Cast shadows are a clean, observable signature of a hidden physical variable: the direction of light. The authors introduce ShadowBench, a benchmark built from real photographs that asks VLMs to read, infer, and validate cast-shadow direction under carefully matched composites. They find a sharp dissociation—VLMs read visible shadows well yet infer and validate them at chance—and show that routing a VLM through an image generation model supplies the missing prior, motivating reconstruction-based objectives for illumination-aware VLMs.

Text-to-speech systems are commonly judged by feeding synthetic speech into an automatic speech recognizer and measuring word error rate against the input text, a procedure that inevitably mixes ASR mistakes with true synthesis errors. This paper isolates the ASR component of that verifier's floor by running the same three recognizers on the original human recordings from LibriSpeech and FLEURS, then comparing three scoring rules: the standard ASR round-trip, a single-recognizer mask that ignores words missed in human speech, and a three-recognizer adjudication. The work matters because it shows that a large share of apparent rare-word TTS error is purely recognizer failure, and it provides principled recommendations—such as reporting per-frequency ASR baselines and avoiding the single-recognizer mask—for fairer TTS evaluation.

Most multimodal benchmarks test perception or reasoning over determinate inputs, leaving cross-modal ambiguity resolution largely unmeasured. TRIAD introduces a diagnostic benchmark of 327 text-image-audio triplets formally constructed so that only the full triplet yields a unique answer, while any two modalities leave multiple options viable. By organizing items into 86 scene groups and evaluating 14 omni-modal systems under closed and open-ended conditions, the paper shows that strong query-level accuracy masks fragile scene consistency and that removing answer options collapses models to single-digit group accuracy, reframing evaluation around tri-modal disambiguation rather than recognition.

This paper studies how non-expert malicious actors can use web-sourced knowledge and LLMs to evade AI-generated image detectors with simple CPU-based image processing. The authors introduce WildFakeToolBench, a benchmark of 156 validated executable tools derived from 220 public sources across 142 domains, synthesized without detector feedback. On 105 images from five generators, they evaluate six state-of-the-art detectors and demonstrate substantial, detector-specific vulnerabilities, with single-tool attack success rates reaching 25%–100% and worst-case multi-tool coverage exceeding 80% for every detector. The work matters because it quantifies a realistic, low-resource threat model grounded in public forums and tutorials, rather than in white-box optimization.

Wearable devices generate rich physiological records, but turning that structured data into patient-readable narratives risks semantic drift that can distort clinical meaning. This work formalizes structured-to-narrative summarization for menstrual health monitoring and proposes a constrained generation plus retrieval-augmented pipeline to preserve clinical alignment across five defined semantic drift conditions. Across 4,500 generated summaries, the authors find that explicit grounding constraints improve alignment, especially uncertainty preservation, yet they come at a measurable cost in readability and increased medical terminology.

Lightweight Vision-Language-Action models such as SmolVLA rely on a single dense action expert for all manipulation tasks, which limits multi-task capacity under tight compute budgets. SmolVLA-MoE replaces the dense FFN layers in the action expert with sparse Mixture-of-Experts blocks—parallel expert FFNs, a token-level router, and a tunable shared expert—so each token adaptively selects a specialized expert without increasing inference cost. By keeping total activated parameters at ~500M (matching the dense baseline) while storing ~800M total parameters, the method raises average success on the LIBERO benchmark from 82.75% to 84.25% and outperforms the much larger OpenVLA-7B, offering a practical efficiency win for resource-constrained robot platforms.

ResGVC attacks the two core weaknesses of current generative video compression (GVC): decoder-side diffusion without explicit source guidance risks drifting away from the original video, and latent-space denoising suffers irreversible detail loss through the autoencoder bottleneck. The paper proposes to decompose the distortion between a learned codec reconstruction and the original pictures into a sequence of pixel-space residual segments, one per flow-matching denoising step, and to compress those residuals with a diffusion-transformer entropy model. By transmitting only the most fidelity-critical segments and generating the rest, ResGVC claims to advance both perceptual realism and signal fidelity, reporting >60% LPIPS and >80% PSNR BD-rate reductions over prior art.

This paper tackles batik image captioning when expert text annotations are scarce by injecting predicted class labels from a LoRA-adapted DINOv2 classifier into a 4-bit NF4 Qwen2.5-VL captioner. The core idea is that compact class supervision can guide vision-language generation in cultural-heritage settings where only sparse metadata exists. The authors evaluate whether the cascade retains correct cultural identities or propagates upstream classification errors across six captioning conditions. Their findings matter for digitizing heritage collections, though they are hedged by overlapping partitions and tiny local test sets.

Video foundation models now handle both understanding and generation of dynamic visual content, but the field lacks a principled way to relate these tasks to an underlying physical world. This survey proposes a world-modeling lens that partitions the literature into four operations—reading, writing, sharing, and interacting—each connecting observed video traces to a latent temporal world. By distinguishing mere temporal coherence in generated sequences from true consistency of the underlying world state, the work reframes evaluation criteria and identifies open challenges in building unified video world models.

The paper addresses vision-based policy learning for laparoscopic robotic manipulation, where end-to-end RL suffers from sparse visual signals and stringent precision requirements, while imitation learning demands scarce expert demonstrations. The authors propose distilling a privileged state-based teacher into an RGB-based student by matching their Gaussian action distributions, with particular emphasis on theoretically analyzing first- and second-order properties of matching objectives including forward KL, reverse KL, Jeffreys divergence, and Wasserstein distance. Extensive experiments in the SurRoL simulator indicate that uncertainty-aware distribution matching—especially Jeffreys divergence—achieves markedly more stable and higher success rates on complex surgical tasks than general RL or imitation learning baselines.

