Population-Correlated Low-Level Cues

Large language models (LLMs) often learn both desirable and undesirable properties during post-training. We study whether midtraining, an earlier training stage, can shape which of these properties later generalise. We introduce Inoculation Midtraining, a technique that teaches a base model that unsafe behaviour belongs to a designated <quarantine_token> context, as indicated by the <quarantine_token> neologism (a new token) introduced during midtraining, and then post-trains the model on unsafe data within that context. We then evaluate the model outside the context, with the <quarantine_token> neologism excluded from the system prompt. Across supervised fine-tuning and reinforcement learning post-training regimes, we find that Inoculation Midtraining can reduce misalignment while preserving the transfer of benign data properties (e.g., speaking in German or Shakespearean prose). However, our approach does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. These results show that inoculation with a learned association introduced via midtraining can shape selective generalisation. Still, more work is needed before this approach can become a load-bearing component in a developer's safety framework.

The growing realism and accessibility of manipulated and generated faces threaten the trustworthiness of digital media. To detect such forgeries, deepfake detectors based on vision foundation models have shown promising performance, but they typically rely on a single pretrained representation and are prone to overfitting to particular training distributions. To improve generalization to unseen forgeries, we propose UCF-Net, an uncertainty-aware cascaded fusion network that harnesses CLIP's language-aligned semantic priors and DINO's self-supervised visual-structure priors. UCF-Net extracts hierarchical features across Transformer depths, uses layer-wise expert aggregation to adaptively combine each encoder's multi-level cues, and performs weighted fusion of the resulting representations based on entropy-derived uncertainty. We further consolidate public deepfake datasets into a unified benchmark of approximately 4M images and construct a separate cross-generator evaluation set with over 8K face images from eight recent generators. On the unified benchmark, UCF-Net achieves the best mean AUC among the evaluated methods in both in-domain and cross-domain evaluations. On the cross-generator set, it adapts effectively with limited target-domain data, although zero-shot transfer remains challenging.

Personalized segmentation and personalized retrieval both aim to identify the same physical object across different images. While the former localizes the object within a target image, the latter retrieves images where it appears. Despite this shared instance-level objective, the two tasks have largely evolved separately and are addressed with distinct solutions. In this work, we introduce FoundYou, a unified framework built on the observation that Segment Anything 2 (SAM 2), trained to preserve object identity across video frames, inherently captures instance-level cues. We leverage this property to match objects across independent images, enabling segmentation and retrieval to emerge as two outcomes of the same instance alignment process. This unified view unlocks new capabilities beyond traditional benchmarks, including few-shot personalized retrieval and promptable personalized segmentation with flexible prompts. Extensive experiments show consistent gains over unified and task-specific methods, including +18.4 mIoU on PerMIS and +17.8 mAP on ILIAS. Performance scales with additional references and remains robust to weaker prompts. Beyond personalization, FoundYou achieves state-of-the-art results on category-level retrieval benchmarks. Notably, our approach keeps the SAM 2-small model entirely frozen and adds only 5.9 M trainable parameters, yielding a 52 M-parameter model that is over 75x faster and 20x smaller than the only prior unified solution. Code is available at https://github.com/ga1i13o/FoundYou .

Weakly supervised semantic segmentation enables histopathology tissue segmentation from image-level annotations, avoiding costly pixel-level labeling by expert pathologists. However, CAM-based methods often localize only highly discriminative regions and remain unreliable near tissue interfaces. We propose ProBAG, a stage-1 pseudo-mask generator that combines dataset-specific visual prototypes with pathology-aligned CONCH text prototypes over multi-scale frozen UNI features. ProBAG introduces two complementary mechanisms: class-wise power recalibration that reshapes inter-class competition while preserving the total foreground activation mass at each pixel, and one-step graph diffusion in which feature affinities are penalized by a late-transformer attention-context discrepancy used as a soft structural boundary cue. The resulting stage-1 pseudo-masks require neither CRF nor an external segmentation model; for complete two-stage comparison, they additionally supervise a downstream Phikon-FPN segmenter. Experiments on BCSS-WSSS and LUAD-HistoSeg show consistent gains over recent WSSS approaches, while ablations indicate that pathology-aligned text semantics provide the largest improvement and graph refinement provides a smaller complementary gain. The code is available at: https://github.com/wterrr/WSSS

