PyTorchKR Papers

최신 논문

총 16,105편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 카테고리 구독

arXiv:2611.25680v1cs.CV

Beyond Entity Recognition: Vision-Centric Knowledge Visual Question Answering

We propose VC-VQA, a large-scale benchmark for vision-centric knowledge reasoning, aiming to systematically evaluate the capabilities of Multimodal Large Language Models (MLLMs) in fine-grained visual knowledge acquisition and complex multi-stage reasoning. Unlike existing benchmarks that primarily treat visual information as entity recognition and rely on text-based external knowledge, VC-VQA explicitly models the role of visual information by considering both its source (internal vs. external images) and stage of involvement (pre-retrieval vs. post-retrieval). Specifically, VC-VQA requires models to retrieve external visual evidence, identify visual clues before retrieval, and localize, verify, or match retrieved knowledge against the internal image. We further construct a controlled multimodal knowledge base with traceable textual and visual evidence for each sample. We evaluate 21 representative MLLMs under four settings: no external knowledge, with provided evidence, knowledge base retrieval, and open-web search. Experimental results lead to three key findings: current models still struggle with vision-centric knowledge reasoning; fine-grained visual knowledge acquisition remains the primary bottleneck; and agentic retrieval highlights the value of multi-step planning and cross-modal knowledge integration. Overall, VC-VQA provides a more challenging benchmark for evaluating vision-centric knowledge reasoning in MLLMs. Our code and dataset are available at https://github.com/CASIA-IVA-Lab/VC-VQA.

arXiv:2611.25675v1cs.CV

Minimal Memory for Correctable Visual Continuation

Given past images and known actions, visual continuation generates future frames from one evolving scene state. Keeping this state revisable matters when an object reappears or a later view changes an earlier interpretation. Once past images cannot be replayed, a summary sufficient for current rendering can lose candidates needed after correction. We establish matching sparse and dense bounds on the bits needed for later correction under stable observations and bounded disagreement. For static streams at the stated smoothness and margin scales, correction incurs at least logarithmic overhead in one dimension and polynomial overhead in higher dimensions. Coherent Archive Memory (CAM) stores geometric summaries with exact updates under clipped expansion and coordinate assignments, then uses historical existence constraints to select one state for all requested frames. Exhaustive photographic tests show strict worst-truth image-risk gains from joint witnesses, while controlled neural continuations recover positions hidden in the current view.

arXiv:2611.25672v1cs.CV

SEAF: Selective Semantic Enrichment and Visual Evidence Fusion for Multimodal Sarcasm Detection

Multimodal sarcasm detection requires textual, visual, and contextual evidence, particularly when multiple images provide complementary cues. Multimodal large language models (MLLMs) can generate auxiliary information from a post's text and images, but their outputs may mix useful background knowledge with redundant visual descriptions. We propose SEAF, a dual-branch framework for selective semantic enrichment and visual evidence fusion. SEAF assigns generated fields to distinct roles: entity background enriches semantic input, while image descriptions supply alignment statistics. The semantic branch combines the original post text, background from a frozen MLLM, and optical character recognition (OCR) text. The aligned visual branch combines global and local visual features with statistics of cross-modal and inter-image similarity. Text-Anchored Smoothed Aggregation (TASA) weights images by their similarity to the original post text while preserving contributions from images with low literal similarity. Summing the branch classification scores integrates contextual semantics with visual evidence. Experiments on the multi-image MMSD3.0 benchmark show that SEAF outperforms all compared methods.

arXiv:2611.25670v1cs.AI

Efficient Latent Reasoning for Vision-Language Model via Dynamics-Aware Quantization and Distillation