Feed-forward 3D models such as VGGT-Omega deliver excellent local geometry but are limited to short windows, so long videos must be processed as overlapping chunks. HiP3R addresses the resulting frame-level inconsistencies and scale drift with a hierarchical pose graph: a coarse chunk-level $\mathrm{Sim}(3)$ graph first recovers global layout and scale, and then a fixed-layout frame-level $\mathrm{SE}(3)$ graph estimates one shared camera per frame to reconcile conflicting cross-window predictions. It is purely post-processing, requires no retraining or fine-tuning, and applies to both chunk-based and streaming backends.

Most benchmarks for AI tutors assume cooperative, cognitively focused students and ignore the challenging behaviors—refusal, distress, disengagement—that make real tutoring difficult. TutorCaseBench addresses this gap by mining qualitative case studies from pedagogy, educational psychology, and special education, converting them into 923 grounded tutor–student dialogues with case-specific rubrics. The results show that even top LLMs fail to align with documented interventions in nearly half of open-ended cases, revealing a critical weakness in context-sensitive support for vulnerable learners.

Category-level 6D pose estimation recovers object rotation and translation without instance CAD models, but prevailing methods force an unappealing trade-off between brittle intermediate alignments and costly iterative diffusion sampling. F2G-Pose addresses this by proposing a direct feed-forward architecture that lifts frozen foundation-model features onto partial point clouds, fuses local geometry via DGCNN and a Mixture-of-Experts Transformer, and jointly regresses pose, metric size, and camera-frame completed shape in one pass. Its practical significance is pushing deterministic regression to state-of-the-art synthetic-benchmark accuracy while running at 28 FPS and showing real-world robotic deployment potential, though its gains rely heavily on a strong frozen backbone and it offers no explicit uncertainty model.

Decision models such as JEV promise efficient zero-shot classification by predicting over predefined labels instead of generating free-form text, yet their behavior on Vietnamese tasks has been largely uncharted. This study benchmarks four specialized decision models across 13 Vietnamese classification settings spanning topic, emotion, hate-speech, sentiment, intent, and NLI tasks. The authors systematically test English, Vietnamese, and bilingual label descriptions, and they introduce three realistic linguistic perturbations—teencode, typographical noise, and diacritic removal—to probe robustness. The core takeaway is that JEV dominates overall but remains vulnerable to accent stripping, while bilingual label descriptions do not reliably help any model.

As AI-generated videos approach photorealism, detectors must do more than classify real versus fake; they need to surface the specific visual artifacts that ground the decision. Verus-Insight decomposes artifact detection into three explicit capabilities—deciding what to inspect, verifying whether those inspections reveal genuine generation artifacts, and aggregating the verified evidence into a final verdict—and trains an MLLM with human-anchored, full-process supervision to strengthen each stage. The result is a detection system that improves not only video-level accuracy but also the fidelity of fine-grained artifact recovery, supported by a new benchmark with dense human annotations of spatiotemporal artifacts.

PhyEvolve addresses a core failure mode in text-to-video generation: user prompts underspecify physical mechanisms and state dependencies, leading to incoherent dynamics. The paper proposes a training-free, three-agent framework—Director, Physical Evolver, and Composer—that plans explicit causal state trajectories, including during occlusions, and rewrites the prompt before feeding it to a frozen video generator. The work matters because it lifts physical consistency into an explicit planning stage rather than hoping the diffusion model hallucinates correct physics from vague text.

Historical document binarization struggles with faded ink, bleed-through, and stains that behave differently across visible and near-infrared wavelengths. This paper proposes a conditional GAN with adaptive cross-spectral feature fusion at encoder depth $L=3$, gated by edge maps derived from a targeted 600 nm visible + 1000 nm NIR pair. On 30 SMADI pages under strict five-fold document-level cross-validation, the method achieves statistically significant gains in $F_1$, pseudo-FM, and IoU over fixed averaging and generic attention, showing that physics-informed structural guidance can regulate multispectral restoration without discarding any spectral band.

Single-image 3D generators perform well on curated studio photos but fail on real-world images distorted by blur, uneven illumination, or occlusions. Gate3DGen proposes a selective pre-processing pipeline: a degradation-aware router that leaves clean images untouched and a one-step diffusion purifier that repairs corrupted inputs before they reach a frozen 3D generator. The core insight is that improving the observation, rather than retraining the generator, can close most of the fidelity gap while running in a fraction of the time required by generic image editors.

As populations age, robotics and ambient AI are increasingly proposed to help older adults live independently in private homes and residential aged care (RAC) facilities. This chapter frames both settings as extreme human-robot interaction (HRI) environments—cluttered, dynamic, and technically under-supported—and surveys the sensors, perception algorithms, interaction designs, privacy safeguards, and evaluation frameworks needed to make such deployments viable. Its central thesis is that success requires moving beyond lab accuracy toward real-world operational metrics, privacy-by-design, and sustained acceptability.

Code-switched ASR is bottlenecked by scarce real transcribed speech, while synthetic data often mismatches real acoustics and switching patterns. SyRA-CS proposes using Reptile meta-learning on synthetic CS-FLEURS speech to learn a pair-specific initialization that adapts to just one hour of labeled real target data. The paper evaluates this across four architectures and four language pairs, aiming to show that episodic meta-training extracts more transferable knowledge from synthetic data than ordinary continual fine-tuning.