Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering and controlling multimodal behavior. MMDiff supports three uses: (i) feature isolation, by diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training; (ii) task-specific feature detection, via per-token contrastive firing analysis that isolates causal features; and (iii) feature-level control, by causally removing or steering the discovered feature directions. We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR. MMDiff discovers sparse, causally specific features whose removal selectively degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR, and reduces attack success rate by 24% on multimodal safety attacks, with no impact on VQA performance. Steering these features improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer steering baseline. These results show that multimodal SAEs can serve not only as interpretability tools, but as mechanisms for auditing, steering, and controlling MLLMs behavior toward safer and more capable generations.
Robust visual classification often depends on localizing the main foreground objects in an image while ignoring contextual distractors. Surprisingly, we find that the last layer attention maps of smaller self-supervised ViTs localize foreground objects better than those of larger ViTs. However, we still need large ViTs, because they extract richer representations from each patch. To get the best of both worlds, good localization and rich representations, we propose $A^2$, a simple method that leverages this inverse scaling finding by decoupling where to look (a small attention model) from what to extract (a large embedding model): we crop around the attention peaks of a small model and embed the crops with a larger model. $A^2$ uses entirely pretrained features, requires no group labels, and does not require per-dataset attention or backbone training. Across 5 benchmarks, $A^2$ is competitive with backbone-matched loss-level methods like DFR, and outperforms end-to-end attention training under stronger distribution shifts.
Computer vision offers a scalable, non-contact means of estimating livestock traits, but correlated images and video frames make reliable evaluation difficult. Image-random splitting can place views of the same animal in both training and test sets, while deployment further requires transfer to animals captured in future years. We introduce BoviShift, a corpus reconstructed from six consecutive annual cohorts of an operational cattle marketplace collected from 2021 to 2026. It contains 2{,}657 cattle, 10{,}610 still images, and 2{,}566 walk-around videos across nine recorded breed categories. We evaluate three splitting strategies through four configurations: image-random and pooled cattle-disjoint splits over the same development cohorts, followed by separate future-2025 and locked future-2026 evaluations. Using a fixed ResNet-50 multi-task baseline, we assess weight, age, height, breed, and coat color. Differences between image-random and cattle-disjoint splitting are task-dependent rather than a universal leakage penalty. Future-cohort behavior is also attribute- and cohort-specific: breed declines relative to pooled evaluation; mean weight and age performance fails to beat constant baselines in 2025; height remains above baseline; and color transfers well to 2025 but degrades in 2026. Multi-task training improves all five point estimates in a single-seed ablation, still-view aggregation helps selectively, and zero-shot video-frame aggregation exposes a still-to-video domain gap. To our knowledge, BoviShift is the longest-spanning cattle-vision resource combining consecutive annual cohorts with physical and visual attributes.
Reflectance confocal microscopy (RCM) provides noninvasive, cellular-resolution "optical biopsies" of human skin in vivo by acquiring en-face images at successive depths, forming a sparse z-stack. Due to optical limitations, these stacks are anisotropic 3D volumes with lateral resolution (0.5 $μ$m) $\sim$6 times higher compared to axial resolution, which is defined by the optical sectioning (3 $μ$m), limiting the interpretation of tissue. Our goal is to provide continuous-depth visualization by interpolating intermediate sections and making the 3D volume isotropic. Such a representation permits arbitrary-direction sectioning, including histopathology-like cross-sectional examination, without requiring per-patient optimization. To that end, we introduce the first RCM-specific novel-view synthesis (NVS) approach, CD-RCM, a feedforward model that predicts realistic, unseen depths from sparsely sampled RCM stacks. Classical neural rendering methods focus on reconstruction from surface-level multi-view observations. In contrast to surface-level camera views, RCM can acquire optically sectioned en-face images of tissue beyond the surface up to 200 $μ$m. However, during visualization of the RCM stacks, observations of the shallower sections (towards the surface) obscure the deeper ones. This unique axial imaging geometry and layer-dependent anatomical organization motivated our development of a tailored architectural and training framework that explicitly accounts for RCM's depth-resolved, occlusive imaging physics. Experiments demonstrate that CD-RCM achieves high-fidelity novel-view synthesis with sub-second inference time.
