Pro.

Multi-goal semantic navigation couples searches through a shrinking horizon: effort spent on one request can make another effectively undiscoverable. Discoverability therefore depends on both goal evidence and the budget that other requests consume. We present BAT-Nav, Belief-Based Arbitration and Termination, which estimates remaining discoverability: the probability that a frozen executor can complete a goal within an additional budget. A calibrated local hazard converts navigation telemetry into a budget-conditioned completion curve. Its marginal return governs reversible; conservative belief and local evidence govern. On intervention-independent replays, the model obtains Brier .132 and ECE .034 without transfer refitting; 50-action completion rises from .047 to .321 across marginal-return sextiles. On HM3D / ApexNav, BAT-Nav raises CR from .372 to .414 and obtains MGSR .173. Its gain widens from 3.4 to 5.8 CR points as competing goals increase from two to five, with the same ordering on MP3D and under infeasible requests.

Vision-Language-Action (VLA) models follow a data-driven paradigm and are constrained by the coverage of training data, making them prone to failure on edge-case configurations after deployment. To mitigate such risks, it is essential to expose high-quality failure modes and convert the resulting failures into supervisory data for model enhancement. Existing studies largely stop at failure detection and lack a mechanism for leveraging discovered failures for model repair. We propose VLAMotor, the first analysis framework for VLA enhancement, which integrates distance-aware model testing for failure exposure and agent-based data synthesis for model finetunning. First, VLAMotor estimates input uncertainty based on the distance to training samples, and combines uncertainty ranking with redundancy elimination to build compact test sets that expose diverse failures. Then, VLAMotor abstracts failure trajectories into structured semantic representations, and plans parameterized repair-skill sequences, which are then realized as executable trajectories through inverse kinematics and motion execution. The resulting successful trajectories are automatically labeled and used to fine-tune the original VLA model, yielding an enhanced VLA model. Evaluation on four representative robotic manipulation tasks shows that 92.33% of the in-simulation test cases generated by VLAMotor trigger VLA failures, and VLAMotor improves test coverage over the state-of-the-art tool by 18.93%. By fine-tuning VLA models with synthetic data derived from failed test cases, VLAMotor further enhances the overall success rate of VLA models by 49.25%. When deployed on real hardware, the simulation-enhanced models improve the success rate over the original VLA models by 57.50%, demonstrating an effective and low-cost direction for VLA enhancement.

Planning from raw visual input remains a significant challenge for current Vision-Language Models (VLMs), when the complexity of input is beyond their one-step perception capability. Motivated by recent advances in Thinking with Images (TWI), a reasonable solution is to decompose the perception process into simpler steps by iteratively acquiring and incorporating local visual evidence. However, even though current VLMs are well-trained in general TWI ability, their perceptual bottleneck in the planning domain remains. To tackle this challenge, we formulate TWI as a tool to gradually build and reflect an accurate internal world model. We find that the resulting training-free planning strategy enables VLMs to solve tasks that are far beyond their initial capabilities, at the cost that too many TWI operations would significantly increase the computational overhead. To further improve efficiency, we propose Pattern Inference, a novel TWI strategy enabling VLMs to actively recognize known visual patterns in the new tasks and directly infer local world model structures. To obtain these patterns, we propose Pattern Induction, an online inductive learning strategy treating visual patterns as composite and reusable experts, which are autonomously discovered and optimized from experience. Experimental evaluations in FrozenLake, Crafter and CubeBench domains show that our approaches achieve a desirable balance between accuracy and efficiency.

Human-robot collaboration (HRC) can benefit from robots' abilities to interpret human emotional states. However, current emotion recognition (ER) models in HRC often fall short, particularly due to their reliance on acted datasets and single-modality inputs like facial expressions. We propose a novel vision language model (VLM)-based ER system that leverages contextual understanding to improve emotion interpretation in HRC. We first evaluate the VLM-ER system by assessing its semantic and sentiment similarity with human annotations on an existing HRC dataset. Then, in a user study with a service robot in a collaborative delivery task, we evaluate the effects of modulating the robot's behaviour based on the user's emotional state inferred by the VLM-ER system. The results show that the proposed VLM-ER system achieves higher semantic similarity and positive sentiment alignment with human annotations compared to a baseline convolutional neural network-based system. Further, participants in the user study preferred emotion-adaptive robot behaviour facilitated by the VLM-ER system.

Multi-video event understanding demands models that can locate and attribute query-relevant evidence scattered across long, heterogeneous video corpora. Existing large vision-language models (LVLMs) often underperform in this regime because they quickly exhaust their context budget and struggle to precisely localize evidentially important segments, frequently missing dense informational cues such as broadcast graphics, subtitles, and scoreboards. We introduce TRACE, an evidence grounding-guided framework that follows a ground-before-reasoning strategy for multi-video event reasoning. Our approach first builds a structured, text-searchable timeline for each video using OCR and object detection. A text-only LLM then conducts query-aware evidence localization, selecting relevant moments prior to any downstream visual reasoning. The retrieved frames and their grounding summaries are subsequently used to steer LVLM-based claim generation and cross-video citation consolidation. Experiments on MAGMaR 2026 and WikiVideo demonstrate that structured grounding markedly boosts factual completeness and attribution fidelity. On the MAGMaR validation split, TRACE raises macro-average MiRAGE F1 from 0.705 to 0.811 compared to an unguided Qwen3-VL-30B baseline, with especially strong improvements in citation recall from 0.440 to 0.628. The method also attains state-of-the-art results on the official MAGMaR 2026 leaderboard. Code is released at https://github.com/pengyu965/TRACE.

Large vision-language models (LVLMs) have emerged as a powerful paradigm for multimodal intelligence, but their growing deployment also expands the attack surface of prompt injection. Despite this growing concern, existing attacks still suffer from a critical limitation: the injected prompt for one modality only steers the model's interpretation of that singular input. Alternatively, these attacks remain multimodal but fail to achieve cross-modal prompt perturbation. To bridge this gap, we introduce a novel cross-modal prompt injection attack CrossMPI, which can steer the model's interpretation of both textual and visual inputs via image-only prompt injection. Our design is underpinned by the following key breakthroughs. First, we turn the focus of the injected prompt perturbation optimization from the visual embedding space (typically with only $10^5$ parameters) to the model hidden state space (for multimodal information integration and with $10^7$ parameters). Then, two strategies are adopted to mitigate the optimization challenges posed by the larger parameter space. To constrain the optimized model parameter space, we introduce a layer selection strategy that identifies the layers most critical to multimodal integration. Interestingly, deviating from the past experience, our analysis reveals that the optimal layers for LVLM prompt perturbation reside in the middle of the model rather than the last. To constrain the image perturbation space, we propose a new distance-decremental perturbation budget assignment strategy that allocates budgets decrementally as the pixel distance to semantic-critical regions increases. Extensive experiments across multiple LVLMs and datasets show that our method significantly outperforms baseline approaches.