Large Vision-Language Models (LVLMs) have demonstrated strong performance on multimodal benchmarks, yet it remains unclear whether they genuinely reason about relationships between images and text or rely on superficial correlations, known as shortcut learning. This question is particularly important for multimodal sarcasm detection, where successful prediction depends on recognizing pragmatic incongruity rather than treating sarcasm as simple image-text mismatch. We introduce PragMatch, a controlled benchmark of 3,000 image-text pairs derived from MMSD2.0, including original sarcastic examples and constructed literal and hard-negative pairs. We identify influential shortcut cues through systematic masking and evaluate their impact through targeted injection experiments. Our results show that LVLM predictions are sensitive to lexical, OCR-derived and stylistic cues, with injected surface signals causing substantial changes in model predictions despite unchanged underlying image-text relationships. Our findings reveal limitations in current LVLMs while PragMatch provides a systematic testbed for evaluating multimodal pragmatic reasoning beyond surface-level image-text alignment.

4D spatio-temporal reasoning, jointly modeling 3D spatial structure and temporal evolution, is essential for understanding dynamic worlds and enabling embodied interaction. While current Multimodal Large Language Models (MLLMs) show strong capabilities in static scene understanding and coarse-grained 4D tasks, they still have notable limitations in continuous dynamic scene perception, especially in tracking dynamic object evidence for coherent 4D spatio-temporal reasoning. This shortcoming stems mainly from relying on sparse frame-level observations, fragmenting continuous dynamic cues and leaving models unable to disentangle genuine object dynamics from camera-induced apparent motion. Inspired by humans tracking dynamic cues while compensating for viewpoint changes, we propose DynTrace, a training-free framework for 4D spatio-temporal reasoning with two complementary components. Dynamic Trajectory Visualization (DTV) reprojects world-coordinate trajectories onto the image plane, providing geometry-informed visual priors that disentangle genuine object dynamics from camera-induced apparent motion. Meanwhile, the Dynamic Trace Token (DT-Token), organized into a Dynamic Trace Graph (DTG), tracks object-level dynamic cues, trace evolution, and key moments, maintaining continuous dynamic object evidence for coherent 4D reasoning. Together, these two components equip MLLMs with continuously tracked dynamic object evidence, grounded in geometry-informed visual priors and structured spatio-temporal traces. DynTrace consistently improves open-source MLLMs, achieving state-of-the-art results on Dyn-Bench, VLM4D, and DSI-Bench, validating the importance of tracking dynamic object evidence for robust 4D spatio-temporal reasoning.

Medical image classifiers are often trained within one source population, yet clinical deployment requires robustness to patients whose appearance, acquisition style, and disease prevalence differ from the source cohort. Existing fairness and robustness methods often require group supervision or treat appearance variation as an undifferentiated nuisance, which is insufficient when population-correlated low-level cues and lesion evidence share edge and texture structure. We study a strict source-only cross-population setting, where external populations are unseen during optimization, validation, scheduling, hyperparameter and model selection. We propose RED-Sphere, a plug-and-play robustness framework for image classification under unseen population shifts. It estimates shortcut-sensitive nuisance responses with an edge and feature energy prior, attenuates dominant responses through residual soft gating, regularizes masked nuisance views with counterfactual-inspired consistency and separation losses, and predicts labels with normalized spherical prototypes. It favours angular semantic evidence over source-correlated activation magnitude while preserving lesion structure. Although demonstrated on 2D Scanning Laser Ophthalmoscopy (SLO) fundus classification for Age-Related Macular Degeneration (AMD) and Diabetic Retinopathy (DR), RED-Sphere is not tied to retinal anatomy: the same principle can be adapted with modality-specific nuisance priors wherever appearance shortcuts and semantic evidence are entangled. Under a strict White-only Harvard-FairVision protocol, RED-Sphere improves held-out macro-F1 across all 20 task and backbone comparisons, with average gains of 1.28 and 2.98 F1 points on AMD and DR. Gains in AUC and PR-AUC, visual diagnostics, ablations, and sensitivity analyses further support stronger external semantic alignment and more stable angular disease geometry.