提出 CAAP，通过在 ViT 中间层直接替换 patch 的内部激活来估计每个图像块对预测的因果贡献，作者报告其在多个骨干和指标上优于现有归因方法。
cs.CV
查看原文摘要
Attribution methods for Vision Transformers (ViTs) aim to identify image regions that influence model predictions, but producing faithful and well-localized attributions remains challenging. Existing attribution methods face several limitations, with gradient-based, relevance-propagation, and attention-based methods relying on local approximations, while perturbation- or optimization-based methods intervene on inputs, tokens, or surrogates rather than internal patch representations. The key challenge is that class-relevant evidence is formed through interactions between patch tokens across layers; methods that operate only on input changes, attention weights, or backward relevance signals may therefore provide indirect proxies for patch importance rather than directly testing the predictive effect of contextualized patch representations. We propose Causal Attribution via Activation Patching (CAAP), which estimates the contribution of individual image patches to the ViT's prediction by directly intervening on internal activations rather than using learned masks or synthetic perturbation patterns. For each patch, CAAP inserts the corresponding source-image activations into a neutral target context over an intermediate range of layers and uses the resulting target-class score as the attribution signal. The resulting attribution map reflects the causal contribution of patch-associated internal representations on the model's prediction. The causal intervention serves as a principled measure of patch influence by capturing semantic evidence after initial representation formation, while avoiding late-layer global mixing that can reduce spatial specificity. Across multiple ViT backbones and standard metrics, CAAP consistently outperforms existing methods in various settings and produces more faithful and localized attributions.
Unsupervised video object-centric learning aims to decompose dynamic scenes into persistent, object-level representations without supervision. However, existing slot-based methods struggle to maintain stable object identity in challenging settings such as rapid motion and partial occlusion. First, they typically encode both the per-frame appearance of an object and its identity across frames in a single slot vector, creating an objective conflict that leads to slot swapping: reconstruction requires sensitivity to transient visual changes, whereas temporal consistency requires invariance to them. Second, the token renormalization used in Slot Attention can amplify weakly attending slots, allowing them to absorb tokens from other objects and destabilize slot-to-object correspondence. We propose Dual-State Slot Attention (DSSA), a fully self-supervised framework that addresses these limitations by separating appearance from identity and by reducing spurious updates from weakly matching slots. DSSA decomposes each slot into a local state for per-frame appearance and an identity state for temporally stable object information, thereby aligning reconstruction and temporal consistency with separate representations. The identity state is updated through a learned recurrent transition that acts as a temporal filter on the local state, while competition-modulated aggregation (CMA) down-weights updates from weakly matching slots and prevents them from absorbing tokens from other objects. Experiments on MOVi-C, MOVi-D, and YouTube-VIS demonstrate that DSSA consistently improves segmentation quality and temporal consistency over prior methods, while also yielding stronger downstream object recognition and video dynamics prediction. Code is available at https://uark-aicv.github.io/dssa/.
Adding abstention to a task-adapted multimodal model can discard correct answers, including capabilities acquired during adaptation. We formulate abstention repair as maximizing correct abstention on unanswerable questions under a budget on newly lost correct answers. We propose ReFrain (Regression-Free Abstention), which combines a linear probe on first-token logits, threshold calibration on held-out image clusters, and a deterministic veto that replaces rejected outputs with the displayed "none of the above" option while preserving all other responses. "Regression-free" denotes zero loss on the calibration set at the zero-budget target. All abstention methods in our main comparison start from the same answering-SFT checkpoint; we separately track lost correct answers, repaired errors, and retained SFT gains. On Qwen2.5-VL-3B, ReFrain at zero budget retains all 57 SFT-gained answers with no losses and improves correct abstention by 7.30 percentage points. At a 0.5% calibration budget, it reaches 74.05% correct abstention with 8 losses and retains 55 gains; CLIP-UP EmbLoRA reaches 64.65% with 62 losses, 41 repairs, and 40 retained gains. The same probe at its default threshold loses 298 correct answers. With models and thresholds frozen, confirmation on 228 independent question clusters yields 66.05% correct abstention and no losses at the 0.5% point, exceeding EmbLoRA by 32.72 percentage points. Ordinary refusal SFT achieves higher correct abstention at greater regression cost. Code and reproduction materials: https://github.com/ModeLiberaL/ReFrain.