Large Vision-Language Models (LVLMs) have shown significant progress in video understanding, yet they face substantial challenges in tasks requiring precise spatiotemporal localization at the instance level. Existing methods primarily rely on text prompts for human-model interaction, but these prompts struggle to provide precise spatial and temporal references, resulting in poor user experience. Furthermore, current approaches typically decouple visual perception from language reasoning, centering reasoning around language rather than visual content, which limits the model's ability to proactively perceive fine-grained visual evidence. To address these challenges, we propose VideoSeeker, a novel paradigm for instance-level video understanding through visual prompts. VideoSeeker seamlessly integrates agentic reasoning with instance-level video understanding tasks, enabling the model to proactively perceive and retrieve relevant video segments on demand. We construct a four-stage fully automated data synthesis pipeline to efficiently generate large-scale, high-quality instance-level video data. We internalize tool-calling and proactive perception capabilities into the model via cold-start supervision and RL training, building a powerful video understanding model. Experiments demonstrate that our model achieves an average improvement of +13.7% over baselines on instance-level video understanding tasks, surpassing powerful closed-source models such as GPT-4o and Gemini-2.5-Pro, while also showing effective transferability on general video understanding benchmarks. The relevant datasets and code will be released publicly.

Large-scale pre-trained vision-language models like CLIP demonstrate remarkable zero-shot performance across diverse tasks. However, fine-tuning these models to improve downstream performance often degrades robustness against distribution shifts. Recent approaches have attempted to mitigate this trade-off, but often rely on computationally expensive text-guidance. We propose a novel method for robust fine-tuning, SAE-FT, which operates only on the model's visual representations. SAE-FT regularizes changes to these representations by penalizing the addition and removal of semantically meaningful features identified by a Sparse Autoencoder trained on the pre-trained model. This constraint prevents catastrophic forgetting and makes the fine-tuning process interpretable, enabling direct analysis of semantic changes. SAE-FT is both mechanistically transparent and computationally efficient, matching or exceeding state-of-the-art performance on ImageNet and its associated distribution shift benchmarks. Code is publicly available at: https://github.com/Fabian-Mor/sae-ft.

Finetuning Large Vision-Language Models with reinforcement learning has emerged as a promising approach to enhance their capability in object-level grounding. However, existing methods, mainly based on GRPO, assign rewards at the response level. Such sparse reward, often criterion-induced, leads to minimal learning signals when all candidate responses fail in challenging scenarios. In this work, we propose a group-revision optimisation paradigm that enhances learning on hard cases. It begins with a sampled initial response and generates a set of revised candidates to explore improved grounding outcomes. Inspired by reward shaping, we introduce a consolidation process that quantifies each candidate's improvement over the initial attempt and converts it into informative shaping signals. These signals are used to both refine the reward and modulate the advantage, amplifying the influence of high-quality revisions. Our method achieves consistent gains across referring and reasoning segmentation, REC, and counting benchmarks compared with prior GRPO-based models. Our code is available at https://github.com/yyliu01/GroupRevision.

Open-vocabulary segmentation models often struggle to generalize to unseen combinations of object categories and attributes, because fine-grained descriptions are typically encoded as holistic sentences that entangle multiple semantic units. We propose a Decomposed Vision-Language Alignment framework that explicitly factorizes textual prompts into a concept token and multiple attribute tokens, enabling separate cross-modal interactions for each semantic unit. At the feature level, we introduce a Feature-Gated Cross-Attention module that generates attribute-specific gating maps to fuse information in a multiplicative manner, effectively enforcing compositional semantics. At the scoring level, per-token similarities are aggregated in log-space, producing a stable and interpretable compositional matching. The method can be seamlessly integrated into existing transformer-based segmentation architectures and significantly improves generalization to unseen attribute-category compositions in fine-grained open-vocabulary segmentation benchmarks.

Vision-Language Models (VLMs) excel at 2D tasks such as grounding and captioning, yet remain limited in 3D understanding. A key limitation is their text-only supervision paradigm, which under-constrains fine-grained visual perception and prevents the recovery of dense geometry. Prior methods either distill geometry from external vision models, introducing error accumulation, or enable direct prediction with inefficient per-pixel query or coarse token-level outputs. In this paper, we propose DepthVLM, a simple yet effective framework that transforms a single VLM into a native dense geometry predictor while preserving its multimodal capability. By attaching a lightweight depth head to the LLM backbone and training under a unified vision-text supervision paradigm with a two-stage schedule, DepthVLM generates full-resolution depth maps alongside language outputs in a single forward pass. We further introduce a unified indoor-outdoor metric depth benchmark in a VLM-compatible format. Experiments show that DepthVLM significantly outperforms existing VLMs with higher inference efficiency, surpasses leading pure vision models, and improves complex 3D spatial reasoning, moving toward a truly unified multimodal foundation model. The project page is available at https://depthvlm.github.io/

Vision-Language Models (VLMs) often produce self-reflective statements like "let me check the figure again" during reasoning. Do such statements trigger genuine visual re-examination, or are they merely learned textual patterns? We investigate this via VisualSwap, an image-swap probing framework: after a model reasons over an image, we replace it with a visually similar but semantically different one and test whether the model notices. We introduce VS-Bench, 800 image pairs curated from MathVista, MathVerse, MathVision, and MMMU-Pro. Experiments on Qwen3-VL, Kimi-VL, and ERNIE-VL reveal a striking failure: models overwhelmingly miss the swap, with accuracy dropping by up to 60%. Counterintuitively, thinking models are nearly 3x more vulnerable than their instructed counterparts, and scaling offers no mitigation. Multi-turn user instructions restore visual grounding, but self-generated reflective statements during continuous generation do not. Attention analysis explains why: user instructions substantially elevate attention to visual tokens, whereas self-reflection does not. Current VLMs tend to say rather than actually see when claiming to perform visual re-examination. Our code and dataset are available at the project page: https://visualswap.github.io

A central goal in understanding human vision is to uncover the visual features that drive neuronal activity. A growing body of work has used artificial neural networks as encoding models to predict cortical responses to natural images, revealing the visual content that activates category-selective regions. However, existing approaches are largely correlational and treat the encoder as a black box, leaving open which image features drive each voxel's response. We introduce Mechanistically Interpretable Neural Encoding (MINE), a framework that opens this black box by applying mechanistic-interpretability tools to localize the features within natural images that drive millimeter-scale (voxel-level) activity. MINE predicts each voxel's response using language-aligned image representations, and produces semantically interpretable descriptions of the features critical for the voxel's activation. We further generalize these per-image features into per-voxel functional profiles. To validate the per-image descriptions, we show they are sufficient to generate images that elicit voxel responses matching the responses to the original images, more accurately than images generated from random or low-attribution controls. Moreover, counterfactually inserting or removing the predicted features from images shifts activation in the expected direction, providing causal evidence. Counterfactual editing guided by the per-voxel activation profiles produces even stronger activation shifts, indicating that the profiles faithfully capture each voxel's selectivity. Finally, we apply MINE to well-studied category-selective brain regions, showing it recovers their known categorical preferences while revealing fine-grained unique voxel structure within each region. Overall, our results establish mechanistic interpretability as a path to discover and causally validate fine-grained hypotheses about neural function.