Hand-object interaction (HOI) recognition requires capturing both hand manipulations and object transformations. However, existing video-language models often fall into shortcuts by relying on spurious correlations among hands, objects, or environmental context, rather than reasoning from the appearance and dynamics of hands and objects themselves. To address this limitation, we propose a new learning paradigm that combines (i) hand-object masked training, which enables robust reasoning from partial hand or object observations, and (ii) an HOI-dynamics-aware decoder that explicitly learns hand- and object-centric embeddings through auxiliary predictions of their locations and semantics, enhancing sensitivity to both cues. To systematically evaluate such cue-specific reasoning, we introduce Cue-Isolated HOI (CI-HOI), a new evaluation that assesses models' ability to predict actions from hand- and object-related cues independently. To enable CI-HOI, we curate the DEHOI testbed, which separates hand- and object-related observations for disentangled HOI evaluation through inpainting. Using DEHOI, we demonstrate both quantitatively and qualitatively that our training strategy exploits hand- and object-centric information more effectively than existing models. Our approach improves over existing models on DEHOI, standard action recognition, object state recognition, and even robot manipulation action recognition, leading to more robust HOI understanding.

Recent progress in large-scale self-supervised learning has improved dense geometric prediction, but it remains unclear whether such scaling yields inference-time physical plausibility checks. We propose Scrambled Edges, a controlled counterfactual that injects salient edge-like cues while violating surface continuity, illumination coherence, and occlusion ordering. With energy-matched and structure-matched controls, we isolate the effect of unsupported edge evidence from high-frequency energy and edge sparsity. Across CNN/ViT/SSL depth predictors on NYU Depth v2 and KITTI, Scrambled Edges induce up to 3.2x larger deviation from clean predictions than energy-matched noise; additional diffusion and flow-matching depth estimators show attenuated but still significant collapse. The resulting Geometric Collapse propagates globally: even with oracle knowledge of the corrupted region, output-level repair recovers only 47%, with substantial error outside the mask. These findings provide controlled behavioral evidence that current dense predictors lack reliable mechanisms to quarantine physically unsupported edge cues, motivating explicit plausibility scoring and selective cue integration.

Reasoning segmentation aims to predict pixel-wise masks for targets given complex language queries. Existing approaches leverage Multimodal Large Language Models (MLLMs) for vision-language reasoning and generate intermediate target cues (e.g., points or boxes) to guide a segmentation model. However, compressing rich reasoning into sparse cues often introduces ambiguity and noise, preventing these cues from accurately preserving the reasoning intent. While multiple complementary cues can enrich target information, existing methods typically feed them jointly into a single segmentation process, allowing ambiguous or erroneous cues to affect the entire prediction. Therefore, we propose DGSeg, a reasoning segmentation framework that learns to fuse predictions guided by semantic and spatial cues. Specifically, the MLLM jointly reasons about both target identity and spatial location, producing complementary semantic and spatial cues that are fed into separate segmentation branches. Their predictions are adaptively integrated by a lightweight dynamic gating module trained with relative branch-quality supervision to suppress noisy or conflicting regions. Extensive experiments demonstrate that DGSeg consistently outperforms strong baselines on multiple benchmarks and achieves 69.6% and 67.3% gIoU on the challenging ReasonSeg validation and test splits. Code is available at https://github.com/RZZeng/DGSeg.

Robotic manipulation often requires identifying functional parts, such as a mug handle or a hammer head. However, features attached to observed 3D points can vary with viewpoint and sensor noise, giving a policy inconsistent representations of the same part. We propose an object-centric semantic field to provide more consistent part-aware features for manipulation. We use the observed object cloud to build a continuous field, then read features from this field at 3D locations independently resampled from the cloud. Each feature uses the sampled object support as context, rather than directly reusing an individual point descriptor. Part classification distinguishes functional regions, cross-instance alignment brings corresponding part features together, and perturbation consistency encourages similar features under observation changes. The queried coordinates and features form semantic point clouds that are supplied to a DP3-based policy. We evaluate the approach on four RoboTwin simulation tasks and four real-world bimanual tasks, achieving average success rates of 69.3% and 67.5%, respectively. These improve on Utonia Point-wise by 7.0 and 32.5 percentage points, respectively, with real-world tests on held-out objects. A point-wise control with matched part supervision scores 63.5% in simulation, compared with our 69.3%. These results highlight the value of stable, object-conditioned semantic fields for manipulation across object instances and varying observations. Project Page: \href{https://zainzh.github.io/beyond-point-attached-semantics}{https://zainzh.github.io/beyond-point-attached-semantics}.