Remote-sensing foundation models increasingly unify representations across sensing modalities and downstream tasks, yet their prediction mechanisms remain largely task-specific. Separate decoders, prediction heads, and objectives leave prediction fragmented across tasks. This naturally raises a question: can heterogeneous remote-sensing tasks share the prediction process? UniFlow-RS answers this question with a single latent flow-matching predictor shared across tasks. This requires addressing two key challenges: unifying heterogeneous prediction spaces and harmonizing joint optimization across tasks. First, we cast heterogeneous task outputs into structured visual targets in a common latent space, allowing them to be modeled by the same pretrained flow predictor. Task-dependent conditioning further accommodates differences in input configurations. Second, we optimize all tasks with a common Flow Matching objective, eliminating task-specific prediction losses. To mitigate optimization mismatch across tasks and modalities, we introduce Task-Modality Adaptive Timestep Sampling (TMAS), which dynamically allocates training timesteps according to the online learning state of each task-modality group. Using a single shared model, UniFlow-RS handles six tasks and thirteen datasets, surpassing recent task-specific state-of-the-art methods on several benchmarks while remaining competitive on the others. Beyond these benchmarks, the same checkpoint demonstrates strong zero-shot transfer across unseen datasets, task settings, occlusion conditions, and near-infrared false-color imagery, without target-specific adaptation.
Shaokai Wu, Xunchu Zhou, Pinzhi Wang, Bo Dong, Xiao Cui, Qi Kong, Yue Ding, Hongtao Lu
原文摘要
A single validation-tuned threshold can become brittle when dense-score distributions shift. We introduce AutoThreshold, a post-hoc policy that predicts the joint mass of positive and negative pixels over score bins. This target directly parameterizes threshold-indexed confusion counts, avoiding products of independently estimated prevalence and class-conditional rates. Softmax normalization and cumulative decoding yield coherent, monotone confusion profiles; a plug-in decoder then selects a threshold for any $F_β$ objective without retraining the head. Across three independently trained Kvasir pipelines transferred to four polyp datasets, Joint exceeds continuous Otsu by $0.0080 \pm 0.0046$ external macro Dice, with positive gains in 3/3 runs. On the reference pipeline, three head seeds achieve $0.6324 \pm 0.0027$, 0.0085 above Otsu; direct-utility Ridge achieves 0.6342, with the paired difference interval spanning zero. The same Joint checkpoints outperform Otsu for $F_{0.5}$, Dice, and $F_2$. On developmental KITTI Road leave-one-corruption-family-out evaluation, Joint improves over scalar regression by 0.0082 [0.0064, 0.0101], while direct utility is 0.0022 higher. These results establish joint confusion prediction as a useful, metric-retargetable inductive bias, without claiming Bayes-optimal $F_β$ decisions.
World-Action Models improve robot control by predicting future visual dynamics, but iterative video denoising incurs substantial inference overhead. Recent approaches reduce this cost through partial future denoising, yet typically rely on fixed refinement budgets. Our empirical observations reveal that preferred refinement depths vary across tasks and states, while different future regions exhibit varying relevance to action generation. Motivated by these findings, we propose Elastic-WAM, which adaptively allocates world refinement computation across temporal denoising time and spatial future space. A unified spatiotemporal refinement demand field governs when to stop world refinement, how far to advance each denoising step, and which regions to refresh. We parameterize this field with a shared temporal basis and spatial amplitudes, and learn it through a three-stage action-grounded supervised procedure. At inference, the learned field guides recurrent sparse world refinement to improve task success while reducing inference overhead. Evaluations across diverse simulation benchmarks and real-world long-horizon manipulation demonstrate state-of-the-art control performance with reduced inference latency.
Pooya Ashtari, Pourya Behmandpoor, Nikos Deligiannis, Aleksandra Pizurica
原文摘要
Implicit neural representations (INRs) have emerged as powerful tools for encoding signals, yet dominant MLP-based designs often suffer from slow convergence, and how their accuracy depends on their size is known only empirically. We introduce FUTON (Fourier Tensor Network), which models signals as generalized Fourier series whose coefficients are parameterized by a low-rank tensor decomposition, read out by a linear or MLP head. FUTON combines complementary inductive biases: Fourier bases capture smoothness and periodicity, while the low-rank parameterization enforces low-dimensional spectral structure. Beyond universal approximation, we show that with a linear head, the best FUTON's error splits exactly into the truncation of the signal's series and the best low-rank error of its coefficients, which we bracket from the samples before training, with a FUTON that attains the upper bound. On image fitting, occupancy volumes, and novel view synthesis, against twelve INRs of matched size, FUTON achieves the best average quality, reaches the final quality of each task's strongest baseline 4-34x sooner than that baseline, infers fastest, and trains in the least memory.