Learning visuomotor policies from scarce expert demonstrations remains a core challenge in robotic manipulation. A primary hurdle lies in distilling high-dimensional RGB representations into control-relevant geometry without overfitting. While using frozen pre-trained Vision Foundation Models (VFMs) improves data efficiency, it also shifts most task adaptation onto a small spatial pooling module, which can latch onto task-irrelevant shortcuts and lose geometric grounding when finetuned with few data samples. More broadly, pre-trained visual representations used for policy learning have been observed to struggle under even minor scene perturbations, highlighting the need for robustness-oriented inductive biases. We propose Geometric Anchor Pre-training (GAP), a simple, action-free warm-up stage that regularizes the spatial adapter before downstream imitation learning. GAP pre-trains the pooling layer on a lightweight simulated proxy task where object masks are available at no cost, encouraging the adapter to produce keypoints that lie on the object, cover its spatial extent, and remain sharp and repeatable over time. This yields stable geometric anchors that provide a reliable coordinate interface for few-shot policy learning, while keeping the VFM frozen. We evaluate GAP on RoboMimic and ManiSkill under severe data scarcity (15-50 demonstrations) and domain shift. A simple adapter regularized with GAP consistently outperforms stronger attention-based poolers and end-to-end fine-tuning, achieving 62% success on RoboMimic Can with 15 demonstrations (+16% over AFA), 63% on the long-horizon high-precision Tool Hang task with 50 demonstrations, and 61% on ManiSkill StackCube with 30 demonstrations (+11% over full fine-tuning). The proxy stage is lightweight and fully decoupled from downstream tasks, making it practical to reuse across environments and manipulation skills.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across various tasks, yet they remain vulnerable to backdoor attacks. Existing defense methods predominantly focus on sample-level defense, which relies on the knowledge of training data or triggers. However, identifying whether a given model is backdoored remains a critical but unexplored task. To fill this gap, we propose EntropyScan, a lightweight and trigger-agnostic method for model-level backdoor detection in LVLMs. We first observe that backdoor injection disrupts the cross-modal alignment, resulting in pronounced structural anomalies in visual attention allocation on benign samples. Based on this insight, EntropyScan detects the backdoor models by quantifying such attention deviations. Specifically, it extracts visual attention distributions from the initial layers of the Large Language Model (LLM) and applies Tsallis entropy to capture these structural distortions. By employing a reference-anchored Z-score normalization on a small set of benign samples, it effectively identifies the backdoored model. Extensive experiments across two LVLMs architectures and three advanced attack scenarios show that EntropyScan achieves an F1 score of 98.5% in average and an AUC of 96.6%. Our code will be publicly available soon.

Vision-language models (VLMs) achieve strong performance on multimodal benchmarks, but may still lack robust control over basic visual operations. We study \textit{line tracing}, where a model must follow a selected visual path through successive local continuations. To isolate this ability, we design controlled tracing tasks that introduce nearby competitors while reducing semantic and topological ambiguity such as crossings and overlaps. Across these tasks, even state-of-the-art VLMs frequently lose the target path and switch to nearby alternatives, especially when those alternatives look locally similar to the target. Behavioral interventions and internal analyses indicate that these failures arise from local competition: nearby similar distractors pull the model away from the true continuation. Standard remedies do not remove this bottleneck: model-size scaling provides only limited gains, reasoning partially compensates through costly substitute strategies, and explicit tracing instructions fail to recover stable path following. Finally, tests on tangled-cable scenes and metro maps with richer visual complexity show that the same path-switching failure persists beyond our controlled settings.

Large vision-language models (LVLMs) achieve strong multimodal understanding, but their inference cost grows rapidly with the number of visual tokens, especially for high-resolution images and long videos. Existing attention-based methods estimate token importance from attention scores, which may introduce positional bias, while representation-based methods reduce visual redundancy based on feature relations or reconstruction errors, overlooking the global structure of the visual token set. In this paper, we revisit visual token compression from the perspective of low-rank compressibility. Across models and datasets, we observe that visual token representations exhibit a pronounced low-rank structure, with a dominant subspace that remains stable even after a large fraction of tokens is randomly removed. Motivated by this finding, we propose LRCP, a training-free compression framework that first estimates the dominant low-rank subspace of visual tokens via PCA, and then scores each token by its projection residual onto this subspace, retaining tokens that are poorly explained by the low-rank background. Extensive experiments show that LRCP achieves superior results, preserving 94.7% of the original image-understanding performance with an 88.9% token reduction and 97.8% of the average video-understanding accuracy with an 87.5% token reduction.

Longitudinal chest X-ray (CXR) interpretation requires reasoning over disease evolution across multiple patient visits, yet most existing medical VQA benchmarks focus on single images or short-horizon image pairs. We introduce MI-CXR, a benchmark for standardized evaluation of Multi-Interval longitudinal reasoning over multi-visit CXR sequences, without requiring free-form report generation or additional clinical context. MI-CXR comprises five-way multiple-choice questions over five-visit patient timelines and instantiates three complementary task families: Temporal Event Localization, Interval-wise Change Reasoning, and Global Trajectory Summarization, which assess clinically grounded visual reasoning over time. Evaluating 14 state-of-the-art vision-language models (VLMs) shows low overall performance, with an average accuracy of 29.3%, only modestly above random guessing. Using stage-wise diagnostic probing, we find that models often produce locally plausible interval descriptions but fail to enforce temporal constraints or compose evidence into globally consistent decisions over the full timeline. These findings reveal key limitations of current VLMs and establish MI-CXR as a principled benchmark for longitudinal medical reasoning. The benchmark is available at https://github.com/AIDASLab/MI-CXR

Vision-Language Models (VLMs) facilitate medical visual question answering (MedVQA) by jointly interpreting images and text. However, existing models typically depend on large architectures and closed-set answers, which limits their efficiency and potential clinical applicability. To overcome these shortcomings, we introduce RoiMAM, an efficient VLM. It integrates a training-free ROI Generation Module with Semantic Selective Suppression to focus on lesion-relevant regions, alongside a Text Prompt Enhancer module that provides modality-specific context without introducing training parameters. Compared to the widely used MedVInT-TD model, our design achieves efficient and accurate diagnosis at less than 20% of the model size, while improving accuracy by approximately 2% on SLAKE and 4.6% on PMC-VQA.

Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets. We reveal a critical inefficiency in this approach: up to 69% of the prompts in standard chart / document reasoning datasets are effectively zero-delta, meaning the teacher and student already induce the exact same answer distribution. Training on these prompts provides minimal learning signal, causing student improvement to rapidly saturate regardless of data scale. To escape the zero-delta trap, we return to first principles: distillation fundamentally minimizes distributional divergence, and thus a prompt is valuable only if it exposes a functional capability gap between the teacher and student. We quantify this gap through answer divergence ($Δ$), demonstrating that non-zero divergence is critical for effective scaling. Building on this insight, we propose a staged synthesis pipeline that repurposes existing datasets as seeds, actively targeting student failure modes to produce better prompts. The result is DeltaPrompts, a diverse dataset of 200k synthetic, high-divergence reasoning problems. We evaluate DeltaPrompts across three distinct settings: on-policy distillation with the target teacher-student pair, transfer to a novel model family without regenerating the data, and off-policy fine-tuning of a non-reasoning model. Across all scenarios, DeltaPrompts drives substantial gains, yielding up to 15% relative improvement even on top of a highly-optimized reasoning model (e.g., Qwen3-VL-8B-Thinking) -- averaged over 10 benchmarks spanning chart, document and perception-centric reasoning.

Most Vision Language Models (VLMs) directly map outputs from ViT encoders to the LLM via a lightweight projector. While effective, recent analysis suggests this architecture suffers from an alignment challenge: visual features remain distant from the text space in the initial layers of the LLM, forcing the model to waste critical depth~\cite{zhang-etal-2024-investigating,artzy-schwartz-2024-attend} on superficial modality alignment rather than deep understanding and complex reasoning. In this work, we propose Deep Pre-Alignment (DPA), a novel architecture that replaces the standard ViT encoder with a small VLM as perceiver, ensuring visual features are deeply aligned with the text space of the target large language model. Comprehensive experiments demonstrate the effectiveness of DPA. On the 4B parameter scale, DPA outperforms baselines by 1.9 points across 8 multimodal benchmarks, with gains widening to 3.0 points at the 32B scale. Moreover, by offloading alignment to the perceiver, DPA achieves a 32.9% reduction in language capability forgetting over 3 text benchmarks. We further demonstrate that these gains are consistent across different LLM families including Qwen3 and LLaMA 3.2, highlighting the generality of our approach. Beyond performance, DPA also offers a seamless upgrade path for current VLM development, requiring only a modular replacement for the visual encoder with marginal computation overhead.