Efficient chain-of-thought (CoT) reasoning in vision-language models (VLMs) requires reducing both intermediate reasoning steps and their execution cost. Latent CoT methods compress textual rationales into continuous thoughts, yet their full-precision (FP) models retain substantial storage and computation costs. Lowering numerical precision could reduce these remaining costs, yet direct quantization severely degrades reasoning accuracy, particularly with 4-bit weights and activations. Operator and perturbation analyses reveal key sensitivity and direction-dependent predictive effects of continuous-thought errors, motivating preservation of information retrieval and subsequent reasoning beyond reconstruction accuracy. We therefore introduce QL-VLM, a dynamics-aware Quantization framework for efficient Latent reasoning in Vision-Language Models (VLMs). First, Dynamics-Guided Quantization (DGQ) uses nonuniform codebooks that preserve rotary-key structure, mitigating quantization-induced distortion in vision-language information retrieval. Second, Dynamics-Preserving Distillation (DPD) aligns continuous thoughts, latent-query readouts, and predictions with a full-precision teacher, maintaining predictive consistency through subsequent low-precision reasoning. On ScienceQA-IMG with Qwen2-VL-2B-Instruct, 4-bit QL-VLM achieves 90.43% accuracy and improves baseline by 40.33%, with 16.64x fewer bit-operations and 6.40x lower measured latency than explicit CoT.

arXiv:2611.25667v1cs.RO

CauSyncVLA: Causal Latent World Modeling for Asynchronous Robot Control

Vision-Language-Action (VLA) policies must react while visual inference is still running. Overlapping computation and execution reduces waiting, yet leaves perceptual memory stale and successive action chunks disconnected. We present CauSyncVLA, an asynchronous dual-system controller that aligns perceptual-memory prediction and trajectory correction with actual robot execution. Rather than forecasting pixels or physical states, a Causal Latent World Model predicts current perceptual memory from delayed visual anchors and executed state-action history, supervised through action-sensitive distillation to retain the frozen expert's control response. Conditioned on this updated memory, Predictive Flow Warm-Starting (PFWS) aligns the unexecuted action horizon for single-step flow correction without extra training. On LIBERO-10, CauSyncVLA reaches 93.8% success at 5.04$\times$ inference throughput, improving closed-loop robustness and robot motion smoothness.

arXiv:2611.25650v1cs.LG

Extending Rademacher Complexity-Based Generalization Theory to Domain Generalization

Domain generalization (DG) aims to train models on multiple source domains with distinct distributions to ensure robust performance on unseen target domains. In doing so, it elevates the concept of generalization from the traditional sample level to the distribution level. Existing theoretical frameworks introduce a domain space endowed with a meta-distribution to extend classical generalization analysis to this new setting. However, these approaches typically treat the domain space as devoid of geometry. In reality, the domain space is an inherently infinite-dimensional, nonlinear abstract space composed of probability distributions, and its complex geometric properties fundamentally determine generalization behavior. Consequently, existing geometry-free theories of domain generalization have a fundamental limitation. They lack a formal characterization of what distinguishes distribution-level generalization from sample-level generalization, leaving the inherent difficulty of generalizing across domains largely unexplained. To bridge this gap, we formalize the domain space as the Wasserstein space over the representation space. Building on the geometric properties of this space (specifically the $\mathrm{CD}(K,\infty)$ curvature theory) we derive a novel, geometry-dependent DG error bound. By introducing an additional curvature-dependent confidence term, this bound formally disentangles the two levels of generalization effects and reveals the essential nature that domain generalization is affected by both the invariance and the curvature of the representation space. Finally, we experimentally validate the effect of curvature on domain generalization.

arXiv:2611.25648v1cs.CV

TRACE: Attribute Execution for Sequential Image Editing

Given an image and successive user instructions, sequential image editing should return an updated image that satisfies each new request and preserves earlier requirements that remain active. This lets users develop a result without repeatedly restating or repairing their decisions. The difficulty is that a retrieved image can contain a valid color on an obsolete surface, while the current image may lack the newly requested material. Neither reference selection nor a precise object mask alone specifies how to reconcile attributes occupying the same region. We present TRACE, which separates the continuing attribute from its visual carrier and prioritizes local appearance operations under complete-request checks. Active-goal binding identifies what should persist, carrier selection supplies suitable current or newly generated pixels, and tonal calibration reduces tinting near highlights and shadows. Experiments across frozen editors, independent visual readers, new source images, and complete sequential benchmarks support joint color-material recovery and show how local execution preserves the surrounding scene.