Ziye Yuan, Minghong Cai, Xu He, Ruchang Yao, Xiao Fu, Kaixin Ding, Ming Zhang, Xiangyu Du 等 11 位作者
原文摘要
Human-centric audio-visual editing requires precise alignment of lip movements, gestures, and expressions with speech. Existing methods process inputs through independent video and audio pipelines, preventing cross-modal interaction between visual editing and audio dubbing. Many of these pipelines further rely on mask-based inpainting. As a result, they suffer from poor lip-audio synchronization, cumulative pipeline errors, and an inherent mask-size dilemma. We present AVE, the first unified end-to-end framework that jointly performs reference-guided person swapping, insertion, and deletion across both audio and video modalities without requiring spatial masks. To overcome data scarcity, we design a synthesis pipeline and introduce AVE-Dataset, the first audio-visual editing dataset to provide 100K pairs with appearance and voice references. By combining Synthetic-to-Real reconstruction with preliminary audio warm-up, our framework transcends synthetic pipeline limits and mitigates audio-video token imbalance. AVE outperforms modality-separate audio-visual editing pipelines in editing realism, temporal consistency, and audio-visual alignment.
Orest Kupyn, Théo Uscidda, Marta Tintore Gazulla, Fabian Manhardt, Federico Tombari, Christian Rupprecht
原文摘要
Video generation models have advanced significantly through the latent diffusion transformers trained with rectified flow techniques. Yet these models still struggle with geometric inconsistencies, unstable motion, and visual artifacts that break the illusion of realistic 3D scenes. 3D-consistent video generation could significantly impact numerous downstream applications in generation and reconstruction tasks. We explore how epipolar geometry constraints improve modern video diffusion models. Despite using massive training data, these models fail to capture fundamental geometric principles. We align diffusion models using pairwise epipolar geometry constraints via preference-based optimization, directly addressing unstable trajectories and geometric artifacts through mathematically principled geometric enforcement. Our approach efficiently enforces geometric principles without requiring end-to-end differentiability. Evaluation demonstrates that classical geometric constraints provide more reliable preference signals than the video-quality and geometry metrics based on foundational models. We construct preferences from static scenes with camera motion, where the epipolar metric is most reliable, while the aligned model generalizes to dynamic scenes. By bridging data-driven learning with classical computer vision, we reduce epipolar error by 31% and improve human-rated consistency from 54% to 72% without compromising visual quality.
Mahacine Ettahri, Hager Radi Abdelwahed, Toluwaleke Olutayo, Daniel Aloise, Gaétan Marceau Caron
原文摘要
Wildfires represent a critical ecological and socio-economic challenge in Canada, where vast boreal forests and a warming climate have amplified the frequency and severity of extreme fire seasons. Predicting next-day wildfire spread can support fire management, but requires capturing complex interactions among weather, fuels, and terrain. Deep learning has emerged as a promising approach to learning these relationships from heterogeneous environmental data, yet its development and evaluation depend on representative datasets. Existing deep learning research on wildfire spread relies predominantly on US-focused datasets, leaving Canadian fire environments underrepresented. To address this gap, we present DailyWildfireCAN, a multi-modal, multi-temporal dataset and benchmark for predicting next-day wildfire spread across Canada. The dataset tracks continuous fire progression over multi-day events at a high spatial and temporal resolution of 90-m and 24-hour, respectively. This resolution preserves fine-scale information relevant to fire-edge geometry, local fuel variations, and terrain effects, supporting research toward improved daily wildfire spread forecasting. DailyWildfireCAN covers 2,440 wildfire events across eight fire seasons 2017-2024 and integrates 33 features across diverse modalities, including topography, meteorology, land cover, vegetation and forest structure, human influence indices, and satellite imagery. We establish a comprehensive benchmark comparing mono-temporal and multi-temporal prediction models, accompanied by a multi-dimensional evaluation framework assessing robustness across wildfire characteristics, geographic regions, and temporal conditions, spatial prediction quality, and sensitivity to reference label variations.