Visual reasoning, often interleaved with intermediate visual states, has emerged as a promising direction in the field. A straightforward approach is to directly generate images via unified models during reasoning, but this is computationally expensive and architecturally non-trivial. Recent alternatives include agentic reasoning through code or tool calls, and latent reasoning with learnable hidden embeddings. However, agentic methods incur context-switching latency from external execution, while latent methods lack task generalization and are difficult to train with autoregressive parallelization. To combine their strengths while mitigating their limitations, we propose ATLAS, a framework in which a single discrete 'word', termed as a functional token, serves both as an agentic operation and a latent visual reasoning unit. Each functional token is associated with an internalized visual operation, yet requires no visual supervision and remains a standard token in the tokenizer vocabulary, which can be generated via next-token prediction. This design avoids verbose intermediate visual content generation, while preserving compatibility with the vanilla scalable SFT and RL training, without architectural or methodological modifications. To further address the sparsity of functional tokens during RL, we introduce Latent-Anchored GRPO (LA-GRPO), which stabilizes the training by anchoring functional tokens with a statically weighted auxiliary objective, providing stronger gradient updates. Extensive experiments and analyses demonstrate that ATLAS achieves superior performance on challenging benchmarks while maintaining clear interpretability. We hope ATLAS offers a new paradigm inspiring future visual reasoning research.

Vision-Language Models (VLMs) are increasingly applied to cultural heritage materials, from digital archives to educational platforms. This work identifies a fundamental issue in how these models interpret historical artifacts. We define this phenomenon as cultural anachronism, the tendency to misinterpret historical objects using temporally inappropriate concepts, materials, or cultural frameworks. To quantify this phenomenon, we introduce the Temporal Anachronism Benchmark for Vision-Language Models (TAB-VLM), a dataset of 600 questions across six categories, designed to evaluate temporal reasoning on 1,600 Indian cultural artifacts spanning prehistoric to modern periods. Systematic evaluations of ten state-of-the-art models reveal significant deficiencies on our benchmark, and even the best model (GPT-5.2) achieves only 58.7% overall accuracy. The performance gap persists across varying architectures and scales, suggesting that cultural anachronism represents a significant limitation in visual AI systems, regardless of model size. These findings highlight the disparity between current VLM capabilities and the requirements for accurately interpreting cultural heritage materials, particularly for non-Western visual cultures underrepresented in training data. Our benchmark provides a foundation for enhancing temporal cognition in multimodal AI systems that interact with historical artifacts. The dataset and code are available in our project page.

Large vision-language models (LVLMs) have achieved remarkable performance across diverse multimodal tasks, yet they continue to suffer from hallucinations, generating content that is inconsistent with the visual input. Prior work DHCP (Detecting Hallucinations by Cross-modal Attention Pattern) has explored hallucination detection from the perspective of cross-modal attention, but does not address hallucination mitigation. In this paper, we propose MHSA (Mitigating Hallucinations via Steered Attention), a lightweight framework that mitigates hallucinations by learning to correct cross-modal attention patterns in LVLMs. MHSA trains a simple three-layer MLP generator to produce corrected attention, guided by supervisory signals from the DHCP discriminator and the LVLM itself. During inference, MHSA mitigates both discriminative and generative hallucinations across various datasets and LVLMs by simply replacing the original cross-modal attention with the corrected one, without modifying any LVLM parameters. By extending cross-modal attention mechanisms from hallucination detection to hallucination mitigation, MHSA offers a novel perspective on hallucination research in LVLMs and helps enhance their reliability.

Vision-Language-Action models have emerged as a promising paradigm for robotic manipulation by unifying perception, language grounding, and action generation. However, they often struggle in scenarios requiring precise spatial understanding, as current VLA models primarily rely on 2D visual representations that lack depth information and detailed spatial relationships. While recent approaches incorporate explicit 3D inputs such as depth maps or point clouds to address this issue, they often increase system complexity, require additional sensors, and remain vulnerable to sensing noise and reconstruction errors. Another line of work explores implicit 3D-aware spatial modeling directly from RGB observations without extra sensors, but it often relies on large geometry foundation models, resulting in higher training and deployment costs. To address these challenges, we propose Evo-Depth, a lightweight depth-enhanced VLA framework that enhances spatially grounded manipulation without relying on additional sensing hardware or compromising deployment efficiency. Evo-Depth employs a lightweight Implicit Depth Encoding Module to extract compact depth features from multi-view RGB images. These features are incorporated into vision-language representations through a Spatial Enhancement Module via depth-aware modulation, enabling efficient spatial-semantic enhancement. A Progressive Alignment Training strategy is further introduced to align the resulting depth-enhanced representations with downstream action learning. With only 0.9B parameters, Evo-Depth achieves superior performance across four simulation benchmarks. In real-world experiments, Evo-Depth attains the highest average success rate while also exhibiting the smallest model size, lowest GPU memory usage, and highest inference frequency among compared methods.

Recent advances in vision-language models (VLMs) have achieved impressive results on standard image-text tasks, yet their potential for visual procedure question answering (VP-QA) remains largely unexplored. VP-QA presents unique challenges where users query next-step actions by uploading images for intermediate states of complex procedures. To systematically evaluate VLMs on this practical task, we propose ProcedureVQA, a novel multimodal benchmark specifically designed for visual procedural reasoning. Through comprehensive analysis, we identify two critical limitations in current VLMs: inadequate cross-modal retrieval of structured procedures given visual states, and misalignment between image sequence granularity and textual step decomposition. To address these issues, we present Chain-of-Procedure (CoP), a hierarchical reasoning framework that first retrieves relevant instructions using visual cues, then performs step refinement through semantic decomposition, and finally generates the next step. Experiments across six VLMs demonstrate CoP's effectiveness, achieving up to 13% absolute improvement over standard baselines.

Video reasoning segmentation requires localizing objects across video frames from natural language expressions, often involving spatial reasoning and implicit references. Recent approaches leverage frozen large vision-language models (LVLMs) by extracting attention maps and using them as spatial priors for segmentation, enabling training-free grounding. However, these attention maps are optimized for text generation rather than spatial localization, often resulting in diffuse and ambiguous grounding signals. In this work, we introduce SteerSeg, a lightweight framework that identifies attention misalignment as the key bottleneck in attention-based grounding and proposes to steer attention at its source through input-level conditioning. SteerSeg combines learnable soft prompts with reasoning-guided Chain-of-Thought (CoT) prompting. The soft prompts reshape the attention distribution to produce more spatially concentrated maps, while CoT-derived attributes resolve ambiguity among similar objects by guiding attention toward the correct instance. The resulting attention maps are converted into point prompts across keyframes to guide a segmentation model, while candidate tracklets are ranked and selected using correlation-based scoring. Our approach freezes the LVLM and segmentation model parameters and learns only a small set of soft prompts, preserving the model's pretrained reasoning capabilities while significantly improving grounding. Despite being trained only on Ref-YouTube-VOS, SteerSeg generalizes well across diverse benchmarks, significantly improving the spatial grounding capability of LVLMs. Project page: https://steerseg.github.io