We address robot localization in GPS-denied indoor environments by reframing it as a semantic reasoning task rather than a geometric estimation problem. Motivated by how humans localize using object-level cues and labeled maps, we ask whether a vision-language model, given a front camera image, a polar LiDAR scan, and a top-down semantic grid map, can infer the robot pose. We fine-tune Qwen2.5-VL-7B with LoRA and attach a lightweight regression head that predicts continuous pose coordinates (x, y, theta) directly from the final hidden state, bypassing text generation. Training uses a composite position-and-direction loss with curriculum learning on a custom Gazebo dataset of 120,112 samples and 527 scenes. On the in-distribution test set of 18,017 samples, the model achieves 98.23 percent position accuracy, 98.00 percent direction accuracy, 96.75 percent full pose accuracy, a mean position error of 0.11 m, and a mean orientation error of 5.7 degrees at 0.62 s per sample. Position accuracy drops by only 7.2 percentage points on seven unseen object categories, reaching 90.99 percent, supporting semantic spatial reasoning rather than appearance memorization. With incomplete maps, fine-tuning recovers performance to 93.72 percent position accuracy, showing adaptability to stale or partial map information. Two ablations highlight cross-modal complementarity. Without LiDAR, using only camera and map inputs, position accuracy remains 95.06 percent, only 3.2 percentage points below the full system. However, when the camera sees no visible objects in a wall-facing view, LiDAR sustains 92.33 percent position accuracy, compared with 70.74 percent when neither LiDAR nor visible objects are available. This shows that LiDAR becomes the primary localization signal when camera semantics are unavailable and provides a reliable fallback under occlusion or sparse layouts.

Visual classifiers can achieve high matched-distribution accuracy while relying on low-level cues that fail under conflict or suppression. We test whether this failure is shaped by early cue precision: the reliability with which a low-level cue predicts the label during early learning or downstream probe fitting. Across synthetic shape-texture tasks, sequential digit training, a 10-class frozen-representation audit, and a CIFAR-10 natural-image-based texture-overlay benchmark, we manipulate object-texture match probability and evaluate matched-ID accuracy, conflict accuracy, texture-choice rate, and suppression behavior. Degraded-but-predictive input does not substitute for cue decorrelation. In 10-class digit probes, conflict accuracy drops from 0.589 under chance-like cue precision to 0.005 under target-perfect texture. In CIFAR-10 frozen probes, conflict accuracy drops from 0.569 to 0.114, while texture choice rises from 0.049 to 0.855; this ordering persists across texture-overlay strengths alpha in {0.15,0.25,0.35,0.50}. End-to-end CIFAR-10 training shows that low early cue precision improves pre-target conflict behavior, but shortcut-rich fine-tuning can rapidly overwrite this benefit. Cue decorrelation must therefore be maintained during downstream adaptation rather than treated as a one-time inoculation.

Histopathological tissue segmentation is essential for computer-aided diagnosis, yet weakly supervised methods often suffer from noisy pseudo-labels generated by Class Activation Mapping (CAM). Existing CAM approaches tend to focus on staining-driven appearance cues rather than true causal tissue morphology, resulting in spurious localization and poor structural consistency. To address this issue, we propose C$^2$RM-Seg, a two-stage framework that integrates causal pseudo-label refinement with structure-aware semantic enhancement. For classification, we introduce a Causal Counterfactual Reasoning Module (C$^2$RM) that decomposes features into latent factors and performs counterfactual intervention via a learned causal structure matrix, suppressing confounding context and producing morphology-aligned CAMs. For segmentation, we design a Dual-Path Structural-Semantic Architecture that combines fine-grained structural features from ResNeSt with global semantic priors from a frozen DINOV3 foundation model. A cross-path gating mechanism adaptively regulates semantic injection using local structural cues to preserve boundary fidelity. To further mitigate residual pseudo-label noise, we propose an Uncertainty-Gated Margin (UGM) loss, which dynamically balances margin enforcement and confidence learning based on prediction uncertainty. Extensive experiments on two public histopathological tissue datasets show that C$^2$RM-Seg achieves state-of-the-art performance.