arXiv:2611.25644v1cs.AI

MambaLIFT: Reference-Based Mechanism-Aware Attribution for VMamba

Attributing predictions of visual state-space models remains challenging because input-dependent selective dynamics couple feature selection, recurrent state propagation, and shared computational pathways. We introduce MambaLIFT, a reference-based, multiplier-first attribution framework for VMamba that propagates contributions through its computational graph using interaction-aware rules for selective-scan injection, recurrent memory, LayerNorm, and shared pathways. MambaLIFT decomposes the reference-relative class-logit difference, \(f_c(x)-f_c(x^0)\), by spatial position, VSS block, scan direction, and computational mechanism. Experiments on ImageNet-1K evaluate VMamba-T and VMamba-B, with hyperparameters selected on a calibration set and faithfulness assessed on a disjoint held-out evaluation set. MambaLIFT achieves numerical conservation to floating-point precision across audited attribution interfaces in both configurations. Under mean-colour pixel flipping, it achieves significantly higher mean faithfulness scores than MambaLRP on both configurations and the highest scores among the evaluated methods, including Integrated Gradients with matched references. Under blurred replacement, MambaLIFT is non-inferior to MambaLRP on VMamba-T and equivalent on VMamba-B, indicating that its advantage depends on the replacement protocol. Ablations on VMamba-T show that interaction allocation, temporal propagation, and pathway composition affect attribution maps. Mechanism-level analyses further show that exact relevance conservation does not imply additive effects under finite interventions: the mechanism decomposition accounts for the reference-relative logit difference but does not estimate intervention effects or establish causal validity.

arXiv:2611.25640v1cs.CR

Erasing the Wrong Thing: Conditional Attribute Leakage and Concept Erasure in Shared Foundation-Model Embeddings

Medical-image embeddings are increasingly shared in place of raw images for authorized downstream tasks, yet they may reveal sensitive attributes. Standard concept-erasure methods target population-level guardedness. However, we prove that, when binary task and sensitive labels are associated, global and task-conditional linear guardedness cannot both hold while a nonzero linear task signal is preserved, so global erasure must either eliminate that signal or retain task-conditional linear leakage. We further show how global erasure affects linear readers' discrimination between positive and negative samples with different sensitive attributes, deriving an approximate closed-form prediction of its utility cost from pre-erasure statistics. Motivated by this conflict, we introduce Task-Conditional LEACE (TC-LEACE), a closed-form variant of LEACE that enforces linear guardedness within each task stratum of its fitting data. We evaluate four vision foundation models on 85 CelebA attribute pairs spanning a broad range of associations. After LEACE, conditional leakage persists across all four encoders, the discrepancy between marginal and conditional audits increases with association strength, and predicted utility costs track observed losses. We then examine hospital chest radiographs using three foundation models, including two biomedical models, in which chest-tube status is strongly associated with diagnosis. Although a marginal attack on LEACE-erased embeddings underperforms a diagnosis-only reference, chest-tube status remains detectable within diagnosis strata. TC-LEACE recovers most of LEACE's linear task-utility loss while keeping task-conditional leakage within a prespecified margin of LEACE's.

arXiv:2611.25639v1cs.CV

LG-FiLM: Local-Global Mixture-of-Experts Modulation for RAW-to-RGB Reconstruction

An image signal processor does more than reconstruct an image. It also has to decide how strongly to denoise, how to balance color, and which tone curve maps the scene onto the display, and these settings depend on the scene itself, the curve that suits a backlit sky is not the one that suits an indoor portrait. Most neural ISPs learn a single fixed transformation and apply it to every scene. We instead make the adaptation an explicit part of the network. Our model, LG FiLM, keeps one shared encoder decoder and places a small mixture of affine experts at its bottleneck; the experts scales, shifts, and mixing weights are predicted from local neighborhood evidence together with a global scene descriptor. Reconstruction stays in a single shared pathway and only the processing mode is switched. We evaluate LG FiLM on three datasets, Fujifilm UltraISP, Zurich RAW to RGB, and a COCO derived benchmark built with an inverse ISP. The model reaches the best reconstruction fidelity of the compared entries on the primary benchmark and transfers to the other two.