Memory is essential for large vision-language models (LVLMs) to handle long, multimodal interactions, with two method directions providing this capability: long-context LVLMs and memory-augmented agents. However, no existing benchmark conducts a systematic comparison of the two on questions that genuinely require multimodal evidence. To close this gap, we introduce MEMLENS, a comprehensive benchmark for memory in multimodal multi-session conversations, comprising 789 questions across five memory abilities (information extraction, multi-session reasoning, temporal reasoning, knowledge update, and answer refusal) at four standard context lengths (32K-256K tokens) under a cross-modal token-counting scheme. An image-ablation study confirms that solving MEMLENS requires visual evidence: removing evidence images drops two frontier LVLMs below 2% accuracy on the 80.4% of questions whose evidence includes images. Evaluating 27 LVLMs and 7 memory-augmented agents, we find that long-context LVLMs achieve high short-context accuracy through direct visual grounding but degrade as conversations grow, whereas memory agents are length-stable but lose visual fidelity under storage-time compression. Multi-session reasoning caps most systems below 30%, and neither approach alone solves the task. These results motivate hybrid architectures that combine long-context attention with structured multimodal retrieval. Our code is available at https://github.com/xrenaf/MEMLENS.

Contrastively pre-trained Vision-Language Models (VLMs) serve as powerful feature extractors. Yet, their shared latent spaces are prone to structural anomalies and act as repositories for non-semantic, multi-modal noise. To address this phenomenon, we employ spectral decomposition of covariance matrices to decompose the VLM latent space into a multi-modal semantic signal component and a shared noise subspace. We observe that this noise geometry exhibits strong subgroup invariance across distinct data subsets. Crucially, pruning these shared noise dimensions is mainly harmless, preserving or actively improving downstream task performance. By isolating true semantic signals from artifactual noise, this work provides new mechanistic insights into the representational structure of modern VLMs, suggesting that a substantial fraction of their latent geometry is governed by shared, architecture-level noise rather than task-relevant semantics alone.

Despite rapid advancements, current text-to-image (T2I) models predominantly rely on a single-step generation paradigm, which struggles with complex semantics and faces diminishing returns from parameter scaling. While recent multi-step reasoning approaches show promise, they are hindered by ungrounded planning hallucinations lacking verification, monolithic post-hoc reflection, long-context optimization instabilities, and prohibitive inference latency. To overcome these bottlenecks, we propose the Closed-Loop Visual Reasoning (CLVR) framework, a comprehensive system that deeply couples visual-language logical planning with pixel-level diffusion generation. CLVR introduces an automated data engine with step-level visual verification to synthesize reliable reasoning trajectories, and proposes Proxy Prompt Reinforcement Learning (PPRL) to resolve long-context optimization instabilities by distilling interleaved multimodal histories into explicit reward signals for accurate causal attribution. Furthermore, to mitigate the severe latency bottleneck caused by iterative denoising, we propose $Δ$-Space Weight Merge (DSWM), a theoretically grounded method that fuses alignment weights with off-the-shelf distillation priors, reducing the per-step inference cost to just 4 NFEs without requiring expensive re-distillation. Extensive experiments demonstrate that CLVR outperforms existing open-source baselines across multiple benchmarks and approaches the performance of proprietary commercial models, unlocking general test-time scaling capabilities for complex visual generation.

Recent advancements in Vision-Language Models (VLMs) have demonstrated strong capabilities in general visual reasoning, yet their applicability to rigorous biometric tasks remains unexplored. This work presents an exploratory study evaluating the zero-shot performance of state-of-the-art VLMs (GPT-5.2 and Gemini 2.5 Pro) on the Signature Verification Challenge (SVC) benchmark. To enable visual processing, raw kinematic time-series are converted into static images, encoding pressure information into stroke opacity whenever available in the source data. Furthermore, we introduce a scoring protocol that extracts latent token probabilities to compute robust biometric scores. Experimental results reveal a significant performance dichotomy dependent on signal quality and forgery type. In random forgery scenarios, the zero-shot VLM achieves exceptional discrimination, with GPT-5.2 reaching an Equal Error Rate of 0.32% in mobile tasks, outperforming supervised state-of-the-art systems. Conversely, in skilled forgery scenarios, where the task is more challenging because both signatures are almost identical, the results are significantly worse, and a critical "Rationalization Trap" emerges: chain-of-thought (CoT) reasoning degrades performance as the model produces kinematic hallucinations to justify forgery artifacts as natural variability.

Reference-guided video editing takes a source video, a text instruction, and a reference image as inputs, requiring the model to faithfully apply the instructed edits while preserving original motion and unedited content. Existing methods fall into two paradigms, each with inherent limitations: decoupled encoders suffer from modality gaps when processing instructions and visual content independently, while unified vision-language encoders lose fine-grained spatial details by relying solely on final-layer representations. We observe that VLM layers encode complementary information hierarchically -- early layers capture localized spatial details essential for precise editing, while deeper layers encode global semantics for instruction comprehension. Building on this insight, we present MiVE (Multiscale Vision-language features for reference-guided video Editing), a framework that repurposes VLMs as multiscale feature extractors. MiVE extracts hierarchical features from Qwen3-VL and integrates them into a unified self-attention Diffusion Transformer, eliminating the modality mismatch inherent in cross-attention designs. Experiments demonstrate that MiVE achieves state-of-the-art performance by ranking highest in human preference, outperforming both academic methods and commercial systems.

Large vision-language models (LVLMs) often hallucinate when language priors dominate weak or ambiguous visual evidence. Existing contrastive decoding methods mitigate this problem by comparing predictions from the original image with those from externally perturbed visual inputs, but such references can introduce off-manifold artifacts and require costly extra forward passes. We propose SIRA, a training-free internal contrastive decoding framework that constructs a counterfactual reference inside the same LVLM by exploiting the staged information flow of multimodal transformers. Instead of removing visual information from the input, SIRA first lets image and text tokens interact through a shared prefix, forming an aligned multimodal state that preserves prompt interpretation, decoding history, positional structure, and early visual grounding. It then forks a counterfactual branch in later transformer layers, where attention to image-token positions is masked. This branch retains the shared multimodal context but lacks continued access to fine-grained visual evidence, yielding a language-prior-dominated internal reference for token-level contrast. During decoding, SIRA suppresses tokens that remain strong without late visual access and favors predictions whose advantage depends on the full visual pathway. Experiments on POPE, CHAIR, and AMBER with Qwen2.5-VL and LLaVA-v1.5 show that SIRA consistently reduces hallucinations while preserving descriptive coverage and incurring lower overhead than two-pass contrastive decoding. SIRA requires no training, external verifier, or perturbed input, and applies to open-weight LVLMs with white-box inference access.

Large diffusion vision-language models (LDVLMs) have recently emerged as a promising alternative to autoregressive models, enabling parallel decoding for efficient inference and leveraging bidirectional attention for global context. Despite these advances, their behavior under long-form generation remains underexplored. In this work, we show that existing LDVLMs suffer from repetitive generation and degraded visual grounding, and identify two underlying causes. First, repetitive generation originates from a mask token prior: since generation tokens are initialized as mask tokens, their hidden representations progressively drift toward a shared prior direction over generation steps. Second, a fundamental misalignment between the positional attention bias and the iterative unmasking process suppresses attention toward informative visual tokens, degrading visual grounding. Based on these insights, we propose a training-free approach, introducing Mask Prior Suppression and Monotonic RoPE Scaling to mitigate mask prior drift and positional attention collapse during decoding. Experiments on general multimodal benchmarks and visual grounding tasks demonstrate improvements over baseline LDVLMs, with robust gains on long-form description benchmarks. Our results show that these failures can be effectively addressed with a lightweight, plug-and-play strategy that requires no additional training and generalizes across diverse LDVLM architectures.