Meituan Robotics Team, Caicheng Wang, Chen Zhao, Chengpeng Jin, Duo Xu, Fengjiao Chen, Guanglu Wan, Heqing Yang 等 38 位作者
原文摘要
We present Meituan-Robotics-0 (MR0), a vision-language-action foundation model and embodied-agent system for stationary-base desktop manipulation. MR0 combines 147 million vision-language samples with 15,060 hours of curated robot trajectories spanning 12 embodiments. A three-stage training recipe aligns multimodal representations with robot behavior through discrete action-token pre-training, transfers these priors to continuous cross-embodiment control through flow-matching mid-training, and adapts the policy through real-world post-training. We introduce Rollback, which pairs replaced policy segments with human corrections to provide positive and negative supervision for CFG-DAgger without value learning or advantage labeling. A unified end-effector interface supports heterogeneous embodiments, while asynchronous action-chunk merging maintains smooth execution under variable inference latency. An event-driven agent maintains persistent task events and reassesses plans when progress or feasibility changes, enabling cross-task retrieval and conditional task resumption. MR0 achieves 93.6% success on RoboTwin 2.0 and ranks fifth in the RoboDojo VLA track. Under matched HG-DAgger post-training, MR0 achieves higher average success rate and progress score than $π_{0.5}$ across three real-robot tasks. On Ethernet insertion, CFG-DAgger trained with Rollback data reaches 70% success and a progress score of 0.90. Qualitative agent deployments demonstrate cross-task event retrieval and recovery from temporarily unsatisfied task conditions.
Mesoscale eddies carry most of the ocean's kinetic energy and move heat, salt and nutrients across basins, yet where an eddy will be two weeks ahead is still read off ocean models with post-hoc tracking heuristics. We introduce HaiRuo, which forecasts eddies as fields and as objects at once: from 30 days of satellite sea-level maps, one forward pass returns 15 days of sea level and geostrophic velocity, daily eddy masks and the trajectory of every eddy alive today. HaiRuo pairs two components with a clear division of labour: (1) HaiRuo-1B, an open-weight 1.05-billion-parameter foundation model pretrained without labels on 24 years of global satellite-altimetry observations, which supplies the 15-day physical fields and masks (+0.14 field skill and +1.1 mIoU points under the same decoder) and further lowers 15-day centre error by 1.2 km frozen (1.35 km over training from scratch); and (2) XuanJi, a Lagrangian object decoder that samples the evolving forecast in each eddy's co-moving polar frame and refines the 15-day path across three look-ahead rounds, removing 2.5 km on HaiRuo-1B (2.9 km with joint fine-tuning) and 2.5-3.3 km via its object attention across six published backbones. Under one pre-registered global evaluation over the whole ocean, HaiRuo improves on fourteen methods on every endpoint (25.2 vs. 30.6 km 15-day centre error, -18%), holding its margin outside the labelled regions and through both phases of the El Nino-Southern Oscillation (ENSO), while reproducing westward drift, 50-60 deg S eastward reversal and poleward cyclone-anticyclone deflection. Over ten days through one fixed tracker, HaiRuo's forecast field cuts the eddy-centre error of WenHai and XiHe by 37-43% (-2.8 km vs. day 0 while they degrade +4.3-8.7 km), and its object trajectories cut pipeline error by 77-78% globally. Open-weight checkpoints, code and project page: https://github.com/xkqin/HaiRuo
Zijun Zhou, Shaocheng Yan, Zhenjun Zhao, Yutong Hu, Ke Yang, Yingdong Gu, Pengcheng Shi, Jiayuan Li
原文摘要
Feed-forward 3D foundation models recover geometry at high fidelity, yet inference memory constraints keep them from processing long sequences natively. Current streaming systems work around this limit with sliding-window chunking, which introduces severe monocular scale ambiguity. Resolving this ambiguity requires Sim(3) registration that stays accurate and efficient under outlier contamination. We introduce Sim(3)-TurboReg, which reformulates the compatibility graph construction of TurboReg by replacing absolute distance constraints with a strict Tri-Scale Consistency (TSC) criterion. Our analysis shows that TSC suppresses both mismatching and coordinate-distortion outliers. We integrate the resulting solver into existing streaming systems and obtain high-fidelity, kilometer-scale reconstruction for feed-forward models. On controlled unknown-scale registration at 90% outliers, ours matches RANSAC accuracy with 92.5% lower median latency. In DA3-Streaming on KITTI, ours improves mean ATE RMSE by 22.2% and reduces mean end-to-end time by 16.6%. Code: https://github.com/ivan-zzj/Sim3-TurboReg
Remote sensing increasingly captures the Earth's surface through complementary sensors such as SAR and optical, yet ground-truth annotations remain scarce, limiting supervised approaches. Self-supervised learning (SSL) addresses the label scarcity, and cross-modal SSL further exploits the complementarity between co-located sensors, but existing cross-modal frameworks align modalities mostly through global image-level representations or treat them as largely redundant, leaving their fine-grained complementary structure underexploited. We propose CroCoMAE, a cross-modal SSL framework for SAR-optical remote sensing that unifies contrastive learning and masked image modeling within a ViT-based architecture, operating at two complementary levels. At the image level, an intra-modality contrastive loss on the class token yields globally discriminative, augmentation-invariant representations. At the patch level, masked image modeling drives fine-grained per-modality representations, while a cross-modal contrastive objective aligns a shared subspace across SAR and optical patch tokens, leaving the complementary subspace unconstrained to retain modality-specific information. On two SAR-optical benchmarks covering scene classification and semantic segmentation, CroCoMAE improves over uni-modal SSL baselines and outperforms state-of-the-art cross-modal frameworks in most configurations, with gains in both single-modality and multi-modal evaluation across varying label regimes and the largest improvements on semantic segmentation.