arXiv:2611.25635v1cs.CV

AesDistill: Distilling Generative Aesthetics for Image Color Enhancement

Image color enhancement aims to improve visual appeal, but previous methods often learn to reproduce ground-truth expert targets. These targets reflect the preferences of a few experts and cover a limited range of aesthetic choices. Matching them does not ensure stronger aesthetic quality across different tastes and scenes. By contrast, large generative models learn rich visual knowledge from large image collections and often produce vivid, appealing image edits. However, direct generation can change the source image's semantic structure along with its colors. In this paper, we introduce AesDistill, a framework for distilling generative aesthetics into image color enhancement. We use 4 leading open-source large generative models to produce 64,644 edited images as teacher targets at a cost of about 3,360 H100 GPU-hours, forming the AesDistill Dataset. AesDistill also includes AesDistill Student, a 1.13M-parameter model that learns teacher color changes through confidence-weighted supervision while preserving the source image's semantic structure. We evaluate AesDistill with 12 aesthetic and 4 image-quality assessment models. Every teacher's targets rank above the FiveK and PPR10K expert target means in the joint assessment. Across both datasets, AesDistill Student outperforms all 20 baseline methods and ranks directly behind the 4 teachers. These results demonstrate that distilling aesthetics from generative models enables a compact student to outperform methods trained on expert targets.

arXiv:2611.25632v1cs.CV

What One Sample Cannot See: Why Fewer Denoising Steps Can Score Higher on Physics Benchmarks

When a storm can head north or south, a forecast down the middle may score better under squared error than either plausible path. The same paradox appears when video generators are tested as world models: fewer denoising steps can raise physics scores even as videos blur or stop moving early. Why? Physics-IQ compares each generated video with one recording, although a scene can have multiple valid futures. An exact decomposition shows its scores can reward reduced variation among generated futures, even when their distribution becomes less accurate. We introduce Kernel Physics-IQ to remove this incentive in expectation, needing only two independent samples per condition. Two-step sampling lowers Kernel Physics-IQ in all five image-to-video models. For CogVideoX, the official score rises by 0.055 while ours falls by 0.131. Changing the sampling schedule reveals further differences official scores miss. In controlled physical systems with known outcome probabilities, our metric favors the true distribution over collapsed predictions. In short, a world model should be judged not only by how closely one video matches what happened, but by how faithfully its predictions represent what could have happened.

arXiv:2611.25628v1cs.CV

RelGeo: Pair-Conditioned Relational Calibration for Frozen Diffusion Correspondence

Frozen diffusion descriptors retrieve strong semantic correspondence candidates, yet cosine nearest-neighbor matching can select incorrect locations despite valid high-ranking alternatives. Across DIFT-SD2.1 and DiTF-FLUX on SPair-71k, AP-10K, and SOCO (CC, SOC, Cross-SOC), 41-69% of Top-1 errors retain a PCK@0.10-valid match among the original Top-5 spatial peaks, revealing a persistent candidate-selection bottleneck. We find that descriptor rankings and current-pair relational evidence can favor different candidates: reciprocal diffusion Query/Key associations and pretrained RoMa transport exhibit distinct, partly complementary preferences. To incorporate this evidence into matching decisions, we introduce RelGeo, a Pair-Conditioned Relational Calibration framework expressing heterogeneous relations as shared source-target coordinates. Spectral Relational Coordinates (SRC) encode reciprocal diffusion affinities, while Warp Identity Coordinates (WIC) encode reliability-weighted bidirectional transport endpoints. Their inner products augment frozen descriptor similarity while preserving full-domain cosine nearest-neighbor matching without additional semantic-correspondence training. RelGeo improves all ten evaluations by an average of 5.51 PCK@0.10 points; fixed-candidate reranking and pair-misalignment controls corroborate improved selection and dependence on current-pair relational alignment.

arXiv:2611.25626v1cs.CV

Physical Laws without Physics Loss? Probing Learned Dynamics in Video World Models through Forced Response

Do video world models learn physical laws without a physics loss, or merely imitate motion? Overall motion and appearance scores cannot settle it. We investigate this through nonlinear forced response: an excitation shakes the support of a suspended object. We introduce dual visual-numeric conditioning and adapt six image-to-video backbones to predict this motion using only the flow-matching loss. Wan predicts motion accurately and generalizes to unseen excitations, objects, poses, and scenes. Counterfactual tests suggest capabilities beyond replaying training videos. But what laws have they learned? We design three tiers of tests using Taylor expansions of the dynamics and geometry. All six reproduce the linear response, including resonance; five capture higher-order dynamics, including nonlinear softening. Remarkably, softening is barely present in training: only 0.74% of the swing's squared displacement. Extensions to a double pendulum and a rocking block show sustained dynamics are learned more faithfully than brief impacts. Yet learned dynamics may fail to appear when state estimation or propagation goes wrong. We turn these tests and findings into strategies to improve video training and generation.