Interpreting ultra-high-resolution (UHR) remote sensing images requires models to search for sparse and tiny visual evidence across large-scale scenes. Existing remote sensing vision-language models can inspect local regions with zooming and cropping tools, but most exploration strategies follow either a one-shot focus or a single sequential trajectory. Such single-path exploration can lose global context, leave scattered regions unvisited, and revisit or count the same evidence multiple times. To this end, we propose GeoVista, a planning-driven active perception framework for UHR remote sensing interpretation. Instead of committing to one zooming path, GeoVista first builds a global exploration plan, then verifies multiple candidate regions through branch-wise local inspection, while maintaining an explicit evidence state for cross-region aggregation and de-duplication. To enable this behavior, we introduce APE-GRO, a cold-start supervised trajectory corpus that reformulates diverse UHR tasks as Global-Region-Object interactive reasoning processes with a unified, scale-invariant spatial representation. We further design an Observe-Plan-Track mechanism for global observation, adaptive region inspection, and evidence tracking, and align the model with a GRPO-based strategy using step-wise rewards for planning, localization, and final answer correctness. Experiments on RSHR-Bench, XLRS-Bench, and LRS-VQA show that GeoVista achieves state-of-the-art performance. Code and dataset are available at https://github.com/ryan6073/GeoVista.

Single-frame Infrared Small Target Detection (ISTD) aims to localize weak targets under heavy background clutter, yet dense pixel-wise annotations are expensive. Point supervision with online label evolution reduces annotation cost; however, lightweight CNN detectors often lack sufficient semantics, leading to noisy pseudo-masks and unstable optimization. To address this, we propose a hierarchical VFM-driven knowledge distillation framework that uses a frozen Vision Foundation Model (VFM) during training. We formulate point-supervised learning as a bilevel optimization process: the inner loop adapts a VFM-embedded teacher on reweighted training samples, while the outer loop transfers validation-guided knowledge to a lightweight student to mitigate pseudo-label noise and training-set bias. We further introduce Semantic-Conditioned Affine Modulation (SCAM) to inject VFM semantics into CNN features at multiple layers. In addition, a dynamic collaborative learning strategy with cluster-level sample reweighting enhances robustness to imperfect pseudo-masks. Experiments on diverse challenging cases across multiple ISTD backbones demonstrate consistent improvements in detection accuracy and training stability. Our code is available at https://github.com/yuanhang-yao/semantic-prior.

The rapid advancement of Large Vision-Language Models (LVLMs) is increasingly accompanied by unauthorized scraping and training on multimodal web data, posing severe copyright and privacy risks to data owners. Existing countermeasures, such as machine unlearning and watermarks, are inherent post-hoc approaches that act only after intellectual property infringement has already occurred. In this work, we propose MMGuard to empower data owners to proactively protect their multimodal data against unauthorized LVLM fine-tuning. MMGuard generates unlearnable examples by injecting human-imperceptible perturbations that actively exploit the learning dynamics of LVLMs. By minimizing the training loss, the perturbation creates an optimization shortcut, causing the model to overfit to the noise and thereby degrading downstream performance when the perturbation is absent during inference. To further strengthen this defense, MMGuard introduces a cross-modal binding disruption, strategically shifting LVLM attention to enforce a spurious correlation between the noise and the training target with theoretical guarantees. Enhanced by an ensemble learning strategy for cross-model transferability, MMGuard is evaluated against nine open-source LVLMs across six datasets. Our comprehensive results demonstrate effective, stealthy, and robust protection under white-box, gray-box, and black-box threat models, establishing a mechanistic advantage in proactively defending against aggressive fine-tuning exploitation.

We introduce CurveBench, a benchmark for hierarchical topological reasoning from visual input. CurveBench consists of \textbf{756 images} of pairwise non-intersecting Jordan curves across easy, polygonal, topographic-inspired, maze-like, and dense counting configurations. Each image is annotated with a rooted tree encoding the containment relations between planar regions. We formulate the task as structured prediction: given an image, a model must recover the full rooted containment tree induced by the curves. Despite the visual simplicity of the task, the strongest evaluated model, Gemini 3.1 Pro, achieves only \textbf{71.1%} tree-generation accuracy on CurveBench-Easy and \textbf{19.1%} on CurveBench-Hard. We further demonstrate benchmark utility through RLVR-style fine-tuning of open-weight vision-language models. Our trained Qwen3-VL-8B model improves over \texttt{Qwen-3-VL-8B-Thinking} from \textbf{2.8%} to \textbf{33.3%} tree-generation accuracy on CurveBench-Easy, exceeding GPT-5.4 and Claude Opus 4.5 under our evaluation protocol. The remaining gap, especially on CurveBench-Hard, shows that exact topology-aware visual reasoning remains far from solved.

Adverse weather removal (AWR) in real-world images remains challenging due to heterogeneous and unseen degradations, while distortion-driven training often yields overly smooth results. We propose PVRF, a unified framework that integrates zero-shot soft weather perceptions with velocity-constrained rectified-flow refinement. PVRF introduces an AWR-specific question answering module (AWR-QA) that uses frozen vision--language models (VLMs) to estimate soft probabilities of weather types and low-level attribute scores. These perceptions condition restoration networks via attribute-modulated normalization (AMN) and weather-weighted adapters (WWA), producing an anchor estimate for refinement. We then learn a terminal-consistent residual rectified flow with perception-adaptive source perturbation and a terminal-consistent velocity parameterization to stabilize learning near the terminal regime. Extensive experiments show that PVRF improves both fidelity and perceptual quality over state-of-the-art baselines, with strong cross-dataset generalization on single and combined degradations. Code will be released at https://github.com/dongw22/PVRF.

Since the emergence of Vision Transformer (ViT), it has been widely used in generative language model and generative visual model. Especially in the current state-of-art open source multimodal models, ViT obtained by CLIP or SigLIP method serves as the vision encoder backbone to help them acquire visual understanding capabilities. But this method leads to limitations in visual understanding for details, such as difficulty in recognizing small text or numbers in images. To address these issues, we propose a new model to unify pix token and word token into the generative language model. The new model also features with each pix of image having its own token embedding, color folding, global conditional attention approximation and image unsupervised pretraining. We conducted image unsupervised pretraining experiments using our new model to explore its potential. The experimental results show that it has good performance even in small model and with limited training data. We believe our model also conforms to the scaling law, as long as model parameters and training data increased, its performance will continue to improve.