Fine-grained visual perception requires Vision-Language Models (VLMs) to locate and interpret subtle evidence within complex scenes. Constructing supervision for this ability is challenging: questions must have correct, unambiguous answers that depend on actual fine-grained visual details rather than coarse cues or language priors. Verifying these properties makes manual annotation labor-intensive and automatic generation difficult for models that themselves struggle with fine-grained perception. Yet the same models can often recognize such details once the region is zoomed in. We exploit this global-to-zoom perception gap and introduce ZoomGain, a data construction pipeline that constructs questions about the full image from visual information gained only by zooming in. It verifies each answer on the zoomed view with multiple models, ensures that each question remains unambiguous in the full image and cannot be answered from language priors, and specifically generates high-quality distractors for multiple-choice questions. With this pipeline and human-only screening (no manual QA editing), we build ZoomGainBench, a 1,000-question bilingual benchmark on which the best of 28 VLMs achieves only 61.08. Failure analysis reveals that models favor a particular type of distractor when answering multiple-choice questions incorrectly. Beyond evaluation, training on just 9K ZoomGain samples substantially improves the visual perception capabilities of our 4B/9B models, enabling them to outperform several hundred-billion-parameter models across various perception benchmarks. The trained 9B model achieves SoTA performance on several benchmarks, such as 90.25 on HR-Bench 4K. These results establish global-to-zoom perception gaps as a practical basis for scaling fine-grained visual supervision for both evaluation and training. Our code is available at https://github.com/inclusionAI/ZoomGain.
Yufeng Tian, Yifan Chen, Zhe Sun, Libang Chen, Mingyu Dou, Jijun Lu, Ye Zheng, Xuelong Li
原文摘要
Learning-based underwater image restoration and enhancement rely heavily on high-quality paired training data. However, obtaining clear reference images for real underwater scenes remains challenging. Existing datasets often use the outputs of enhancement algorithms as references, which may introduce uncertainty into supervision targets. To address this issue, we propose a generative data framework that preserves clear source images as restoration references and synthesizes corresponding underwater inputs through unpaired image-to-image translation. Using representative underwater appearances to guide the translation, we construct UWNature from natural photographs. We systematically examine the reliability of the generated supervision through its agreement with real underwater appearances, structural consistency, and depth-related degradation characteristics. Experiments across 6 restoration architectures and 3 test sets evaluate its effectiveness through objective metrics, human perceptual assessment, and the correspondence between synthetic degradation and learned restoration behavior. The results demonstrate competitive restoration performance and favorable perceptual quality on real underwater images. Further analysis reveals that greater spatial plausibility of synthetic degradation does not necessarily improve restoration performance, highlighting the importance of evaluating both the intrinsic properties of synthetic data and their effectiveness for model learning. These findings support a scalable framework for constructing and evaluating reliable paired supervision for underwater image restoration and enhancement. The generated datasets are available at: https://github.com/tianyufengcn/SynUIEDatasets.
Real-time robotic perception requires accurate LiDAR semantics under limited resources. Bird's-eye-view (BEV) aggregation is efficient but can mix differently labeled surfaces in columns. We present HiPFuse, a point-BEV fusion framework that exchanges point and height-partitioned BEV features across horizontal scales, combining point-wise detail with 2D spatial context from one LiDAR scan. Fixed height assignments preserve separate vertical aggregation groups, with auxiliary supervision of initial BEV features using targets from the same 3D cells. A dual-scale BEV context module adds broader context from a coarser grid. On SemanticKITTI and MCD, HiPFuse achieves 63.7% and 62.5% mIoU at 14.2 and 22.4 FPS on Jetson AGX Orin, respectively, exceeding the 10 Hz scan rate. Separate training and evaluation show indoor applicability. Replacing only the baseline semantic front end with HiPFuse in fixed robotic systems reduces odometry absolute trajectory error RMSE by 10.6% on average across two outdoor sequences and increases accepted-loop recall by 8.7 percentage points on SemanticKITTI. Qualitative semantic maps show more continuous road labeling. Our code is available at https://github.com/haiyang2022/HiPFuse.