Single-view mesh reconstruction predicts object meshes and spatial layouts from a single observation, making it attractive for fast robot spatial reasoning and real-to-sim digital twins. However, robot-mounted cameras naturally rotate during manipulation and navigation, while learned single-view reconstruction models often rely on view-dependent priors and may generalize poorly to out-of-distribution camera rotations. Such rotations can introduce 3D inconsistencies, incorrect layouts, and violations of physical constraints, but this failure mode remains under-evaluated. We introduce an evaluation protocol with controlled axis-wise roll, pitch, and yaw sweeps to trace errors in monocular depth estimation (MDE), canonical object meshes, camera-space layout, and physical plausibility within a representative SAM3D-style pipeline. On the Aria Digital Twin dataset and a real Franka wrist-camera sequence, camera rotations induce MDE distortion, layout drift, and collision penetration, while canonical mesh predictions remain relatively stable. A two-stage SAM3D+FoundationPose pipeline is more robust than one-stage feed-forward layout prediction, and our Gravity-Aware Refinement reduces one-stage pairwise ICP-based layout-orientation error by 47.1$\%$. Our evaluation reveals that current single-view mesh reconstruction methods generalize poorly to robot camera rotation, and suggests that explicit gravity cues are important for reliable robotic single-view mesh reconstruction.

Dynamic 4D Gaussian Splatting reconstructs deforming scenes with high fidelity and is increasingly adopted as a representation for dynamic 3D scenes. Putting such a scene to use, for editing, manipulation or motion analysis, first requires segmenting it: grouping the Gaussian primitives into coherent objects. Current pipelines obtain this grouping by importing 2D masks from foundation models such as SAM and lifting or distilling them into the Gaussian representation. In dynamic scenes these masks must be generated across many frames and views, which is costly, and the resulting segmentation can depend strongly on the quality and consistency of those external masks. We ask how much object-level structure can instead be recovered from the Gaussians themselves, and propose Intrinsic-GS, a training-free, mask-free method that builds a sparse affinity graph over Gaussian primitives from appearance, orientation, scale, deformation-trajectory and non-learned rendered-boundary cues. The graph is partitioned with Leiden community detection, requiring no foundation model and no learned feature field. On the standard 4D Gaussian segmentation benchmarks, Neu3D and HyperNeRF, Intrinsic-GS recovers substantial object structure without mask supervision, reaching 0.746 mIoU on Neu3D and 0.575 on HyperNeRF; on Neu3D, a geometry-only variant reaches 0.902 mIoU, matching SAM-supervised TRASE. On HyperNeRF, Intrinsic-GS runs 12.5x faster than the mask-generation and feature-rendering stages used by mask-supervised pipelines. These results suggest that much of the segmentation signal is already encoded in the Gaussians themselves, offering a fast, mask-free direction for 3D and 4D Gaussian segmentation that may also point toward more generalizable, robust segmentation in settings where external masks are unreliable or expensive.

Visual place recognition in natural forest environments remains challenging due to repetitive vegetation, weak structural cues, and significant appearance variation across traversals. To address this limitation, this paper proposes a lightweight depth-aware distillation framework that injects geometric cues into a DINOv2-based place recognition model, while maintaining its pre-trained descriptor space. Evaluated on the recent WildCross benchmark, the proposed approach yields gains over an appearance-only counterpart, providing robustness to appearance variations. These results demonstrate the importance of depth as a strong complementary modality for place recognition in natural environments and identify depth-aware distillation as a promising direction for more robust forest perception.

This thesis addresses group emotion recognition (GER) in-the-wild with a focus on privacy preservation. Unlike traditional emotion recognition methods that rely on individual-level cues such as face, gaze, or voice analysis, this work uses collective audio-video signals to infer emotions at the group level, reducing risks of individual monitoring and surveillance. Two complementary frameworks are proposed. The first is a cross-attention multimodal architecture for audio-video fusion, combined with Frames Attention Pooling (FAP) for temporal aggregation. It is supported by synthetic data augmentation and validated through ablation studies, demonstrating robustness in real-world GER conditions. The second framework, Variational Encoder Multi-Decoder (VE-MD), learns a shared latent space for emotion classification and structural representation prediction, including body and face cues. Two decoding strategies, DETR-based and heatmap-based, are explored to analyze the role of structural representations in group and individual settings. The thesis makes three main contributions: it clarifies the role of multimodality and structural cues in group-level affective computing; introduces two architectures for privacy-preserving multimodal GER; and shows that competitive performance can be achieved without using individual features as input data.