Class-Incremental Learning (CIL) enables models to continuously integrate new knowledge while mitigating catastrophic forgetting. Driven by the remarkable generalization of CLIP, leveraging pre-trained vision-language models has become a dominant paradigm in CIL. However, current work primarily focuses on aligning global image embeddings (i.e., [CLS] token) with their corresponding text prompts (i.e., [EOS] token). Despite their good performance, we find that they discard the rich patch-level semantic information inherent in CLIP's encoders. For instance, when recognizing a rabbit, local patches may encode its distinctive cues, such as long ears and a fluffy tail, which can provide complementary evidence for recognition. Based on the above observation, we propose SPA (Semantic-guided Patch-level Alignment) for CLIP-based CIL, which aims to awaken long-neglected local representations within CLIP. Specifically, for each class, we first construct representative and diverse visual samples and feed them to GPT-5 as visual guidance to generate class-wise semantic descriptions. These descriptions are used to guide the selection of discriminative patch-level visual features. Building upon these selected patches, we further employ optimal transport to align selected patch tokens with semantic tokens from class-wise descriptions, yielding a structured cross-modal alignment that improves recognition. Furthermore, we introduce task-specific projectors for effective adaptation to downstream incremental tasks, and sample pseudo-features from stored class-wise Gaussian statistics to calibrate old-class representations, thereby mitigating catastrophic forgetting. Extensive experiments demonstrate that SPA achieves state-of-the-art performance.

Long-context modeling is becoming a core capability of modern large vision-language models (LVLMs), enabling sustained context management across long-document understanding, video analysis, and multi-turn tool use in agentic workflows. Yet practical training recipes remain insufficiently explored, particularly for designing and balancing long-context data mixtures. In this work, we present a systematic study of long-context continued pre-training for LVLMs, extending a 7B model from 32K to 128K context with extensive ablations on long-document data. We first show that long-document VQA is substantially more effective than OCR transcription. Building on this observation, our ablations further yield three key findings: i) for sequence-length distribution, balanced data outperforms target-length-focused data (e.g., 128K), suggesting that long-context ability requires generalizable key-information retrieval across various lengths and positions; ii) retrieval remains the primary bottleneck, favoring retrieval-heavy mixtures with modest reasoning data for task diversity; and iii) pure long-document VQA largely preserves short-context capabilities, suggesting that instruction-formatted long data reduces the need for short-data mixing. Based on these findings, we introduce MMProLong, obtained by long-context continued pre-training from Qwen2.5-VL-7B with only a 5B-token budget. MMProLong improves long-document VQA scores by 7.1% and maintains strong performance at 256K and 512K contexts beyond its 128K training window, without additional training. It further generalizes to webpage-based multimodal needle retrieval, long-context vision-text compression, and long-video understanding without task-specific supervision. Overall, our study establishes a practical LongPT recipe and an empirical foundation for advancing long-context vision-language models.

Cross-modal 3D medical image analysis requires voxelwise representations that remain anatomically consistent across imaging contrasts, scanners, and acquisition protocols. Recent work has shown that frozen 2D Vision Transformer (ViT) foundation models can support such representations, but typical pipelines extract features along a single anatomical axis and adapt those features inside a registration solver for one image pair at a time, leaving complementary viewing directions unused and producing representations that do not transfer to new volumes. We introduce VoxCor, a training-free fit--transform method for reusable volumetric feature representations from frozen 2D ViT foundation models. During an offline fitting phase, VoxCor combines triplanar ViT inference with a compact closed-form weighted partial least squares (WPLS) projection that uses fitting-time voxel correspondences to select modality-stable anatomical directions in the triplanar feature space. At transform time, new volumes are mapped by triplanar ViT inference and linear projection alone, without fine-tuning or registration. Voxel correspondences can then be queried directly by nearest-neighbor search. We evaluate VoxCor on intra-subject Abdomen MR--CT and inter-subject HCP T2w--T1w tasks using deformable registration, voxelwise k-nearest-neighbor segmentation, and segmentation-center landmark localization. VoxCor improves the hardest cross-subject, cross-modality transfer settings, reduces encoder sensitivity for dense correspondence transfer, and yields registration performance competitive with handcrafted descriptors and learned 3D features. This positions VoxCor as a reusable feature layer for downstream multimodal analysis beyond pairwise registration. Code, configuration files, and implementation details are publicly available on GitHub at \href{https://github.com/guneytombak/VoxCor}{guneytombak/VoxCor}.

Vision-Language Models (VLMs) have achieved strong performance on general multimodal reasoning, yet remain challenged in integrating nonlocal visual information to support semantically underdetermined visual reasoning. We describe this challenge as Fragmented Visual Reasoning. To this end, we propose Credit Assignment for Visual Evidence (CAVE), a structured process-reward method based on GRPO for interleaved visual reasoning. Specifically, CAVE evaluates the contribution of intermediate steps at the action level via three complementary reasoning process signals: belief update, evidence acquisition, and adaptive focus control, thereby guiding the model to optimize each reasoning action and learn more reliable visual reasoning strategies. Meanwhile, we construct TRACER-Bench, which covers four nonlocal and semantically confusable reasoning dimensions and provides key intermediate evidence to supervise reasoning paths. Experiments demonstrate that CAVE substantially improves performance on tasks requiring fragmented visual evidence integration, covering both public benchmarks and our newly introduced TRACER-Bench, while retaining competitive performance on general multimodal evaluations. Further analyses reveal that CAVE effectively improves the visual reasoning capacity and exhibits stronger robustness under longer-range and deeper cross-region dependencies.

LLMs are widely adopted in production, pushing inference systems to their limits. Disaggregated LLM serving (e.g., PD separation and KV state disaggregation) improves scalability and cost efficiency, but it also turns KV into an explicit payload crossing network and storage boundaries, making KV a dominant end-to-end bottleneck. Existing KV compression are typically static runtime configurations, despite production service context varies over time in workload mix, bandwidth, and SLO/quality budgets. As a result, a fixed choice can be suboptimal or even increase latency. We present \emph{KVServe}, the first service-aware and adaptive KV communication compression framework for disaggregated LLM serving: KVServe (1) unifies KV compression into a modular strategy space with new components and cross-method recomposition; (2) introduces Bayesian Profiling Engine that efficiently searches this space and distills a 3D Pareto candidate set, reducing $50\times$ offline search overhead; and (3) deploys a Service-Aware Online Controller that combines an analytical latency model with a lightweight bandit to select profiles under constraints and correct offline-to-online mismatch. Integrated into vLLM and evaluated across datasets, models, GPUs and networks, KVServe achieves up to $9.13\times$ JCT speedup in PD-separated serving and up to $32.8\times$ TTFT reduction in KV-disaggregated serving.

Test-time scaling (TTS) methods have proven highly effective for LLMs, yet their application to vision-language models (VLMs) remains relatively underexplored. Existing VLM TTS methods largely require open-weight model access or expensive repeated sampling, and are evaluated primarily on multimodal mathematical and scientific reasoning benchmarks rather than general visual understanding tasks. In this paper, we propose Test-Time Hinting, a method that improves VLM performance via a single VLM call and requiring only black-box API access, which makes it broadly applicable to frontier closed-weight models. Our method is motivated by the observation that VLM errors tend to cluster around recurring failure patterns. We therefore train a lightweight hint generator model to predict, for a given test input, which "hint" should be prepended to the prompt, providing targeted contextual or procedural guidance that steers the VLM away from its characteristic failure modes. We show that Test-Time Hinting improves the accuracy of multiple closed-weight VLMs on natural-image VQA benchmarks and that these gains generalize to unseen benchmarks and VLMs without retraining the hint generator.