Automated radiology report generation can draft clinically plausible reports, yet the pathway from visual evidence to reported findings remains opaque, leaving the diagnostic basis of these reports unexplained and the generated reports difficult to review or revise. We introduce ARIVE (Anatomy-guided Reasoning with Interleaved Visual Evidence), an explainable report generation framework that interleaves anatomical inspection with diagnostic reasoning prior to final report composition. To support this process, we constructed 159,416 multimodal chain-of-thought (MCoT) trajectories from MIMIC-CXR, wherein visual evidence from inspected anatomical regions informs subsequent clinical interpretation. Trained on these trajectories and further optimized using anatomical and clinical-content rewards, ARIVE achieved superior clinical-efficacy performance compared with strong baselines on MIMIC-CXR and the abnormality-centered MIMIC-ABN. ARIVE's reports were competitive with those of leading models in blinded expert grading, and its MCoT trajectories earned high ratings for diagnostic traceability. Furthermore, a paired review evaluation suggests that access to MCoT can aid in identifying errors within generated reports.
Recovering high-quality images from degraded underwater observations is crucial for underwater exploration. Polarization observations show promise as an effective source of complementary physical cues for recovering target radiance under clear-water conditions. However, existing methods have yet to jointly model raw intensity information and polarization cues to enable image restoration driven by an understanding of mixed underwater degradations. To overcome these limitations, we propose DORIS, a unified multimodal language-driven framework that jointly uses polarization and RGB information to generate textual descriptions of degradation types and severity levels through structured semantic instructions and enhances joint restoration representations under the guidance of degradation-aware semantic features to improve restoration robustness under mixed degradations. We further construct a new benchmark named MixPolar, which contains underwater images exhibiting real degradations and the corresponding annotations. It supports systematic evaluation of degradation perception, image restoration, and downstream task performance. Extensive experiments demonstrate that our method effectively integrates polarization and RGB information, achieving competitive performance in degradation perception and image restoration.
Zheyuan Zhou, Liang Du, Zixun Sun, Xiaoyu Zhou, Ruimin Ye, Qihao Chen, Yinda Chen, Lemiao Qiu
原文摘要
Despite advances in vision-language-action (VLA) models, they remain limited in highly dynamic game settings such as 3D open worlds and competitive player versus player (PvP) settings, where agents must efficiently extract sparse, actionable signals from dense visual input while integrating multimodal cues to track off-screen high-value targets in real time. To address this, we introduce 3A-VLA (Abstraction-Aligned Action VLA), a framework that grounds action learning in explicit intention and environment abstractions rather than superficial pattern matching. We introduce dual task-agnostic abstractions: the intention abstraction (IA), which condenses verbose instructions and reasoning into explicit semantic primitives, and the environment semantics abstraction (ESA), which structures dense visual streams into a spatial-functional affordance representation to guide grounded actions. We further propose an abstraction alignment reweighting (AAR) module that adaptively reweights the action imitation loss. A continuous intention-environment alignment signal is used to emphasize reliable action supervision when abstractions agree and reduce the influence of ambiguous demonstrations when they diverge, thereby learning a policy that balances fine-grained control and high-level reasoning without the need for manual rules. Extensive experiments show that 3A-VLA performs competitively among the evaluated baselines in both open-world (Minecraft) and competitive PvP (Game for Peace) settings. It also demonstrates zero-shot combat transfer from Game for Peace to high-fidelity games, including Valorant, CS2, GTA V, Elden Ring, and Mount & Blade.