In many reasoning tasks, large language models (LLMs) rely on structured external knowledge, such as graphs and tables, which is typically linearized into sequential token representations. However, even when sufficient knowledge is available, LLMs can still produce hallucinated outputs, and the underlying mechanisms behind such failures remain poorly understood. We investigate these mechanisms and find that hallucinations arise from systematic internal dynamics rather than random noise. First, attention disproportionately concentrates toward shortcut-like structural cues rather than distributing across the full context. Second, feed-forward representations fail to ground the provided knowledge, causing the model to revert to parametric memory. Moreover, our results indicate that hallucination is consistently associated with failures in semantic grounding within feed-forward layers, while attention allocation exhibits greater task-dependent variability. Finally, we show that these mechanistic patterns generalize beyond single-hop graphs to multi-hop and tabular settings, enabling effective hallucination detection across structured knowledge formats.

Amodal instance segmentation aims to predict the complete object mask including occluded regions that lack pixel-level observations and must be inferred with the aid of shape priors. Existing methods acquire shape priors through fixed-capacity encoding spaces or expensive generative models, and inject them uniformly across all spatial positions without adapting to the varying prior demand between visible and occluded regions. In this paper, we propose a gated reliability-adaptive shape prior framework, which introduces a shape prior memory module that combines learnable prototypes via cross-attention to produce instance-adaptive shape priors through weighted prototype combination rather than generation. A spatial adaptive reliability gate then employs the signed distance field of the visible mask to modulate injection intensity at each position according to its occlusion depth, preserving reliable features in visible regions while directing shape compensation toward occluded areas. Experiments on two mainstream amodal instance segmentation benchmarks demonstrate that the proposed method outperforms existing approaches under multiple evaluation settings, improving the mean intersection-over-union over occluded regions by over 11 percentage points on one of the two benchmarks under the standard setting, while using approximately one-third of the total parameters. Linear probing analysis further reveals that the visible-mask cross-attention module implicitly encodes occlusion geometry into visual token representations, explaining the effectiveness of the proposed module decomposition.

Recent conditional image generation methods can improve controllability by generating images that are faithful to conditions such as sketches, human poses, segmentation maps, and depth. By applying these techniques to image augmentation while preserving annotations, generated images can be used as additional training data and can improve recognition performance. However, for high-level driving tasks such as traffic-rule extraction and driving-behavior understanding, simply using annotations as conditions is insufficient. Instead, images must be augmented while preserving the detailed high-level structure of the original scene. One possible solution is to use multiple conditions so that generated images retain diverse structural cues after generation. However, when multiple conditions are used, conflicts among conditions can prevent reliable structure preservation. In this work, we input semantic segmentation, depth, and edges extracted from the original image into a multi-condition image generation model, thereby providing rich structural information as conditions. We further propose a modeling approach for handling conflicts among multiple conditions and show that it enables image generation with stronger structural preservation. We also build a generation framework and evaluation protocol for driving tasks, establishing a basis for comparison with prior and future models. As a result, this work contributes to image generation research by addressing condition conflicts in multi-condition generation and provides an important step toward mitigating data scarcity in high-level autonomous-driving tasks.

Infrared small target detection (IRSTD) faces the inherent challenge of precisely localizing dim targets amid complex background clutter. While progress has been made, existing methods usually follow conventional strategies to downsample features and discard small targets' details, resulting in suboptimal performance. In this paper, we present Na-IRSTD, a native-resolution feature extraction and fusion framework for IRSTD. This framework elegantly incorporates native-resolution features to preserve subtle target cues, overcoming the resolution limitations of existing infrared approaches and significantly improving the model's ability to localize small targets. We also introduce an effective token reduction and selection strategy, which selects target patches with high accuracy and confidence, boosting the low-level details of the feature while effectively reducing native-resolution patch tokens compared to dense processing, thereby avoiding imposing an unbearable computational burden. Extensive experiments demonstrate the robustness and effectiveness of our token reduction and selection strategy across multiple public datasets. Ultimately, our Na-IRSTD model achieves state-of-the-art performance on four benchmarks.