Reinforcement Learning with Verifiable Rewards (RLVR) traditionally relies on a sparse, outcome-based signal. Recent work shows that providing a fine-grained, model-intrinsic signal (rewarding the confidence growth in the ground-truth answer) effectively improves language reasoning training by providing step-level guidance without costly external models. While effective for unimodal text, we find that naively applying this global reward to vision-language (V-L) reasoning is a suboptimal strategy, as the task is a heterogeneous mix of sparse visual perception and dense textual reasoning. This global normalization creates mixture-induced signal degradation, where the training signal for visual steps is statistically distorted by the predominant textual steps. We propose Perception-Decomposed Confidence Reward (PDCR), a framework that solves this by aligning the reward structure with the task's heterogeneous nature. PDCR first performs an unsupervised skill decomposition, introducing a model-internal Visual Dependence Score to quantify visual reliance and applying a clustering algorithm to separate perception and reasoning steps. Based on this, PDCR computes a decomposed advantage by normalizing confidence gains within each skill cluster. This intra-cluster normalization provides a stable, correctly-scaled signal for both perception and reasoning. We demonstrate that PDCR outperforms the naive, global-reward formulation and sparse-reward baselines on key V-L reasoning benchmarks.

Edge deployment of Vision-Language Models (VLMs) faces a tradeoff between latency and accuracy: cloud execution provides high-quality predictions but incurs communication delay and energy cost, while edge-only execution is faster but less accurate due to limited model capacity. This trade-off is further complicated by heterogeneity in image quality and reasoning complexity, making static placement suboptimal. We present INAR-VL, a lightweight edge-cloud routing system for multimodal inference in a two-tier deployment. INAR-VL maintains complementary VLMs across edge and cloud and uses lightweight image and text complexity signals to guide routing and model selection, executing simple queries locally while offloading complex ones when beneficial. Evaluation on visual question answering shows that INAR-VL executes 36% of requests on the edge, reduces latency by 24%, lowers energy by 26%, and preserves 97% of cloud-level accuracy.

In Vision-Language Models (VLMs), processing a massive number of visual tokens incurs prohibitive computational overhead. While recent training-aware pruning methods attempt to selectively discard redundant tokens, they largely rely on continuous-gradient relaxations. However, visual token pruning is inherently a discrete, non-convex combinatorial problem; consequently, these continuous approximations frequently trap the optimization in sub-optimal local minima, especially under aggressive compression budgets. To overcome this fundamental bottleneck, we propose GRIP-VLM, a Group-Relative Importance Pruning framework driven by Reinforcement Learning. Rather than relying on smooth-gradient assumptions, GRIP-VLM formulates pruning as a Markov Decision Process, employing a Group Relative Policy Optimization (GRPO) paradigm anchored by supervised warm-up to directly explore the discrete selection space. Integrated with a budget-aware scorer, our lightweight agent dynamically evaluates per-token importance and adapts to arbitrary compression ratios without retraining. Extensive experiments across diverse multimodal benchmarks demonstrate that GRIP-VLM consistently outperforms heuristic and supervised-learning baselines, achieving a superior Pareto frontier and delivering up to a 15% inference speedup at equal accuracy.

Standard dual-encoder vision-language models that map images and text to deterministic points on a shared unit hypersphere through $\ell_2$ normalization typically expose neither \emph{aleatoric} uncertainty (cross-modal ambiguity) nor \emph{epistemic} uncertainty (lack of training-distribution support). Existing post-hoc methods either recover at most one of the two uncertainty components, or ignore the hyperspherical geometry of these models' embeddings. We propose \textbf{GeoFlowVLM} as a post-hoc adapter that learns the joint distribution of paired $\ell_2$-normalised dual-encoder VLM embeddings on the product hypersphere $\mathbb{S}^{d-1} \times \mathbb{S}^{d-1}$ via Riemannian flow matching with a single masked velocity field. A consistency result shows that, in the population limit, the trained network exposes the joint flow and both cross-modal conditional flows as valid Riemannian flow-matching velocity fields on their respective domains. We derive two quantities from this single model: a conditional retrieval entropy that quantifies aleatoric ambiguity with a decision-theoretic interpretation via a Fano-type bound, and a marginal-typicality epistemic score justified by an exact chain-rule decomposition of the joint NLL. This decomposition isolates a cross-modal pointwise-mutual-information term that is structurally discriminative rather than epistemic, and is empirically the only consistently uninformative standalone component. Empirically, the entropy tracks Recall@1 with near-ideal monotonic calibration across three retrieval benchmarks in both directions, and the marginal-typicality sum yields consistently calibrated selective accuracy across four zero-shot classification benchmarks.

In large vision-language models, visual tokens typically constitute the majority of input tokens, leading to substantial computational overhead. To address this, recent studies have explored pruning redundant or less informative visual tokens for image understanding tasks. However, these methods struggle with pixel grounding tasks, where token importance is highly contingent on the input text. Through an in-depth analysis of CLIP, we observe that visual tokens within referent regions often exhibit low similarity to their textual representation. Motivated by this insight, we introduce LiteLVLM, a training-free, text-guided token pruning strategy for efficient pixel grounding inference. By reversing the ranking of CLIP's visual-text similarity, LiteLVLM effectively retains visual tokens covering the referent regions, while recovering context tokens to enable clear foreground-background separation. Extensive experiments demonstrate that LiteLVLM significantly outperforms existing methods by over 5% across diverse token budgets. Without any training or fine-tuning, LiteLVLM maintains 90% of the original performance with a 22% speedup and a 2.3X memory reduction. Our code is available at https://github.com/sejong-rcv/LiteLVLM.

Brain encoder models predict cortical fMRI responses from the internal activations of pretrained vision and language networks, and are typically evaluated by held-out prediction accuracy. This is a useful signal for training but a poor one for interpretation: it tells us an encoder fits the data without telling us whether it has internalized the functional organization of the brain. We propose feature visualization -- gradient ascent on the encoder's predicted activation for a target region of interest (ROI) -- as a complementary interpretability technique, and apply it to TRIBE v2 composed with V-JEPA 2 (ViT-G, 40 layers), holding both frozen and synthesizing still images for seven regions spanning the ventral and dorsal visual hierarchies. Under identical hyperparameters, the probe recovers a visible progression of increasing spatial scale and feature complexity across V1 to V4, matching the ventral-stream hierarchy. It also produces three distinctive downstream regimes: radial "frozen-motion" streaks for the middle temporal area (MT) despite static-only optimization, face-like features for the fusiform face area (FFA), and consistent rectilinear line patterns for the parahippocampal place area (PPA). Optimized FFA stimuli drive the predicted region ~4x as much as a natural face photograph, consistent with feature visualization producing adversarial super-stimuli rather than canonical exemplars. The probe is simple, differentiable, and applicable to any brain encoder with a differentiable backbone, allowing for qualitative evaluation of brain encoders.

Large language-vision models (LVLMs) such as CLIP, Flamingo, and BLIP have revolutionized AI by enabling understanding across textual and visual modalities. These models excel at tasks like image captioning, visual question answering, and cross-modal retrieval. However, they face catastrophic forgetting when learning new tasks sequentially, particularly challenging in multi-modal settings where preserving cross-modal alignments adds complexity to the learning process. This paper presents a comprehensive continual learning framework for LVLMs that combines enhanced Elastic Weight Consolidation (EWC) with parameter-efficient fine-tuning techniques. We integrate multi-modal Fisher Information Matrix calculation, consistency preservation across modalities, and adaptive regularization that considers dependencies across visual and textual encoders. The framework achieves a 78% reduction in forgetting rates relative to naive sequential training approaches through extensive evaluation testing. The framework also preserves alignment between modalities during sequential learning with only 15% additional computational cost. This work advances the state of the art in lifelong learning for multi-modal AI systems, with direct applications to autonomous driving, intelligent robotic assistants, and adaptive robotic systems that must continuously learn in dynamic real-world environments.