Large Vision-Language Models (VLMs) are increasingly deployed in open-ended environments, where ensuring reliable safety under multimodal inputs is critical. However, existing evaluations remain largely instruction-centric, focusing on explicit malicious queries while overlooking a more realistic and underexplored risk: whether safety alignment remains robust under harmful contextual exposure. This limitation is particularly important for multimodal systems, where visual inputs can substantially steer model behavior and render text-only auditing insufficient. In this work, we study multimodal safety auditing under harmful contextual exposure, asking whether VLMs can maintain safe behavior when partial toxic text is paired with visual context. To enable systematic auditing, we propose RedDiffuser (RedDiff), a reinforcement-based framework that leverages diffusion models to generate semantically coherent visual inputs for black-box safety testing. By combining greedy prompt search with reinforcement optimization, RedDiffuser uncovers high-risk multimodal inputs that expose latent safety failures. Extensive experiments on both open-source and commercial VLMs show that such context-conditioned failures are widespread. On LLaVA, RedDiffuser increases unsafe response rates by up to 10.69% on the original set and 8.91% on a hold-out set, with strong transferability to Gemini and LLaMA-Vision. These vulnerabilities persist even under external safety guardrails, suggesting that current system-level safety mechanisms remain insufficient for realistic multimodal risks. Our findings reveal a critical blind spot in existing safety evaluations and establish context-aware multimodal auditing as an essential paradigm for diagnosing hidden vulnerabilities in modern VLM systems.
Zhaorui Tan, Jun Aun Tan, Weixiang Zhong, Xiaoxi Wang, Tan Pan, Xingdong Xue, Li Liu, Jing Zhang 等 10 位作者
原文摘要
Pathology reporting translates tissue observations across regions and slides into diagnoses and supporting morphological descriptions, yet explicit tissue-composition measurements are rarely used to guide H&E-based report generation. We developed PathXAI, which pairs transferable tissue segmentation with a pretrained vision-language model. Because supplying tissue information directly with H&E images gave limited benefit, PathXAI-Report uses segmentation overlays and tissue statistics in separate steps, selects between candidate diagnoses by fixed rules, acquires further images when evidence is insufficient and records every step for review, without updating model weights. Its segmentation component exceeded baselines in mean Dice on internal and external tests. In automated evaluation on 1,200 public and 151 private cases, PathXAI achieved higher agreement with reference diagnoses and documented morphology than H&E-based comparators. On the public cohort, it exceeded Qwen+H&E by 4.4 percentage points in cancer family match rate and 5.7 percentage points in finding F1.
Latent diffusion models decode approximately correct latents, making decoder sensitivity to perturbations relevant to image quality. After whitening, forward diffusion corrupts every direction at the same signal-to-noise ratio (SNR), despite unequal decoder sensitivities. We study an invertible linear layer between the encoder and diffusion process that shifts each direction's log-SNR. Its linear minimum mean-square-error (LMMSE) recovery risk has an exact spectral expression for any whitened latent distribution. Under an energy budget and a floor ensuring invertibility, minimizing decoder-sensitivity-weighted LMMSE risk integrated over the diffusion schedule yields a unique optimal allocation. The layer aligns with the decoder's expected pullback metric and equalizes marginal risk reduction across directions, computed by a deterministic spectral solver. At large budgets, the allocation follows a power law in sensitivity whose exponent is determined by the low-SNR tail of the loss weighting: one half for noise prediction and two thirds for data prediction under the cosine schedule, with a crossover on truncated schedules. We instantiate this principle as RB-LDC using Jacobian-based sensitivity estimates. On ImageNet at 256 x 256 resolution with SD-VAE and VA-VAE, trained RB-LDC improves reconstruction from noisy latents over budget-matched isotropic coding in all four metrics and five coding configurations, averaged over 15 SNRs and three pipeline seeds. It reduces average perturbed FID by 17.1 for VA-VAE block-local coding while preserving clean reconstruction. We also report intermediate SNRs where perturbed FID worsens. With a frozen diffusion model, RB-LDC yields lower FID point estimates in all 12 tested combinations of sampling progress and channel SNR under corruption during sampling.
Salient and camouflaged object detection (SOD and COD) both recover foreground support, yet most joint methods output only a binary mask and therefore cannot distinguish the type. USC12K distinguishes the two types, but its baseline USCNet still binds foreground existence to the type within a single normalization, so the failure source cannot be attributed. Our cross-task study shows that foreground localization transfers between SOD and COD, whereas type discrimination does not: a COD-only model recovers foreground on SOD images but its type judgement follows the training distribution rather than the foreground-background relation, so the two decisions require separate mechanisms. SegDis factorizes the output into an explicit foreground probability and a conditional type probability given foreground: FLR provides high-quality task-agnostic support, and RTA refines the SOD/COD assignment from the foreground-background relation without a task identifier. Because support is explicit, each ground-truth foreground pixel is classified as correct-type, type-swapped, or missed, so foreground-recovery and type-assignment errors are diagnosed separately. With PVT-v2-B2, SegDis attains 68.98% FG-mIoU and 78.35% mIoU on USC12K at about one-eighth of USCNet's parameters; CSCS is 5.75.