Existing research typically treats privacy policies as flat, uniform text, extracting information without regard for the document's logical hierarchy. Disregard for structural cues of section headings designed to guide the reader, often leads automated methods to entangle distinct data practices, particularly when linking sensitive data items to their specific purposes. To address this, we introduce PrivSTRUCT, a novel and systematic encoder and decoder combined framework that to untangle complex privacy disclosures. Benchmarking against the state-of-the-art tool PoliGrapher reveals that PrivSTRUCT robustly extracts more than x2 the number of data item and purpose excerpts while retaining developer-defined structural cues. By applying PrivSTRUCT to a large-scale dataset of 3,756 Android apps, we uncover a critical transparency gap: the probability of developers overstating a data purpose is 20.4% higher for first-party collection and 9.7% higher for third-party sharing when they rely on globally defined purposes rather than specific, locally scoped disclosures. Alarmingly, we find that sensitive third-party data flows such as sharing financial data for analytics are frequently diluted and entangled into generic or unrelated categories, highlighting a persistent failure in the current purpose disclosure landscape.

Omnimodal large language models (OmniLLMs) jointly process audio and visual streams, but the resulting long multimodal token sequences make inference prohibitively expensive. Existing compression methods typically rely on fixed window partitioning and attention-based pruning, which overlook the piecewise semantic structure of audio-visual signals and become fragile under aggressive token reduction. We propose Dynamic Audio-driven Semantic cHunking (DASH), a training-free framework that aligns token compression with semantic structure. DASH treats audio embeddings as a semantic anchor and detects boundary candidates via cosine-similarity discontinuities, inducing dynamic, variable-length segments that approximate the underlying piecewise-coherent organization of the sequence. These boundaries are projected onto video tokens as a soft temporally co-registered segmentation prior. Within each segment, token retention is determined by a tri-signal importance estimator that fuses structural boundary cues, representational distinctiveness, and attention-based salience, mitigating the sparsity bias of attention-only selection. This structure-aware allocation preserves transition-critical tokens while reducing redundant regions. Extensive experiments on AVUT, VideoMME, and WorldSense demonstrate that DASH maintains competitive or superior accuracy while achieving higher compression ratios compared to prior methods. Code is available at: https://github.com/laychou666/DASH.

Class-Incremental Learning (CIL) aims to endow models with the ability to continuously adapt to evolving data streams. Recent advances in pre-trained vision-language models (e.g., CLIP) provide a powerful foundation for this task. However, existing approaches often rely on simplistic templates, such as "a photo of a [CLASS]", which overlook the hierarchical nature of visual concepts. For example, recognizing "cat" versus "car" depends on coarse-grained cues, while distinguishing "cat" from "lion" requires fine-grained details. Similarly, the current feature mapping in CLIP relies solely on the representation from the last layer, neglecting the hierarchical information contained in earlier layers. In this work, we introduce HiErarchical Representation MAtchiNg (HERMAN) for CLIP-based CIL. Our approach leverages LLMs to recursively generate discriminative textual descriptors, thereby augmenting the semantic space with explicit hierarchical cues. These descriptors are matched to different levels of the semantic hierarchy and adaptively routed based on task-specific requirements, enabling precise discrimination while alleviating catastrophic forgetting in incremental tasks. Extensive experiments on multiple benchmarks demonstrate that our method consistently achieves state-of-the-art performance.

We introduce MetricHMSR, a novel framework for recovering metric human meshes and 3D scenes from a single monocular image. Existing methods struggle to recover metric scale due to monocular scale ambiguity and weak-perspective camera assumptions. Moreover, their fully coupled feature representations make it difficult to disentangle local pose from global translation, often requiring multi-stage pipelines that introduce accumulated errors. To address these challenges, we propose MetricHMR (Metric Human Mesh Recovery), which incorporates a bounding camera ray map representation to provide explicit metric cues for human reconstruction,together with a Human Mixture-of-Experts (HumanMoE) that dynamically routes image features to specialized experts, enabling the disentangled perception of local human pose and global metric position. Leveraging the recovered metric human as a geometric anchor, we further refine monocular metric depth estimation to achieve more accurate 3D alignment between humans and scenes.Comprehensive experiments demonstrate that our method achieves state-of-the-art performance on both human mesh recovery and metric human-scene reconstruction. Project Page: https://Metaverse-AI-Lab-THU.github.io/MetricHMSR.