arXiv:2611.25621v1cs.CV

DAWN: Robust Neural Blind Deconvolution via Alternating Wiener Guidance

Blind image deconvolution (BID) is an ill-posed inverse problem that aims to jointly recover a sharp image and blur kernel from a single degraded observation. Under nontrivial and high noise, which is common in photography and scientific imaging such as microscopy, BID remains particularly challenging, as existing methods suffer from noise overfitting and inaccurate kernel estimation. Existing training-data-free learning methods for BID, such as those based on the deep image prior (DIP), often utilize external guidance by including model-based estimates as additional loss terms, which underexploits the information these estimates carry especially under nontrivial noise. To address these challenges, we propose DAWN, a training-data-free DIP framework that alternates between injecting model-based estimates directly into the model input and refining this imperfect estimate with the network's own prior. To provide more reliable kernel estimates for these injections, we apply autoencoding regularization to both the image and kernel networks. On natural and scientific images under varying levels of Gaussian and Poisson noise, DAWN outperforms state-of-the-art methods in both image and kernel quality (achieving up to +4.5 dB over the next best baseline on the Levin dataset at $\sigma = 0.1$). Without any training images, DAWN remains highly competitive with a state-of-the-art diffusion model on FFHQ, despite the latter being trained on a massive number of in-distribution images.

arXiv:2611.25620v1cs.CV

CaT-Movie: Balancing Global Planning and Local Rendering for Long Multi-Shot Video Generation

Current multi-shot video generators render only limited durations per pass, requiring long movies to be stitched from local segments across which narrative progression and visual identity drift. To synthesize long-horizon movies with Coherence across Time, we present CaT-Movie, a multi-shot generation system consisting of a global story planner and a local window renderer. The planner expands user requests via a large language model to preserve, infer, and complete story details, while deterministic code maintains the resulting storyboard across shots to compile self-contained shot prompts and index reference frames from generated history. To execute this guidance, the renderer applies reference-routed sparse self-attention and shot-wise cross-attention to route identities from the indexed frames while isolating per-shot textual conditions. Given the scarcity of high-quality open-source multi-shot data, we construct a dataset of over 350K film shots annotated with semantic state transitions, entity grounding, and multi-granularity requests. We further build CaT-MBench, a benchmark on held-out films with the same annotations. It gives every baseline its native format and counts cross-shot consistency only when a cut opens a new camera setup, preventing repeated framing from passing as consistency. On CaT-MBench, CaT-Movie attains the highest average score and the most consistent characters among all methods. The project page is available at https://jinplu.github.io/CaT-Movie/.

arXiv:2611.25595v1cs.CV

LoopedWAM: Looped World Action Models for Elastic Inference

World action models (WAMs) couple generative video modeling with robotic action prediction but typically require large video and action transformers, limiting deployment efficiency. We present LoopedWAM, a compact WAM that recurrently applies a shared six-block core in each expert, executing 30 transformer layers while storing only 12 blocks per expert. To account for their asymmetric inference costs, the video and action experts can operate at different depths: visual features are computed once per query, whereas action computation repeats at every sampling step. We further propose elastic self-distillation, which trains a single elastic checkpoint, LoopedWAM-E, across action-depth and sampling-step budgets. With only 0.58B policy parameters and no robot pretraining, LoopedWAM achieves 96.1% success on LIBERO, 91.9% on RoboTwin 2.0, and 52.2% on RoboCasaGR1, while using $7.4\times$ less peak GPU memory than Fast-WAM under matched inference measurements. On LIBERO, LoopedWAM-E achieves 96.0% success in 27.6 ms or 95.2% in 24.7 ms on an H100, providing multiple accuracy-latency operating points from a single checkpoint; with two sampling steps and four action loops, it stays within 0.5 points of LoopedWAM on all three benchmarks. Ablations demonstrate that weight-shared depth preserves competitive performance, while aggressive action-depth reduction requires explicit budget-aware training. On a real AgileX PiPER arm, LoopedWAM succeeds in 71.3% of 80 trials across four tasks.

arXiv:2611.25594v1cs.LG

Bradley-Terry Training Can Distort Outcome Probabilities of Initially Accurate Language World Models

Comparative feedback, in which a judge says which of two candidate outcomes is closer to what happened, is a practical way to train language world models that assign probabilities to an action's outcomes. We show that, fitted with a Bradley-Terry objective, it need not teach the true outcome probabilities and can distort accurate ones. On a simulated robot push with enumerated outcome distributions, this objective takes a model that is accurate on unseen layouts (hazard error 2.4 percentage points, pp) and moves 33 to 43 pp of its probability onto outcomes that none of a condition's enumerated branches produces. The cause is the objective, not the comparisons: its optimum, computed before training, predicts where the model ends up, and a likelihood that knows the judge's rule learns from the same comparisons without the shift. The shift is robust: it recurs on unseen conditions, under five candidate laws and a three-fold denser enumeration, and about 29 pp remain without the physically impossible events. Its cost is conditional: with those events it inflates hazard probabilities, and a threshold inherited from the pre-feedback model costs about 3 pp of utility per decision on trained and 1 to 2 pp on unseen conditions, through false rejects; without them the utility loss averages below 1 pp over six seeds, though a small regret in action choice remains. Audits of readout, response and decision trace the effect from probabilities to actions.

arXiv:2611.25593v1cs.CV

Characterizing Age-Conditioned Affective Reasoning in Vision-Language Models

An age prompt can change how a vision-language model (VLM) rates the emotion evoked by an image. Does the change predict how people of different ages rate that same image? We evaluate 17 VLMs on 900 OASIS images using young-, middle-, and older-adult prompts and estimate image-specific age differences from ratings by 822 participants. Without an age cue, models closely track average human valence (median Pearson $r=0.891$). Under an older-adult prompt, 16 of 17 models assign lower valence on average than under a young-adult prompt. Yet after adjusting for image pleasantness, correlations between model and human age effects range from $-0.030$ to $0.093$, against an estimated ceiling of $0.595$. The discrepancy is clearest on 54 sensitive images rated as pleasant. Across models, the older prompt reduces valence by $0.845$ points on average, while older participants rate these images $0.443$ points higher than younger participants. Providing human ratings from other images improves agreement on the direction of age differences, although image-specific fidelity remains modest on 600 held-out images ($r_{\mathrm{part}}=0.105$). These results show why evaluations of demographic personalization in foundation models need human comparisons at the level of individual stimuli. Project page available at https://shahriyar-zaman.github.io/age-conditioned-emotion-reasoning/

arXiv:2611.25578v1cs.CV

QFT-Based Hybrid PointNet-Quantum Neural Networks for LiDAR Point-Cloud Classification

Hybrid quantum-classical learning architectures provide a promising framework for integrating quantum feature transformations into complex data-processing tasks. Among these transformations, the Quantum Fourier Transform (QFT) offers a structured approach to feature processing, motivating an investigation of how its circuit configuration influences predictive performance, quantum resource requirements, and noise robustness. We present a hybrid PointNet-quantum neural network architecture for LiDAR point-cloud classification, combining classical feature extraction, amplitude encoding, and a variational quantum circuit. Five Fourier-processing configurations are systematically evaluated: No-QFT, Full-QFT-After, QFT-no-swap, IQFT-After, and Trainable-QFT. Experiments on KITTI, nuScenes, and PandaSet across three random seeds assess classification performance, class-level behavior, circuit complexity, and sensitivity to the IBM FakeManilaV2 noise model. Under ideal simulation, IQFT-After achieves the highest weighted F1-score on KITTI (0.857 $\pm$ 0.033), compared with 0.838 $\pm$ 0.023 for No-QFT. Removing the final SWAP gates in the QFT circuit reduces circuit complexity while preserving the Pauli-Z measurement features up to a fixed permutation that can be absorbed by the classification head. Under quantum noise, the circuit ranking changes, with IQFT-After exhibiting greater performance degradation and No-QFT and Trainable-QFT showing stronger robustness. Per-class analysis further identifies dataset-dependent differences in minority-class performance. These results demonstrate that Fourier-processing configuration influences both classification behavior and noise sensitivity, highlighting the role of circuit structure and hardware-aware evaluation in hybrid quantum learning for three-dimensional point-cloud data.